Upload folder using huggingface_hub

.gitattributes

@@ -1,35 +1 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
-*.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
-*.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
 *.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text

De_DiffusionV2_Image.py

@@ -0,0 +1,263 @@
+from dataclasses import dataclass
+from typing import Optional, Tuple
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from transformers.utils import ModelOutput
+from transformers.modeling_utils import PreTrainedModel
+import torchvision.transforms as transforms
+import os
+from safetensors.torch import load_file
+
+from .build_unfreeze import load_sd_model, load_Florence2_model
+from .utils import initiate_time_steps, normalize
+
+
+class MLP(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super(MLP, self).__init__()
+        self.layers = nn.Sequential(
+            nn.Linear(input_dim, output_dim),
+            nn.GELU(),
+            nn.Linear(output_dim, output_dim),
+        )
+        
+    def forward(self, x):
+        return self.layers(x)
+
+@dataclass
+class SDOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+
+class SDModel(PreTrainedModel):
+    def __init__(
+        self,
+        config = None,
+        training_args = None,
+    ):
+        super().__init__(config)
+        self.training_args = training_args
+        if self.training_args.fp32:
+            self._dtype = torch.float32
+        else:
+            self._dtype = torch.bfloat16
+        # Change device to _device to avoid conflict with nn.Module
+        self._device = torch.device(config.device if hasattr(config, 'device') else "cuda" if torch.cuda.is_available() else "cpu")
+        
+        self.vae, self.tokenizer, self.text_encoder, self.unet, self.scheduler = load_sd_model(training_args)
+        torch.cuda.empty_cache()
+        self.unet.eval()
+        self.text_encoder.eval()
+        self.model, self.processor = load_Florence2_model(training_args)
+        self.config = config
+        # Move models to appropriate device
+        self.unet = self.unet.to(self._dtype).to(self._device)
+        self.text_encoder = self.text_encoder.to(self._dtype).to_empty(device=self._device)
+        self.model = self.model.to(self._dtype).to_empty(device=self._device)
+        self.vae = self.vae.to(torch.float32).to_empty(device=self._device)
+
+        self.batch_size = self.training_args.batch_size 
+
+        hidden_dim = 1024 
+        self.language_proj = nn.Sequential(
+            nn.Linear(1024, hidden_dim, dtype=self._dtype),
+            nn.GELU(),
+            nn.Linear(hidden_dim, 1024, dtype=self._dtype)
+        ).to_empty(device=self._device)
+        for param in self.language_proj.parameters():
+            param.requires_grad = True
+        # Add learnable queries for decoder
+        self.num_queries = self.training_args.learnable_token_length # 77 #  # 128
+        self.query_embed = nn.Parameter(torch.randn(1, self.num_queries, 1024, dtype=self._dtype))
+        self.query_embed.requires_grad = True
+        
+        self.unet.enable_gradient_checkpointing()
+
+    def _unet_pred_noise(self, x_start, t, noise, context):
+        # Convert timesteps to long tensor
+        t = t.to(dtype=torch.long)
+        
+        # Ensure consistent dtype for all tensors
+        dtype = self.unet.dtype
+        x_start = x_start.to(dtype)
+        noise = noise.to(dtype)
+        context = context.to(dtype)
+        
+        # Add noise
+        nt = t.shape[0]
+        noised_latent = self.scheduler.add_noise(x_start, noise, t)
+        
+        # Get prediction
+        pred_noise = self.unet(
+            noised_latent, 
+            t, 
+            encoder_hidden_states=context.expand(nt, -1, -1)
+        ).sample
+
+        return pred_noise
+    
+    def generate_images(self, images):
+        batch_size = self.training_args.eval_batch_size
+        prompt = ["<MORE_DETAILED_CAPTION>"] * batch_size
+        inputs = self.processor(text=prompt, images=images, return_tensors="pt").to(self._device).to(self._dtype)
+        # Get embeddings
+        if inputs["input_ids"] is not None:
+            inputs_embeds = self.model.language_model.get_input_embeddings()(inputs["input_ids"]).to(self._dtype)
+        if inputs["pixel_values"] is not None:
+            image_features = self.model._encode_image(inputs["pixel_values"]).to(self._dtype)
+            inputs_embeds, attention_mask = self.model._merge_input_ids_with_image_features(image_features, inputs_embeds)
+        if inputs_embeds is not None:
+            attention_mask = attention_mask.to(inputs_embeds.dtype)
+        encoder_outputs = self.model.language_model.model.encoder(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+        
+        # Prepare decoder inputs
+        decoder_input_embeds = self.query_embed.expand(batch_size, -1, -1)  # [batch_size, 128, 1024]
+        decoder_attention_mask = torch.ones(
+            (batch_size, self.num_queries), 
+            dtype=self._dtype, 
+            device=self._device
+        )
+        
+        encoder_hidden_states = encoder_outputs.last_hidden_state.to(self._dtype)
+        decoder_input_embeds = decoder_input_embeds.to(self._dtype)
+        attention_mask = attention_mask.to(self._dtype)
+        
+        # Run decoder
+        decoder_outputs = self.model.language_model.model.decoder(
+            inputs_embeds=decoder_input_embeds,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+        
+        last_decoder_hidden_state = decoder_outputs.last_hidden_state  # [batch_size, 128, 1024]
+        conditional_context = self.language_proj(last_decoder_hidden_state)
+
+        un_token = self.tokenizer("", padding="max_length", truncation=True,max_length=77, return_tensors="pt").input_ids.to(self._device)
+        un_context_embeddings = self.text_encoder(un_token).last_hidden_state
+        un_context_embeddings = un_context_embeddings.expand(batch_size, -1, -1)
+        if self.training_args.use_text_encoder:
+            context_embeddings = self.text_encoder(
+                inputs_embeds=conditional_context.to(self._dtype)
+            ).last_hidden_state # 1, 77 , 1024
+
+        latent_shape = (batch_size, 4, self.training_args.image_size // 8, self.training_args.image_size // 8)
+        latents = torch.randn(latent_shape, device=self._device, dtype=self._dtype)
+
+        scheduler = self.scheduler
+        scheduler.set_timesteps(self.training_args.num_inference_steps)
+        with torch.no_grad():
+            for t in scheduler.timesteps:
+                latent_model_input = torch.cat([latents, latents], dim=0)
+                latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+
+                combined_embeddings = torch.cat([un_context_embeddings, context_embeddings], dim=0).to(self._dtype)
+                noise_pred = self.unet(
+                    latent_model_input, t, encoder_hidden_states=combined_embeddings
+                )[0]
+
+                noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2, dim=0)
+                noise_pred = noise_pred_uncond + self.training_args.guidance_scale * (noise_pred_cond - noise_pred_uncond)
+
+                latents = scheduler.step(noise_pred, t, latents)[0]
+            # Decode latents with VAE
+
+        scaled_latents = latents / 0.18215
+        with torch.no_grad():
+            decoded_latents = self.vae.decode(scaled_latents.to(torch.float32))[0]
+
+        return decoded_latents
+    
+    def get_conditional_context(self, images, batch_size=None):
+        if batch_size is None:
+            batch_size = self.batch_size
+        prompt = ["<MORE_DETAILED_CAPTION>"] * batch_size
+        inputs = self.processor(text=prompt, images=images, return_tensors="pt").to(self._device).to(self._dtype)
+        # Get embeddings
+        if inputs["input_ids"] is not None:
+            inputs_embeds = self.model.language_model.get_input_embeddings()(inputs["input_ids"]).to(self._dtype)
+        if inputs["pixel_values"] is not None:
+            image_features = self.model._encode_image(inputs["pixel_values"]).to(self._dtype)
+            inputs_embeds, attention_mask = self.model._merge_input_ids_with_image_features(image_features, inputs_embeds)
+        if inputs_embeds is not None:
+            attention_mask = attention_mask.to(inputs_embeds.dtype)
+        encoder_outputs = self.model.language_model.model.encoder(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+        
+        # Prepare decoder inputs
+        decoder_input_embeds = self.query_embed.expand(batch_size, -1, -1)  # [batch_size, 128, 1024]
+        decoder_attention_mask = torch.ones(
+            (batch_size, self.num_queries), 
+            dtype=self._dtype, 
+            device=self._device
+        )
+        
+        encoder_hidden_states = encoder_outputs.last_hidden_state.to(self._dtype)
+        decoder_input_embeds = decoder_input_embeds.to(self._dtype)
+        attention_mask = attention_mask.to(self._dtype)
+        
+        # Run decoder
+        decoder_outputs = self.model.language_model.model.decoder(
+            inputs_embeds=decoder_input_embeds,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+        
+        last_decoder_hidden_state = decoder_outputs.last_hidden_state  # [batch_size, 128, 1024]
+        return last_decoder_hidden_state
+    
+    def forward(
+        self,
+        image=None,
+        filename=None,
+        **kwargs,
+    ) -> SDOutput:
+        images_for_language_model = image
+        normalize_images = normalize(image, rescale=True)
+        x0=self.vae.encode(normalize_images.to(torch.float32)).latent_dist.sample()
+        latent = x0 * 0.18215
+        
+        # prepare_total_timesteps
+        total_timestep = self.scheduler.num_train_timesteps
+
+        # Initiate timesteps and noise
+        timesteps = initiate_time_steps(0, total_timestep, self.batch_size, self.config).long()
+        timesteps = timesteps.to(self._device)
+        c, h, w = latent.shape[1:]
+        if not self.config.tta.use_same_noise_among_timesteps:
+            noise = torch.randn((self.batch_size, c, h, w), device=self._device, dtype=self._dtype)
+        else:
+            noise = torch.randn((1, c, h, w), device=self._device, dtype=self._dtype)
+            noise = noise.repeat(self.batch_size, 1, 1, 1)
+
+        # prepare context for language model
+        conditional_context = self.get_conditional_context(images_for_language_model)
+        conditional_context = self.language_proj(conditional_context) # [b, 159, 1024]
+
+        if self.training_args.use_text_encoder:
+            text_encoder_output = self.text_encoder(input_ids=None, inputs_embeds=conditional_context.to(self._dtype))
+            pred_noise = self._unet_pred_noise(x_start=latent, t=timesteps, noise=noise, context=text_encoder_output.last_hidden_state.to(self._dtype)).to(self._dtype)
+        else:
+            pred_noise = self._unet_pred_noise(x_start=latent, t=timesteps, noise=noise, context=conditional_context.to(self._dtype)).to(self._dtype)
+        # Compute diffusion loss
+        if self.config.tta.loss == "l1":
+            loss = torch.nn.functional.l1_loss(pred_noise, noise)
+        else:
+            loss = torch.nn.functional.mse_loss(pred_noise, noise)
+            
+        return SDOutput(loss=loss)

De_DiffusionV2_stage2.py

@@ -0,0 +1,418 @@
+import sys
+import os
+sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..')))
+
+from dataclasses import dataclass
+from typing import Optional, Tuple, Dict, Any, Union
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from transformers.utils import ModelOutput
+from transformers import PreTrainedModel, PretrainedConfig
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoConfig
+from safetensors.torch import load_file
+import torchvision.transforms as transforms
+
+from .De_DiffusionV2_Image import SDModel
+from .sd_config import SDConfig
+import argparse
+
+def handle_module_prefix(state_dict):
+    """Handle 'module.' prefix in state dict keys."""
+    if any(k.startswith('module.') for k in state_dict.keys()):
+        return {k.replace('module.', ''): v for k, v in state_dict.items()}
+    return state_dict
+
+def create_model_args(args):
+    """Create model arguments needed by SDModel."""
+    model_args = argparse.Namespace()
+    model_args.use_text_encoder = args.use_text_encoder
+    model_args.batch_size = args.batch_size
+    model_args.eval_batch_size = args.batch_size
+    model_args.distributed_strategy = 'none'
+    model_args.fp32 = args.fp32
+    model_args.learnable_token_length = args.learnable_token_length
+    model_args.num_inference_steps = args.num_inference_steps
+    model_args.image_size = args.image_size
+    model_args.guidance_scale = args.guidance_scale
+    model_args.unfreeze_florence2_all = False
+    model_args.unfreeze_florence2_language_model = False
+    model_args.unfreeze_florence2_language_model_decoder = False
+    return model_args
+
+def load_model_checkpoint(model, model_path, device):
+    """Load model checkpoint."""
+    checkpoint = torch.load(model_path, map_location="cpu")
+    
+    # Handle different checkpoint formats
+    if isinstance(checkpoint, dict) and 'model_state_dict' in checkpoint:
+        state_dict = checkpoint['model_state_dict']
+    else:
+        state_dict = checkpoint
+        
+    state_dict = handle_module_prefix(state_dict)
+    model.load_state_dict(state_dict, strict=False)
+
+    return model
+
+def initialize_diffusion_model(args):
+    """Initialize the diffusion model."""
+    config = SDConfig()
+    diffusion_model_args = create_model_args(args)
+    diffusion_model = SDModel(config, diffusion_model_args)
+    _dtype = torch.float32 if diffusion_model_args.fp32 else torch.bfloat16
+
+    # Delete components that aren't needed for inference
+    if hasattr(diffusion_model, 'vae'):
+        del diffusion_model.vae
+    if hasattr(diffusion_model, 'unet'):
+        del diffusion_model.unet
+    
+    # Clear CUDA cache
+    torch.cuda.empty_cache()
+
+    diffusion_model = diffusion_model.to(_dtype)
+    
+    # Freeze parameters that shouldn't be trained
+    for param in diffusion_model.language_proj.parameters():
+        param.requires_grad = False
+    diffusion_model.query_embed.requires_grad = False
+
+    return diffusion_model
+
+class VLV_Config(PretrainedConfig):
+    model_type = "VLV_decoder" 
+
+    def __init__(self, hidden_size=128, **kwargs):
+        super().__init__(**kwargs)
+        pass
+
+class VLV_MODEL(PreTrainedModel):
+    config_class = VLV_Config
+    model_type = "VLV_decoder" 
+
+    def __init__(self, config):
+        super().__init__(config)
+        """Load the CLIPDecoder model."""
+        # Initialize the diffusion model first
+        device = "cuda"
+        de_diffusion_model = initialize_diffusion_model(config)
+        clip_decoder_model = CLIPDecoder(
+            language_model=config.qwen_model,
+            qwen2_config=config.qwen2_config,
+            VLV_model=de_diffusion_model,
+            device=device,
+            bf16=config.mixed_precision
+        )
+        
+        # Load the trained weights
+        # clip_decoder_model = load_model_checkpoint(clip_decoder_model, config.clip_decoder_checkpoint, device)
+        
+        # Set to evaluation mode
+        clip_decoder_model.eval()
+
+        self.model = clip_decoder_model
+        self.max_new_tokens = config.max_length
+        self.num_beams = config.num_beams
+        self.transform = self.get_transform(config.image_size)
+
+    def get_transform(self, image_size):
+        """Transformation pipeline for input images."""
+        return transforms.Compose([
+            transforms.Resize(image_size),
+            transforms.CenterCrop((image_size, image_size)),
+            transforms.PILToTensor(),
+        ])
+
+    def forward(self, valid_images, max_length):
+        valid_images = [self.transform(img) for img in valid_images]
+        if hasattr(self.model, 'module'):
+            outputs = self.model.module.generate(
+                valid_images, 
+                max_new_tokens=max_length, 
+                num_beams=self.num_beams, 
+                early_stopping=True
+            )
+        else:
+            outputs = self.model.generate(
+                valid_images, 
+                max_new_tokens=max_length, 
+                num_beams=self.num_beams, 
+                early_stopping=True
+            )
+        return outputs
+
+
+class MLP(nn.Module):
+    def __init__(self, input_dim, output_dim):
+        super(MLP, self).__init__()
+        self.layers = nn.Sequential(
+            nn.Linear(input_dim, output_dim),
+            nn.GELU(),
+            nn.Linear(output_dim, output_dim),
+        )
+        
+    def forward(self, x):
+        return self.layers(x)
+
+
+@dataclass
+class CLIPDecoderOutput(ModelOutput):
+    """
+    Output class for the CLIP Decoder model.
+    """
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    generated_ids: Optional[torch.LongTensor] = None
+    generated_text: Optional[list] = None
+
+
+class CLIPDecoder(nn.Module):
+
+
+    def __init__(
+        self, 
+        language_model: str,
+        qwen2_config: dict,
+        VLV_model: SDModel,
+        device: torch.device,
+        bf16: bool = False,
+        args: argparse.Namespace = None
+    ):
+        """
+        Initialize the CLIP Decoder model.
+        
+        Args:
+            language_model: Path to the language model
+            VLV_model: The VLV model instance
+            device: The device to run the model on
+            bf16: Whether to use bfloat16 precision
+        """
+        super(CLIPDecoder, self).__init__()
+
+        self._dtype = torch.bfloat16 if bf16 =="bf16" else torch.float32
+        self.qwen2_tokenizer = AutoTokenizer.from_pretrained(language_model)
+        self.qwen2_config = AutoConfig.from_pretrained(language_model)
+        self.qwen2_model = AutoModelForCausalLM.from_config(self.qwen2_config)
+        self.qwen2_model.save_pretrained("QWEN_new_2", safe_serialization=False)
+        # self.qwen2_tokenizer = AutoTokenizer.from_pretrained(language_model)
+        # self.qwen2_model = AutoModelForCausalLM.from_pretrained(language_model,torch_dtype=self._dtype,device_map=None,low_cpu_mem_usage=True)
+        self.VLV_model = VLV_model # fp32 in this case
+        self.device = device
+        self.mlp = MLP(input_dim=1024, output_dim=self.qwen2_model.config.hidden_size)
+        self.ignore_token_id = -100
+
+        # self.qwen2_model = self.qwen2_model.to_empty(device=self.device)
+        # self.VLV_model = self.VLV_model.to_empty(device=self.device)
+        # self.mlp = self.mlp.to_empty(device=self.device)
+        # self.qwen2_model.gradient_checkpointing_enable()
+        # after you load qwen2_model
+        # for p in self.qwen_embed_tokens.parameters():
+        #     p.requires_grad = False
+    
+    
+    def get_conditional_context(self, images, batch_size):
+        """
+        Get conditional context from images using the diffusion model.
+        
+        Args:
+            images: Input images
+            batch_size: Batch size
+            
+        Returns:
+            Decoder hidden states from the diffusion model
+        """
+        prompt = ["<MORE_DETAILED_CAPTION>"] * batch_size
+        inputs = self.VLV_model.processor(text=prompt, images=images, return_tensors="pt").to(self.device).to(torch.float32)
+
+        self.VLV_model = self.VLV_model.to(inputs["input_ids"].device)
+        self.qwen2_model = self.qwen2_model.to(inputs["input_ids"].device)
+        self.mlp = self.mlp.to(inputs["input_ids"].device)
+        self.VLV_model.model.language_model.model = self.VLV_model.model.language_model.model.to(inputs["input_ids"].device)
+        
+        if inputs["input_ids"] is not None:
+            inputs_embeds = self.VLV_model.model.language_model.get_input_embeddings()(inputs["input_ids"]).to(self.device)
+        
+        if inputs["pixel_values"] is not None:
+            image_features = self.VLV_model.model._encode_image(inputs["pixel_values"]).to(self.device)
+            inputs_embeds, attention_mask = self.VLV_model.model._merge_input_ids_with_image_features(
+                image_features, inputs_embeds
+            )
+        
+        if inputs_embeds is not None:
+            attention_mask = attention_mask.to(inputs_embeds.dtype)
+            
+        encoder_outputs = self.VLV_model.model.language_model.model.encoder(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+        
+        decoder_inputs_embeds = self.VLV_model.query_embed.expand(batch_size, -1, -1)
+        decoder_attention_mask = torch.ones(
+            (batch_size, self.VLV_model.num_queries),
+            dtype=torch.float32,
+            device=self.device
+        )
+
+        encoder_hidden_states = encoder_outputs.last_hidden_state.to(torch.float32)
+        decoder_input_embeds = decoder_inputs_embeds.to(torch.float32)
+        attention_mask = attention_mask.to(torch.float32)
+
+        decoder_outputs = self.VLV_model.model.language_model.model.decoder(
+            inputs_embeds=decoder_input_embeds,
+            attention_mask=decoder_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=attention_mask,
+            output_hidden_states=True,
+            return_dict=True
+        )
+            
+        return decoder_outputs.last_hidden_state
+    
+    def process_image(self, images, batch_size):
+        """
+        Process images to get clip text embeddings.
+        
+        Args:
+            images: Input images
+            batch_size: Batch size
+            
+        Returns:
+            Processed clip text embeddings and attention mask
+        """
+        decoder_hidden_states = self.get_conditional_context(images, batch_size)
+        context_embeds = self.VLV_model.language_proj(decoder_hidden_states)
+        clip_text_embeds = self.VLV_model.text_encoder(inputs_embeds=context_embeds).last_hidden_state
+        # clip_text_embeds = clip_text_embeds.to(self._dtype)
+        clip_text_embeds = self.mlp(clip_text_embeds)
+        clip_text_embeds_attention_mask = torch.ones(
+            (batch_size, self.VLV_model.num_queries),
+            dtype=torch.long,
+            device=self.device
+        )
+        
+        return clip_text_embeds, clip_text_embeds_attention_mask
+    
+    def prepare_generation_inputs(self, clip_text_embeds, clip_text_attention_mask=None):
+        """
+        Prepare inputs for text generation.
+        
+        Args:
+            clip_text_embeds: Processed clip text embeddings
+            clip_text_attention_mask: Attention mask for clip text embeddings
+            
+        Returns:
+            Dictionary of generation inputs
+        """
+        if clip_text_attention_mask is None:
+            clip_text_attention_mask = torch.ones(
+                (clip_text_embeds.shape[0], clip_text_embeds.shape[1]),
+                dtype=torch.long,
+                device=clip_text_embeds.device
+            )
+            
+        return {
+            "inputs_embeds": clip_text_embeds,
+            "attention_mask": clip_text_attention_mask
+        }
+    
+    def generate(self, images, max_new_tokens=300, num_beams=4, early_stopping=True):
+        """
+        Generate text from images.
+        
+        Args:
+            images: Input images
+            max_new_tokens: Maximum number of tokens to generate
+            num_beams: Number of beams for beam search
+            early_stopping: Whether to stop early in beam search
+            
+        Returns:
+            CLIPDecoderOutput with generated ids and text
+        """
+        batch_size = len(images)
+        clip_text_embeds, clip_text_attention_mask = self.process_image(images, batch_size)
+        generation_inputs = self.prepare_generation_inputs(clip_text_embeds, clip_text_attention_mask)
+
+        generation_inputs["inputs_embeds"] = generation_inputs["inputs_embeds"].to(torch.bfloat16)
+        generation_inputs["attention_mask"] = generation_inputs["attention_mask"].to(torch.bfloat16)
+    
+        generated_ids = self.qwen2_model.generate(
+            inputs_embeds=generation_inputs["inputs_embeds"],
+            attention_mask=generation_inputs["attention_mask"],
+            max_new_tokens=max_new_tokens,
+            num_beams=num_beams,
+            early_stopping=early_stopping
+        )
+        
+        generated_text = self.qwen2_tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
+        
+        return CLIPDecoderOutput(
+            generated_ids=generated_ids,
+            generated_text=generated_text
+        )
+    
+    def forward(self, images, captions=None):
+        """
+        Forward pass for training.
+        
+        Args:
+            images: Input images
+            captions: Target captions (optional, for training)
+            
+        Returns:
+            CLIPDecoderOutput with loss and logits
+        """
+        batch_size = images.shape[0]
+        
+        # Process images
+        clip_text_embeds, clip_text_attention_mask = self.process_image(images, batch_size)
+        
+        # If no captions provided, return embeddings for generation
+        if captions is None:
+            return CLIPDecoderOutput(
+                last_hidden_state=clip_text_embeds
+            )
+        assert len(captions) == batch_size
+        # Process captions for training
+        qwen_input_ids = self.qwen2_tokenizer(
+            text=captions,
+            truncation=True,
+            return_tensors="pt",
+            padding="max_length",
+            max_length=300,
+            return_token_type_ids=False,
+        ).input_ids
+        assert len(captions) == batch_size
+        qwen_attention_mask = qwen_input_ids.ne(self.qwen2_tokenizer.pad_token_id).to(torch.long).to(self.device)
+        
+        # Prepare labels for training
+        labels = qwen_input_ids
+        labels[labels == self.qwen2_tokenizer.pad_token_id] = self.ignore_token_id
+        labels = labels.to(self.device)
+        # Get embeddings for captions to create the full input sequence
+        labels_for_embeddings = labels.clone()
+        labels_for_embeddings[labels_for_embeddings == self.ignore_token_id] = self.qwen2_tokenizer.pad_token_id
+        clip_text_embeds_qwen = self.qwen2_model.get_input_embeddings()(labels_for_embeddings)
+        
+        # Concatenate the embeddings and prepare attention mask
+        inputs_embeds = torch.cat((clip_text_embeds, clip_text_embeds_qwen), dim=1)
+        clip_seq_len = clip_text_embeds.shape[1]
+        clip_ignore_labels = torch.full((labels.shape[0], clip_seq_len), self.ignore_token_id).to(labels)
+        combined_labels = torch.cat((clip_ignore_labels, labels), dim=1)
+        
+        attention_mask = torch.cat((
+            clip_text_attention_mask,
+            qwen_attention_mask
+        ), dim=1)
+        
+        # Forward through language model
+        outputs = self.qwen2_model(
+            inputs_embeds=inputs_embeds,
+            labels=combined_labels,
+            attention_mask=attention_mask,
+            use_cache=False
+        )
+        return outputs

build_unfreeze.py

@@ -0,0 +1,81 @@
+import torch
+from diffusers import AutoencoderKL, UNet2DConditionModel, DDPMScheduler, EulerDiscreteScheduler
+from transformers import CLIPTokenizer, CLIPImageProcessor, CLIPTextConfig
+from .modeling_clip import CustomCLIPTextModel
+from .processing_florence2 import Florence2Processor
+from .modeling_florence2 import Florence2ForConditionalGeneration
+from .configuration_florence2 import Florence2Config
+
+from diffusers import StableDiffusionPipeline
+from transformers import AutoProcessor, AutoModelForCausalLM 
+
+
+def load_sd_model(training_args):
+    """Load Stable Diffusion model"""
+
+    repo_id = "stabilityai/stable-diffusion-2-1"
+
+    text_encoder = CustomCLIPTextModel.from_pretrained(repo_id, subfolder="text_encoder")
+    tokenizer = CLIPTokenizer.from_pretrained(repo_id, subfolder="tokenizer")
+    vae = AutoencoderKL.from_pretrained(repo_id, subfolder="vae",revision=None)
+    scheduler = DDPMScheduler.from_pretrained(repo_id, subfolder="scheduler")
+    unet = UNet2DConditionModel.from_pretrained(repo_id, subfolder="unet",revision=None)
+
+    for m in [vae, text_encoder, unet]:
+        for param in m.parameters():
+            param.requires_grad = False
+
+    return (vae, tokenizer, text_encoder, unet, scheduler)
+
+
+def load_Florence2_model(training_args):
+    config = Florence2Config.from_pretrained("microsoft/Florence-2-large")
+    config.vision_config.model_type = "davit"
+    config._attn_implementation = "eager"
+    # # Load the model
+    model = Florence2ForConditionalGeneration(config=config)
+    processor = AutoProcessor.from_pretrained("microsoft/Florence-2-large", trust_remote_code=True)
+
+    # freeze the model
+    if training_args.unfreeze_florence2_all:
+        for param in model.parameters():
+            param.requires_grad = True
+    elif training_args.unfreeze_florence2_language_model:
+        for param in model.parameters():
+            param.requires_grad = False
+        for param in model.language_model.parameters():
+            param.requires_grad = True
+        for param in model.language_model.lm_head.parameters():
+            param.requires_grad = False
+
+        model.language_model.lm_head.weight = torch.nn.Parameter(
+        model.language_model.lm_head.weight.detach().clone())
+
+        for p in model.language_model.lm_head.parameters():
+            p.requires_grad = False
+
+
+    elif training_args.unfreeze_florence2_language_model_decoder:
+        # Create a separate embedding layer for decoder
+        original_embeddings = model.language_model.model.shared
+        new_decoder_embeddings = torch.nn.Embedding(
+            num_embeddings=original_embeddings.num_embeddings,
+            embedding_dim=original_embeddings.embedding_dim,
+            padding_idx=original_embeddings.padding_idx
+        )
+        # Copy the weights
+        new_decoder_embeddings.weight.data = original_embeddings.weight.data.clone()
+        
+        # Replace the decoder embeddings
+        model.language_model.model.encoder.embed_tokens = original_embeddings
+        model.language_model.model.decoder.embed_tokens = new_decoder_embeddings
+        for param in model.parameters():
+            param.requires_grad = False
+        for param in model.language_model.model.decoder.parameters():
+            param.requires_grad = True
+        model.language_model.model.decoder.embed_tokens.weight.requires_grad = False
+    else:
+        for param in model.parameters():
+            param.requires_grad = False
+
+    return model, processor

config.json

@@ -0,0 +1,55 @@
+{
+  "architectures": [
+    "VLV_MODEL"
+  ],
+  "auto_map": {
+    "AutoConfig": "De_DiffusionV2_stage2.VLV_Config",
+    "AutoModel": "De_DiffusionV2_stage2.VLV_MODEL",
+    "AutoModelForCausalLM": "De_DiffusionV2_stage2.VLV_MODEL"
+  },
+  "model_type": "VLV_decoder",
+  "batch_size": 1,
+  "deepspeed": true,
+  "distributed": true,
+  "fp32": true,
+  "guidance_scale": 2.0,
+  "hidden_size": 128,
+  "image_size": 768,
+  "learnable_token_length": 77,
+  "local_rank": 0,
+  "mixed_precision": "bf16",
+  "num_inference_steps": 50,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.1",
+  "use_text_encoder": true,
+  "verbose": true,
+  "qwen_model": "Qwen/Qwen2.5-3B",
+  "qwen2_config":{
+    "architectures": [
+      "Qwen2ForCausalLM"
+    ],
+    "attention_dropout": 0.0,
+    "bos_token_id": 151643,
+    "eos_token_id": 151643,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "initializer_range": 0.02,
+    "intermediate_size": 11008,
+    "max_position_embeddings": 32768,
+    "max_window_layers": 36,
+    "model_type": "qwen2",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 36,
+    "num_key_value_heads": 2,
+    "rms_norm_eps": 1e-06,
+    "rope_theta": 1000000.0,
+    "sliding_window": 32768,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "transformers_version": "4.40.1",
+    "use_cache": true,
+    "use_mrope": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  }  
+}

configuration_florence2.py

@@ -0,0 +1,340 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and the HuggingFace Inc. team. All rights reserved.
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import warnings
+""" Florence-2 configuration"""
+
+from typing import Optional
+
+from transformers import AutoConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class Florence2VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2VisionModel`]. It is used to instantiate a Florence2VisionModel
+    according to the specified arguments, defining the model architecture. Instantiating a configuration with the 
+    defaults will yield a similar configuration to that of the Florence2VisionModel architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        drop_path_rate (`float`, *optional*, defaults to 0.1):
+            The dropout rate of the drop path layer.
+        patch_size (`List[int]`, *optional*, defaults to [7, 3, 3, 3]):
+            The patch size of the image.
+        patch_stride (`List[int]`, *optional*, defaults to [4, 2, 2, 2]):
+            The patch stride of the image.
+        patch_padding (`List[int]`, *optional*, defaults to [3, 1, 1, 1]):
+            The patch padding of the image.
+        patch_prenorm (`List[bool]`, *optional*, defaults to [false, true, true, true]):
+            Whether to apply layer normalization before the patch embedding layer.
+        enable_checkpoint (`bool`, *optional*, defaults to False):
+            Whether to enable checkpointing.
+        dim_embed (`List[int]`, *optional*, defaults to [256, 512, 1024, 2048]):
+            The dimension of the embedding layer.
+        num_heads (`List[int]`, *optional*, defaults to [8, 16, 32, 64]):
+            The number of attention heads.
+        num_groups (`List[int]`, *optional*, defaults to [8, 16, 32, 64]):
+            The number of groups.
+        depths (`List[int]`, *optional*, defaults to [1, 1, 9, 1]):
+            The depth of the model.
+        window_size (`int`, *optional*, defaults to 12):
+            The window size of the model.
+        projection_dim (`int`, *optional*, defaults to 1024):
+            The dimension of the projection layer.
+        visual_temporal_embedding (`dict`, *optional*):
+            The configuration of the visual temporal embedding.
+        image_pos_embed (`dict`, *optional*):
+            The configuration of the image position embedding.
+        image_feature_source (`List[str]`, *optional*, defaults to ["spatial_avg_pool", "temporal_avg_pool"]):
+            The source of the image feature.
+    Example:
+
+    ```python
+    >>> from transformers import Florence2VisionConfig, Florence2VisionModel
+
+    >>> # Initializing a Florence2 Vision style configuration
+    >>> configuration = Florence2VisionConfig()
+
+    >>> # Initializing a model (with random weights)
+    >>> model = Florence2VisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "florence2_vision"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        drop_path_rate=0.1,
+        patch_size=[7, 3, 3, 3],
+        patch_stride=[4, 2, 2, 2],
+        patch_padding=[3, 1, 1, 1],
+        patch_prenorm=[False, True, True, True],
+        enable_checkpoint=False,
+        dim_embed=[256, 512, 1024, 2048],
+        num_heads=[8, 16, 32, 64],
+        num_groups=[8, 16, 32, 64],
+        depths=[1, 1, 9, 1],
+        window_size=12,
+        projection_dim=1024,
+        visual_temporal_embedding=None,
+        image_pos_embed=None,
+        image_feature_source=["spatial_avg_pool", "temporal_avg_pool"],
+        **kwargs,
+    ):
+        self.drop_path_rate = drop_path_rate
+        self.patch_size = patch_size
+        self.patch_stride = patch_stride
+        self.patch_padding = patch_padding
+        self.patch_prenorm = patch_prenorm
+        self.enable_checkpoint = enable_checkpoint
+        self.dim_embed = dim_embed
+        self.num_heads = num_heads
+        self.num_groups = num_groups
+        self.depths = depths
+        self.window_size = window_size
+        self.projection_dim = projection_dim
+        self.visual_temporal_embedding = visual_temporal_embedding
+        self.image_pos_embed = image_pos_embed
+        self.image_feature_source = image_feature_source
+
+        super().__init__(**kwargs)
+
+
+
+class Florence2LanguageConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2LanguagePreTrainedModel`]. It is used to instantiate a BART
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the BART
+    [facebook/bart-large](https://huggingface.co/facebook/bart-large) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 51289):
+            Vocabulary size of the Florence2Language model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Florence2LanguageModel`].
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimensionality of the layers and the pooler layer.
+        encoder_layers (`int`, *optional*, defaults to 12):
+            Number of encoder layers.
+        decoder_layers (`int`, *optional*, defaults to 12):
+            Number of decoder layers.
+        encoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        decoder_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        decoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        encoder_ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        classifier_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for classifier.
+        max_position_embeddings (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        encoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        decoder_layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        scale_embedding (`bool`, *optional*, defaults to `False`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+        num_labels (`int`, *optional*, defaults to 3):
+            The number of labels to use in [`Florence2LanguageForSequenceClassification`].
+        forced_eos_token_id (`int`, *optional*, defaults to 2):
+            The id of the token to force as the last generated token when `max_length` is reached. Usually set to
+            `eos_token_id`.
+
+    Example:
+
+    ```python
+    >>> from transformers import Florence2LanguageConfig, Florence2LanguageModel
+
+    >>> # Initializing a Florence2 Language style configuration
+    >>> configuration = Florence2LanguageConfig()
+
+    >>> # Initializing a model (with random weights)
+    >>> model = Florence2LangaugeModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "florence2_language"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {"num_attention_heads": "encoder_attention_heads", "hidden_size": "d_model"}
+
+    def __init__(
+        self,
+        vocab_size=51289,
+        max_position_embeddings=1024,
+        encoder_layers=12,
+        encoder_ffn_dim=4096,
+        encoder_attention_heads=16,
+        decoder_layers=12,
+        decoder_ffn_dim=4096,
+        decoder_attention_heads=16,
+        encoder_layerdrop=0.0,
+        decoder_layerdrop=0.0,
+        activation_function="gelu",
+        d_model=1024,
+        dropout=0.1,
+        attention_dropout=0.0,
+        activation_dropout=0.0,
+        init_std=0.02,
+        classifier_dropout=0.0,
+        scale_embedding=False,
+        use_cache=True,
+        num_labels=3,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        is_encoder_decoder=True,
+        decoder_start_token_id=2,
+        forced_eos_token_id=2,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.encoder_ffn_dim = encoder_ffn_dim
+        self.encoder_layers = encoder_layers
+        self.encoder_attention_heads = encoder_attention_heads
+        self.decoder_ffn_dim = decoder_ffn_dim
+        self.decoder_layers = decoder_layers
+        self.decoder_attention_heads = decoder_attention_heads
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.activation_function = activation_function
+        self.init_std = init_std
+        self.encoder_layerdrop = encoder_layerdrop
+        self.decoder_layerdrop = decoder_layerdrop
+        self.classifier_dropout = classifier_dropout
+        self.use_cache = use_cache
+        self.num_hidden_layers = encoder_layers
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+
+        super().__init__(
+            num_labels=num_labels,
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            is_encoder_decoder=is_encoder_decoder,
+            decoder_start_token_id=decoder_start_token_id,
+            forced_eos_token_id=forced_eos_token_id,
+            **kwargs,
+        )
+
+        # ensure backward compatibility for BART CNN models
+        if self.forced_bos_token_id is None and kwargs.get("force_bos_token_to_be_generated", False):
+            self.forced_bos_token_id = self.bos_token_id
+            warnings.warn(
+                f"Please make sure the config includes `forced_bos_token_id={self.bos_token_id}` in future versions. "
+                "The config can simply be saved and uploaded again to be fixed."
+            )
+
+class Florence2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Florence2ForConditionalGeneration`]. It is used to instantiate an
+    Florence-2 model according to the specified arguments, defining the model architecture. 
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`Florence2VisionConfig`,  *optional*):
+            Custom vision config or dict
+        text_config (`Union[AutoConfig, dict]`, *optional*):
+            The config object of the text backbone. 
+        ignore_index (`int`, *optional*, defaults to -100):
+            The ignore index for the loss function.
+        vocab_size (`int`, *optional*, defaults to 51289):
+            Vocabulary size of the Florence2model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`~Florence2ForConditionalGeneration`]
+        projection_dim (`int`, *optional*, defaults to 1024):
+            Dimension of the multimodal projection space.
+
+    Example:
+
+    ```python
+    >>> from transformers import Florence2ForConditionalGeneration, Florence2Config, CLIPVisionConfig, BartConfig
+
+    >>> # Initializing a clip-like vision config
+    >>> vision_config = CLIPVisionConfig()
+
+    >>> # Initializing a Bart config
+    >>> text_config = BartConfig()
+
+    >>> # Initializing a Florence-2 configuration
+    >>> configuration = Florence2Config(vision_config, text_config)
+
+    >>> # Initializing a model from the florence-2 configuration
+    >>> model = Florence2ForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "florence2"
+    is_composition = False
+
+    def __init__(
+        self,
+        vision_config=None,
+        text_config=None,
+        ignore_index=-100,
+        vocab_size=51289,
+        projection_dim=1024,
+        **kwargs,
+    ):
+        self.ignore_index = ignore_index
+        self.vocab_size = vocab_size
+        self.projection_dim = projection_dim
+        if vision_config is not None:
+            vision_config = PretrainedConfig(**vision_config)
+        self.vision_config = vision_config
+        self.vocab_size = self.vocab_size
+
+        self.text_config = text_config
+        if text_config is not None:
+            self.text_config = Florence2LanguageConfig(**text_config)
+
+
+        super().__init__(**kwargs)
+

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b184db6b67379fca61c92bc26004b60bdd5af7f894369dd869b0c95cae3eba73
+size 4957566496

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6ebd062fa9eec9bb4beb9eaf10a55db4858053aefe1f50fd5d3f40f0a9ef04f
+size 4995486764

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:066571ad0b78fcba6cbbe1dacbbd000081e7bdd59634e78ab5a265fc1cefd892
+size 1553821280

modeling_clip.py

@@ -0,0 +1,155 @@
+from transformers import CLIPTokenizer, CLIPImageProcessor, CLIPTextModel, CLIPPreTrainedModel, CLIPTextConfig
+from transformers.models.clip.modeling_clip import CLIPTextEmbeddings, CLIPEncoder, CLIPAttention, CLIPMLP, CLIPEncoderLayer, _create_4d_causal_attention_mask, _prepare_4d_attention_mask, BaseModelOutputWithPooling
+from typing import Optional, Union, Tuple
+import torch
+from torch import nn
+
+class CustomCLIPTokenizer(CLIPTokenizer):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # Inherit everything from the original tokenizer
+        # No additional initialization needed unless you want to add specific features
+
+class CustomCLIPImageProcessor(CLIPImageProcessor):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # Inherit everything from the original processor
+        # No additional initialization needed unless you want to add specific features
+
+class CustomCLIPTextTransformer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = CLIPTextEmbeddings(config)
+        self.encoder = CLIPEncoder(config)
+        self.final_layer_norm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+        
+        # For `pooled_output` computation
+        self.eos_token_id = config.eos_token_id
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+            
+        if input_ids is None and inputs_embeds is None:
+            raise ValueError("You must provide either input_ids or inputs_embeds")
+
+
+        if inputs_embeds is not None:
+            inputs_embeds = self.embeddings(inputs_embeds=inputs_embeds)
+        else:
+            inputs_embeds = self.embeddings(input_ids=input_ids, position_ids=position_ids)
+
+        # CLIP's text model uses causal mask, prepare it here.
+        causal_attention_mask = _create_4d_causal_attention_mask(
+            inputs_embeds.size()[:-1], inputs_embeds.dtype, device=inputs_embeds.device
+        )
+        
+        # expand attention_mask
+        if attention_mask is not None:
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            attention_mask = _prepare_4d_attention_mask(attention_mask, inputs_embeds.dtype)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            causal_attention_mask=causal_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.final_layer_norm(last_hidden_state)
+
+        # Update the pooled output computation to work with both input types
+        if input_ids is not None:
+            # Use input_ids to find the EOS token position
+            if self.eos_token_id == 2:
+                pooled_output = last_hidden_state[
+                    torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                    input_ids.to(dtype=torch.int, device=last_hidden_state.device).argmax(dim=-1),
+                ]
+            else:
+                pooled_output = last_hidden_state[
+                    torch.arange(last_hidden_state.shape[0], device=last_hidden_state.device),
+                    (input_ids.to(dtype=torch.int, device=last_hidden_state.device) == self.eos_token_id)
+                    .int()
+                    .argmax(dim=-1),
+                ]
+        else:
+            # When using inputs_embeds, use the last token as the pooled output
+            pooled_output = last_hidden_state[:, -1]
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+class CustomCLIPTextModel(CLIPPreTrainedModel):
+    config_class = CLIPTextConfig
+    _no_split_modules = ["CLIPTextEmbeddings", "CLIPEncoderLayer"]
+
+    def __init__(self, config: CLIPTextConfig):
+        super().__init__(config)
+        self.text_model = CustomCLIPTextTransformer(config)
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.text_model.embeddings.token_embedding
+
+    def set_input_embeddings(self, value):
+        self.text_model.embeddings.token_embedding = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        print("input_ids shape:", input_ids)
+        print("attention_mask shape:", attention_mask)
+        print("position_ids shape:", position_ids)
+        if inputs_embeds is not None:
+            print("inputs_embeds shape:", inputs_embeds.shape)
+
+        print("output_attentions:", output_attentions)
+        print("output_hidden_states:", output_hidden_states)
+        print("return_dict:", return_dict)
+
+        return self.text_model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )

processing_florence2.py

@@ -0,0 +1,1147 @@
+# coding=utf-8
+# Copyright 2024 Microsoft and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for Florence-2.
+"""
+
+import re
+import logging
+from typing import List, Optional, Union
+import numpy as np
+import math
+
+import torch
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput, is_valid_image
+from transformers.processing_utils import ProcessorMixin
+from transformers.tokenization_utils_base import (
+    PaddingStrategy,
+    PreTokenizedInput,
+    TextInput,
+    TruncationStrategy,
+)
+from transformers import BartTokenizer, BartTokenizerFast
+from transformers.utils import TensorType
+
+
+logger = logging.getLogger(__name__)
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_url
+def is_url(val) -> bool:
+    return isinstance(val, str) and val.startswith("http")
+
+# Copied from transformers.models.idefics2.processing_idefics2.is_image_or_image_url
+def is_image_or_image_url(elem):
+    return is_url(elem) or is_valid_image(elem)
+
+
+def _is_str_or_image(elem):
+    return isinstance(elem, (str)) or is_image_or_image_url(elem)
+
+
+class Florence2Processor(ProcessorMixin):
+    r"""
+    Constructs a Florence2 processor which wraps a Florence2 image processor and a Florence2 tokenizer into a single processor.
+
+    [`Florence2Processor`] offers all the functionalities of [`CLIPImageProcessor`] and [`BartTokenizerFast`]. See the
+    [`~Florence2Processor.__call__`] and [`~Florence2Processor.decode`] for more information.
+
+    Args:
+        image_processor ([`CLIPImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`BartTokenizerFast`], *optional*):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "CLIPImageProcessor"
+    tokenizer_class = ("BartTokenizer", "BartTokenizerFast")
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+    ):
+        if image_processor is None:
+            raise ValueError("You need to specify an `image_processor`.")
+        if tokenizer is None:
+            raise ValueError("You need to specify a `tokenizer`.")
+        if not hasattr(image_processor, "image_seq_length"):
+            raise ValueError("Image processor is missing an `image_seq_length` attribute.")
+
+        self.image_seq_length = image_processor.image_seq_length
+
+        tokens_to_add = {
+                'additional_special_tokens': \
+                    tokenizer.additional_special_tokens + \
+                    ['<od>', '</od>', '<ocr>', '</ocr>'] + \
+                    [f'<loc_{x}>' for x in range(1000)] + \
+                    ['<cap>', '</cap>', '<ncap>', '</ncap>','<dcap>', '</dcap>', '<grounding>', '</grounding>', '<seg>', '</seg>', '<sep>', '<region_cap>', '</region_cap>', '<region_to_desciption>', '</region_to_desciption>', '<proposal>', '</proposal>', '<poly>', '</poly>', '<and>']
+            }
+        tokenizer.add_special_tokens(tokens_to_add)
+
+        self.tasks_answer_post_processing_type = {
+            '<OCR>': 'pure_text',
+            '<OCR_WITH_REGION>': 'ocr',
+            '<CAPTION>': 'pure_text',
+            '<DETAILED_CAPTION>': 'pure_text',
+            '<MORE_DETAILED_CAPTION>': 'pure_text',
+            '<OD>': 'description_with_bboxes',
+            '<DENSE_REGION_CAPTION>': 'description_with_bboxes',
+            '<CAPTION_TO_PHRASE_GROUNDING>': "phrase_grounding",
+            '<REFERRING_EXPRESSION_SEGMENTATION>': 'polygons',
+            '<REGION_TO_SEGMENTATION>': 'polygons',
+            '<OPEN_VOCABULARY_DETECTION>': 'description_with_bboxes_or_polygons',
+            '<REGION_TO_CATEGORY>': 'pure_text',
+            '<REGION_TO_DESCRIPTION>': 'pure_text',
+            '<REGION_TO_OCR>': 'pure_text',
+            '<REGION_PROPOSAL>': 'bboxes'
+        }
+
+        self.task_prompts_without_inputs = {
+            '<OCR>': 'What is the text in the image?',
+            '<OCR_WITH_REGION>': 'What is the text in the image, with regions?',
+            '<CAPTION>': 'What does the image describe?',
+            '<DETAILED_CAPTION>': 'Describe in detail what is shown in the image.',
+            '<MORE_DETAILED_CAPTION>': 'Describe with a paragraph what is shown in the image.',
+            '<OD>': 'Locate the objects with category name in the image.',
+            '<DENSE_REGION_CAPTION>': 'Locate the objects in the image, with their descriptions.',
+            '<REGION_PROPOSAL>': 'Locate the region proposals in the image.'
+        }
+
+        self.task_prompts_with_input = {
+            '<CAPTION_TO_PHRASE_GROUNDING>': "Locate the phrases in the caption: {input}",
+            '<REFERRING_EXPRESSION_SEGMENTATION>': 'Locate {input} in the image with mask',
+            '<REGION_TO_SEGMENTATION>': 'What is the polygon mask of region {input}',
+            '<OPEN_VOCABULARY_DETECTION>': 'Locate {input} in the image.',
+            '<REGION_TO_CATEGORY>': 'What is the region {input}?',
+            '<REGION_TO_DESCRIPTION>': 'What does the region {input} describe?',
+            '<REGION_TO_OCR>': 'What text is in the region {input}?',
+        }
+
+        self.post_processor = Florence2PostProcesser(tokenizer=tokenizer)
+
+
+        super().__init__(image_processor, tokenizer)
+    
+    def _construct_prompts(self, text):
+        # replace the task tokens with the task prompts if task token is in the text
+        prompts = []
+        for _text in text:
+            # 1. fixed task prompts without additional inputs
+            for task_token, task_prompt in self.task_prompts_without_inputs.items():
+                if task_token in _text:
+                    assert _text == task_token, f"Task token {task_token} should be the only token in the text."
+                    _text = task_prompt
+                    break
+            # 2. task prompts with additional inputs 
+            for task_token, task_prompt in self.task_prompts_with_input.items():
+                if task_token in _text:
+                    _text = task_prompt.format(input=_text.replace(task_token, ''))
+                    break
+            prompts.append(_text)
+        return prompts
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput]] = None,
+        images: ImageInput = None,
+        tokenize_newline_separately: bool = True,
+        padding: Union[bool, str, PaddingStrategy] = False,
+        truncation: Union[bool, str, TruncationStrategy] = None,
+        max_length=None,
+        return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
+        do_resize: bool = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        data_format: Optional["ChannelDimension"] = "channels_first",  # noqa: F821
+        input_data_format: Optional[
+            Union[str, "ChannelDimension"]  # noqa: F821
+        ] = None,
+        resample: "PILImageResampling" = None,  # noqa: F821
+        do_convert_rgb: bool = None,
+        do_thumbnail: bool = None,
+        do_align_long_axis: bool = None,
+        do_rescale: bool = None,
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to BartTokenizerFast's [`~BartTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwrags` arguments to
+        CLIPImageProcessor's [`~CLIPImageProcessor.__call__`] if `images` is not `None`. Please refer to the doctsring
+        of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+            tokenize_newline_separately (`bool`, defaults to `True`):
+                Adds a separately tokenized '\n' at the end of the prompt.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
+                Select a strategy to pad the returned sequences (according to the model's padding side and padding
+                index) among:
+                - `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+                  sequence if provided).
+                - `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
+                  acceptable input length for the model if that argument is not provided.
+                - `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
+                  lengths).
+            max_length (`int`, *optional*):
+                Maximum length of the returned list and optionally padding length (see above).
+            truncation (`bool`, *optional*):
+                Activates truncation to cut input sequences longer than `max_length` to `max_length`.
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`. If `suffix`
+              is provided, the `input_ids` will also contain the suffix input ids.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **labels** -- Labels compatible with training if `suffix` is not None
+        """
+
+        return_token_type_ids = False
+
+        if images is None:
+            raise ValueError("`images` are expected as arguments to a `Florence2Processor` instance.")
+        if text is None:
+            logger.warning_once(
+                "You are using Florence-2 without a text prompt."
+            )
+            text = ""
+
+        if isinstance(text, List) and isinstance(images, List):
+            if len(images) < len(text):
+                raise ValueError(
+                    f"Received {len(images)} images for {len(text)} prompts. Each prompt should be associated with an image."
+                )
+        if _is_str_or_image(text):
+            text = [text]
+        elif isinstance(text, list) and _is_str_or_image(text[0]):
+            pass
+
+        pixel_values = self.image_processor(
+            images,
+            do_resize=do_resize,
+            do_normalize=do_normalize,
+            return_tensors=return_tensors,
+            image_mean=image_mean,
+            image_std=image_std,
+            input_data_format=input_data_format,
+            data_format=data_format,
+            resample=resample,
+            do_convert_rgb=do_convert_rgb,
+        )["pixel_values"]
+
+        if max_length is not None:
+            max_length -= self.image_seq_length  # max_length has to account for the image tokens
+
+        text = self._construct_prompts(text)
+
+        inputs = self.tokenizer(
+            text,
+            return_tensors=return_tensors,
+            padding=padding,
+            max_length=max_length,
+            truncation=truncation,
+            return_token_type_ids=return_token_type_ids,
+        )
+
+        return_data = {**inputs, "pixel_values": pixel_values}
+
+        if return_token_type_ids:
+            labels = inputs["input_ids"].masked_fill(inputs["token_type_ids"] == 0, -100)
+            return_data.update({"labels": labels})
+        return BatchFeature(data=return_data)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.batch_decode with CLIP->Florence2
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.decode with CLIP->Florence2
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BartTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    # Copied from transformers.models.clip.processing_clip.CLIPProcessor.model_input_names with CLIP->Florence2
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+
+    def post_process_generation(self, text=None, sequence=None, transition_beam_score=None, task=None, image_size=None):
+        """
+        Post-process the output of the model to each of the task outputs.
+
+        Args:
+            text (`str`): The text to post-process.
+            task (`str`): The task to post-process the text for.
+            image_size (`Tuple[int, int]`): The size of the image. height x width.
+        """
+
+        task_answer_post_processing_type = self.tasks_answer_post_processing_type.get(task, 'pure_text')
+        task_answer = self.post_processor(
+            text=text,
+            sequence=sequence,
+            transition_beam_score=transition_beam_score,
+            image_size=image_size,
+            parse_tasks=task_answer_post_processing_type,
+        )[task_answer_post_processing_type]
+
+        if task_answer_post_processing_type == 'pure_text':
+            final_answer = task_answer
+            # remove the special tokens
+            final_answer = final_answer.replace('<s>', '').replace('</s>', '')
+        elif task_answer_post_processing_type in ['od', 'description_with_bboxes', 'bboxes']:
+            od_instances = task_answer
+            bboxes_od = [_od_instance['bbox'] for _od_instance in od_instances]
+            labels_od = [str(_od_instance['cat_name']) for _od_instance in od_instances]
+            final_answer = {'bboxes': bboxes_od, 'labels': labels_od}
+            if len(od_instances) and 'score' in od_instances[0]:
+                scores_od = [_od_instance['score'] for _od_instance in od_instances]
+                final_answer['scores'] = scores_od
+        elif task_answer_post_processing_type in ['ocr']:
+            bboxes = [_od_instance['quad_box'] for _od_instance in task_answer]
+            labels = [str(_od_instance['text']) for _od_instance in task_answer]
+            final_answer = {'quad_boxes': bboxes, 'labels': labels}
+        elif task_answer_post_processing_type in ['phrase_grounding']:
+            bboxes = []
+            labels = []
+            for _grounded_phrase in task_answer:
+                for _bbox in _grounded_phrase['bbox']:
+                    bboxes.append(_bbox)
+                    labels.append(_grounded_phrase['cat_name'])
+            final_answer = {'bboxes': bboxes, 'labels': labels}
+        elif task_answer_post_processing_type in ['description_with_polygons', 'polygons']:
+            labels = []
+            polygons = []
+            for result in task_answer:
+                label = result['cat_name']
+                _polygons = result['polygons']
+                labels.append(label)
+                polygons.append(_polygons)
+            final_answer = {'polygons': polygons, 'labels': labels}
+        elif task_answer_post_processing_type in ['description_with_bboxes_or_polygons']:
+            bboxes = []
+            bboxes_labels = []
+            polygons = []
+            polygons_labels = []
+            for result in task_answer:
+                label = result['cat_name']
+                if 'polygons' in result:
+                    _polygons = result['polygons']
+                    polygons.append(_polygons)
+                    polygons_labels.append(label)
+                else:
+                    _bbox = result['bbox']
+                    bboxes.append(_bbox)
+                    bboxes_labels.append(label)
+            final_answer = {'bboxes': bboxes, 'bboxes_labels': bboxes_labels, 'polygons': polygons, 'polygons_labels': polygons_labels}
+        else:
+            raise ValueError('Unknown task answer post processing type: {}'.format(task_answer_post_processing_type))
+
+        final_answer = {
+            task: final_answer}
+        return final_answer 
+
+class BoxQuantizer(object):
+    def __init__(self, mode, bins):
+        self.mode = mode
+        self.bins = bins
+
+    def quantize(self, boxes: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        xmin, ymin, xmax, ymax = boxes.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            quantized_xmin = (
+                xmin / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_ymin = (
+                ymin / size_per_bin_h).floor().clamp(0, bins_h - 1)
+            quantized_xmax = (
+                xmax / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_ymax = (
+                ymax / size_per_bin_h).floor().clamp(0, bins_h - 1)
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        quantized_boxes = torch.cat(
+            (quantized_xmin, quantized_ymin, quantized_xmax, quantized_ymax), dim=-1
+        ).int()
+
+        return quantized_boxes
+
+    def dequantize(self, boxes: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        xmin, ymin, xmax, ymax = boxes.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            # Add 0.5 to use the center position of the bin as the coordinate.
+            dequantized_xmin = (xmin + 0.5) * size_per_bin_w
+            dequantized_ymin = (ymin + 0.5) * size_per_bin_h
+            dequantized_xmax = (xmax + 0.5) * size_per_bin_w
+            dequantized_ymax = (ymax + 0.5) * size_per_bin_h
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        dequantized_boxes = torch.cat(
+            (dequantized_xmin, dequantized_ymin,
+             dequantized_xmax, dequantized_ymax), dim=-1
+        )
+
+        return dequantized_boxes
+
+
+class CoordinatesQuantizer(object):
+    """
+    Quantize coornidates (Nx2)
+    """
+
+    def __init__(self, mode, bins):
+        self.mode = mode
+        self.bins = bins
+
+    def quantize(self, coordinates: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
+        x, y = coordinates.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            quantized_x = (x / size_per_bin_w).floor().clamp(0, bins_w - 1)
+            quantized_y = (y / size_per_bin_h).floor().clamp(0, bins_h - 1)
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        quantized_coordinates = torch.cat(
+            (quantized_x, quantized_y), dim=-1
+        ).int()
+
+        return quantized_coordinates
+
+    def dequantize(self, coordinates: torch.Tensor, size):
+        bins_w, bins_h = self.bins  # Quantization bins.
+        size_w, size_h = size       # Original image size.
+        size_per_bin_w = size_w / bins_w
+        size_per_bin_h = size_h / bins_h
+        assert coordinates.shape[-1] == 2, 'coordinates should be shape (N, 2)'
+        x, y = coordinates.split(1, dim=-1)  # Shape: 4 * [N, 1].
+
+        if self.mode == 'floor':
+            # Add 0.5 to use the center position of the bin as the coordinate.
+            dequantized_x = (x + 0.5) * size_per_bin_w
+            dequantized_y = (y + 0.5) * size_per_bin_h
+
+        elif self.mode == 'round':
+            raise NotImplementedError()
+
+        else:
+            raise ValueError('Incorrect quantization type.')
+
+        dequantized_coordinates = torch.cat(
+            (dequantized_x, dequantized_y), dim=-1
+        )
+
+        return dequantized_coordinates
+
+
+class Florence2PostProcesser(object):
+    r"""
+    Florence-2 post process for converting text prediction to various tasks results. 
+
+    Args:
+        config: A dict of configs.
+        tokenizer: A tokenizer for decoding text to spans.
+        sample config:
+            UNIFIED_POST_PROCESS:
+                # commom configs
+                NUM_BBOX_HEIGHT_BINS: 1000
+                NUM_BBOX_WIDTH_BINS: 1000
+                COORDINATES_HEIGHT_BINS: 1000
+                COORDINATES_WIDTH_BINS: 1000
+                # task specific configs, override the common configs
+                PRASE_TASKS:
+                    - TASK_NAME: 'video_dense_caption'
+                      PATTERN: 'r<time_(\d+)><time_(\d+)>([a-zA-Z0-9 ]+)'
+                      SCORE_MODE: 'avg_cat_name_scores'
+                      NUM_BINS: 100
+                    - TASK_NAME: 'od'
+                      PATTERN: 'r<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>([a-zA-Z0-9 ]+)'
+                      SCORE_MODE: 'avg_cat_name_scores'
+
+    Returns:
+        parsed_dict (dict): A dict of parsed results.
+    """
+    def __init__(
+        self,
+        tokenizer=None
+    ):
+        parse_tasks = []
+        parse_task_configs = {}
+        config = self._create_default_config()
+        for task in config['PARSE_TASKS']:
+            parse_tasks.append(task['TASK_NAME'])
+            parse_task_configs[task['TASK_NAME']] = task
+
+        self.config = config
+        self.parse_tasks = parse_tasks
+        self.parse_tasks_configs = parse_task_configs
+
+        self.tokenizer =  tokenizer
+        if self.tokenizer is not None:
+            self.all_special_tokens = set(self.tokenizer.all_special_tokens)
+
+        self.init_quantizers()
+        self.black_list_of_phrase_grounding = self._create_black_list_of_phrase_grounding()
+
+    def _create_black_list_of_phrase_grounding(self):
+        black_list = {}
+
+        if 'phrase_grounding' in self.parse_tasks and self.parse_tasks_configs['phrase_grounding']['FILTER_BY_BLACK_LIST']:
+            black_list =  set(
+                ['it', 'I', 'me', 'mine',
+                 'you', 'your', 'yours',
+                 'he', 'him', 'his',
+                 'she', 'her', 'hers',
+                 'they', 'them', 'their', 'theirs',
+                 'one', 'oneself',
+                 'we', 'us', 'our', 'ours',
+                 'you', 'your', 'yours',
+                 'they', 'them', 'their', 'theirs',
+                 'mine', 'yours', 'his', 'hers', 'its',
+                 'ours', 'yours', 'theirs',
+                 'myself', 'yourself', 'himself', 'herself', 'itself',
+                 'ourselves', 'yourselves', 'themselves',
+                 'this', 'that',
+                 'these', 'those',
+                 'who', 'whom', 'whose', 'which', 'what',
+                 'who', 'whom', 'whose', 'which', 'that',
+                 'all', 'another', 'any', 'anybody', 'anyone', 'anything',
+                 'each', 'everybody', 'everyone', 'everything',
+                 'few', 'many', 'nobody', 'none', 'one', 'several',
+                 'some', 'somebody', 'someone', 'something',
+                 'each other', 'one another',
+                 'myself', 'yourself', 'himself', 'herself', 'itself',
+                 'ourselves', 'yourselves', 'themselves',
+                 'the image', 'image', 'images', 'the', 'a', 'an', 'a group',
+                 'other objects', 'lots', 'a set',
+                 ]
+            )
+
+        return black_list
+    
+    def _create_default_config(self):
+        config = {
+            'NUM_BBOX_HEIGHT_BINS': 1000,
+            'NUM_BBOX_WIDTH_BINS': 1000,
+            'BOX_QUANTIZATION_MODE': 'floor',
+            'COORDINATES_HEIGHT_BINS': 1000,
+            'COORDINATES_WIDTH_BINS': 1000,
+            'COORDINATES_QUANTIZATION_MODE': 'floor',
+            'PARSE_TASKS': [
+                {
+                    'TASK_NAME': 'od',
+                    'PATTERN': r'([a-zA-Z0-9 ]+)<loc_(\\d+)><loc_(\\d+)><loc_(\\d+)><loc_(\\d+)>',
+                    'SCORE_MODE': 'avg_loc_scores'
+                },
+                {
+                    'TASK_NAME': 'ocr',
+                    'PATTERN':  r'(.+?)<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>',
+                    'AREA_THRESHOLD': 0.00
+                },
+                {
+                    'TASK_NAME': 'phrase_grounding',
+                    'FILTER_BY_BLACK_LIST': True
+                },
+                {
+                    'TASK_NAME': 'pure_text',
+                },
+                {
+                    'TASK_NAME': 'description_with_bboxes',
+                    'SCORE_MODE': 'avg_loc_scores'
+                },
+                {
+                    'TASK_NAME': 'description_with_polygons',
+                },
+                {
+                    'TASK_NAME': 'polygons',
+                },
+                {
+                    'TASK_NAME': 'bboxes',
+                },
+                {
+                    'TASK_NAME': 'description_with_bboxes_or_polygons',
+                }
+            ]
+        }
+
+        return config
+
+    def init_quantizers(self):
+        # we have box_quantizer (od, grounding) and coordinates_quantizer (ocr, referring_segmentation)
+        num_bbox_height_bins = self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
+        num_bbox_width_bins = self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
+        box_quantization_mode = self.config.get('BOX_QUANTIZATION_MODE', 'floor')
+        self.box_quantizer = BoxQuantizer(
+            box_quantization_mode,
+            (num_bbox_width_bins, num_bbox_height_bins),
+        )
+        
+        num_bbox_height_bins = self.config['COORDINATES_HEIGHT_BINS'] if 'COORDINATES_HEIGHT_BINS' in self.config else self.config.get('NUM_BBOX_HEIGHT_BINS', 1000)
+        num_bbox_width_bins = self.config['COORDINATES_WIDTH_BINS'] if 'COORDINATES_WIDTH_BINS' in self.config else self.config.get('NUM_BBOX_WIDTH_BINS', 1000)
+        box_quantization_mode = self.config.get('COORDINATES_QUANTIZATION_MODE') if 'COORDINATES_QUANTIZATION_MODE' in self.config else self.config.get('BOX_QUANTIZATION_MODE', 'floor')
+        self.coordinates_quantizer = CoordinatesQuantizer(
+            box_quantization_mode,
+            (num_bbox_width_bins, num_bbox_height_bins),
+        )
+
+    def decode_with_spans(self, tokenizer, token_ids):
+        filtered_tokens = tokenizer.convert_ids_to_tokens(
+            token_ids, skip_special_tokens=False)
+        assert len(filtered_tokens) == len(token_ids)
+        sub_texts = []
+        for token in filtered_tokens:
+            if token in self.all_special_tokens:
+                sub_texts.append(token)
+            else:
+                if isinstance(tokenizer, (BartTokenizer, BartTokenizerFast)):
+                    sub_text = tokenizer.convert_tokens_to_string([token])
+                else:
+                    raise ValueError(f'type {type(tokenizer)} not supported')
+                sub_texts.append(sub_text)
+
+        text = ''
+        spans = []
+        for sub_text in sub_texts:
+            span = (len(text), len(text) + len(sub_text))  # [start index, end index).
+            text += sub_text
+            spans.append(span)
+        return text, spans
+
+    def parse_od_from_text_and_spans(
+        self,
+        text,
+        pattern,
+        image_size,
+        phrase_centric=False
+    ):
+        parsed = list(re.finditer(pattern, text))
+
+        instances = []
+        for i in range(len(parsed)):
+            # Prepare instance.
+            instance = {}
+
+            if phrase_centric:
+                bbox_bins = [int(parsed[i].group(j)) for j in range(2, 6)]
+            else:
+                bbox_bins = [int(parsed[i].group(j)) for j in range(1, 5)]
+            instance['bbox'] = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            if phrase_centric:
+                instance['cat_name'] = parsed[i].group(1).lower().strip()
+            else:
+                instance['cat_name'] = parsed[i].group(5).lower().strip()
+            instances.append(instance)
+
+        return instances
+
+    def parse_ocr_from_text_and_spans(self, 
+                                     text, 
+                                     pattern, 
+                                     image_size,
+                                     area_threshold=-1.0,
+        ):
+        bboxes = []
+        labels = []
+        text = text.replace('<s>', '')
+        # ocr with regions
+        parsed = re.findall(pattern, text)
+        instances = []
+        image_width, image_height = image_size
+
+        for ocr_line in parsed:
+            ocr_content = ocr_line[0]
+            quad_box = ocr_line[1:]
+            quad_box = [int(i) for i in quad_box]
+            quad_box = self.coordinates_quantizer.dequantize(
+                torch.tensor(np.array(quad_box).reshape(-1, 2)),
+                size=image_size
+            ).reshape(-1).tolist()
+
+            if area_threshold > 0:
+                x_coords = [i for i in quad_box[0::2]]
+                y_coords = [i for i in quad_box[1::2]]
+
+                # apply the Shoelace formula
+                area = 0.5 * abs(sum(x_coords[i] * y_coords[i + 1] - x_coords[i + 1] * y_coords[i] for i in range(4 - 1)))
+
+                if area < (image_width * image_height) * area_threshold:
+                    continue
+
+            bboxes.append(quad_box)
+            labels.append(ocr_content)
+            instances.append({
+                'quad_box': quad_box,
+                'text': ocr_content,
+            })
+        return instances
+
+    def parse_phrase_grounding_from_text_and_spans(self, text, pattern, image_size):
+        # ignore <s> </s> and <pad>
+        cur_span = 0
+        if text.startswith('<s>'):   
+            cur_span += 3
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        pattern = r"([^<]+(?:<loc_\d+>){4,})"
+        phrases = re.findall(pattern, text)
+    
+        # pattern should be text pattern and od pattern
+        pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
+        box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
+
+        instances = []
+        for pharse_text in phrases:
+            phrase_text_strip = pharse_text.replace('<ground>', '', 1)
+            phrase_text_strip = pharse_text.replace('<obj>', '', 1)
+
+            if phrase_text_strip == '':
+                cur_span += len(pharse_text)
+                continue
+
+            # Prepare instance.
+            instance = {}
+
+            # parse phrase, get string 
+            phrase = re.search(pattern, phrase_text_strip)
+            if phrase is None:
+                cur_span += len(pharse_text)
+                continue
+
+            # parse bboxes by box_pattern
+            bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
+            if len(bboxes_parsed) == 0:
+                cur_span += len(pharse_text)
+                continue
+
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            if phrase in self.black_list_of_phrase_grounding:
+                cur_span += len(pharse_text)
+                continue
+
+            # a list of list 
+            bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
+            instance['bbox'] = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            # exclude non-ascii characters
+            phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
+            instance['cat_name'] = phrase
+
+            instances.append(instance)
+
+        return instances
+
+    def parse_description_with_bboxes_from_text_and_spans(
+            self, 
+            text, 
+            spans=None,
+            scores=None,
+            score_mode=None,
+            pattern=None, 
+            image_size=None, 
+            allow_empty_phrase=False
+        ):
+        def find_matched_token_indices(cur_span, token_spans):
+            inds = []
+            for i, token_span in enumerate(token_spans):
+                if not (token_span[1] <= cur_span[0] or token_span[0] >= cur_span[1]):
+                    inds.append(i)
+            return inds
+
+        cur_span = 0
+        if text.startswith('<s>'):   
+            cur_span += 3
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        if allow_empty_phrase:
+            pattern = rf"(?:(?:<loc_\d+>){{4,}})"
+        else:
+            pattern = r"([^<]+(?:<loc_\d+>){4,})"
+        phrases = re.findall(pattern, text)
+    
+        # pattern should be text pattern and od pattern
+        pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_)'
+        box_pattern = r'<loc_(\d+)><loc_(\d+)><loc_(\d+)><loc_(\d+)>'
+
+        instances = []
+        for pharse_text in phrases:
+            phrase_text_strip = pharse_text.replace('<ground>', '', 1)
+            phrase_text_strip = pharse_text.replace('<obj>', '', 1)
+
+            if phrase_text_strip == '' and not allow_empty_phrase:
+                cur_span += len(pharse_text)
+                continue
+
+            # parse phrase, get string 
+            phrase = re.search(pattern, phrase_text_strip)
+            if phrase is None:
+                cur_span += len(pharse_text)
+                continue
+
+            phrase_span = phrase.span()
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            # parse bboxes by box_pattern
+            bboxes_parsed = list(re.finditer(box_pattern, pharse_text))
+            if len(bboxes_parsed) == 0:
+                cur_span += len(pharse_text)
+                continue
+
+            # a list of list 
+            bbox_bins = [[int(_bboxes_parsed.group(j)) for j in range(1, 5)] for _bboxes_parsed in bboxes_parsed]
+
+            bboxes = self.box_quantizer.dequantize(
+                boxes=torch.tensor(bbox_bins),
+                size=image_size
+            ).tolist()  
+
+            if score_mode == 'avg_loc_scores':
+                if spans is None or scores is None:
+                    all_scores = None
+                else:
+                    bbox_end_spans = [_bboxes_parsed.span(0) for _bboxes_parsed in bboxes_parsed]
+                    all_scores = []
+                    for _spans in bbox_end_spans:
+                        token_inds = find_matched_token_indices((_spans[0] + cur_span, _spans[1]+ cur_span), spans)
+                        loc_scores = [scores[token_i] for token_i in token_inds]
+                        score = sum(loc_scores) / len(loc_scores)
+                        all_scores.append(score)
+            elif score_mode == 'avg_cat_name_scores': 
+                if spans is None or scores is None:
+                    all_scores = None
+                else:
+                    cat_name_token_inds = find_matched_token_indices((phrase_span[0] + cur_span, phrase_span[1]+cur_span), spans)
+                    cat_name_scores = [scores[token_i] for token_i in cat_name_token_inds]
+                    score = sum(cat_name_scores) / len(cat_name_scores)
+                    all_scores = [score] * len(bboxes)
+            elif score_mode is None:
+                all_scores = None
+            else:
+                raise ValueError('Unknown score mode: {}'.format(score_mode))
+
+            phrase = phrase.encode('ascii',errors='ignore').decode('ascii')
+            for _idx, _bboxes in enumerate(bboxes):
+                # Prepare instance.
+                instance = {}
+                instance['bbox'] = _bboxes
+                # exclude non-ascii characters
+                instance['cat_name'] = phrase
+                if all_scores is not None:
+                    instance['score'] = math.exp(all_scores[_idx])
+                instances.append(instance)
+            
+            cur_span += len(pharse_text)
+
+        return instances
+
+    def parse_description_with_polygons_from_text_and_spans(self, text, pattern, image_size, 
+                                                            allow_empty_phrase=False,
+                                                            polygon_sep_token='<sep>',
+                                                            polygon_start_token='<poly>',
+                                                            polygon_end_token='</poly>',
+                                                            with_box_at_start=False,
+                                                            ):
+        
+        # ref_seg format: '<expression><x1><y1><x2><y2><><><sep><><><><>'
+        # ignore <s> </s> and <pad>
+
+        text = text.replace('<s>', '')
+        text = text.replace('</s>', '')
+        text = text.replace('<pad>', '')
+
+        if allow_empty_phrase:
+            pattern = rf"(?:(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
+        else:
+            # [^<]+: This part matches one or more characters that are not the < symbol. 
+            # The ^ inside the square brackets [] is a negation, meaning it matches anything except <.
+            #
+            pattern = rf"([^<]+(?:<loc_\d+>|{re.escape(polygon_sep_token)}|{re.escape(polygon_start_token)}|{re.escape(polygon_end_token)}){{4,}})"
+        phrases = re.findall(pattern, text)
+
+        phrase_string_pattern = r'^\s*(.*?)(?=<od>|</od>|<box>|</box>|<bbox>|</bbox>|<loc_|<poly>)'
+        box_pattern =  rf'((?:<loc_\d+>)+)(?:{re.escape(polygon_sep_token)}|$)'
+
+        # one polygons instance is separated by polygon_start_token and polygon_end_token
+        polygons_instance_pattern = rf'{re.escape(polygon_start_token)}(.*?){re.escape(polygon_end_token)}'
+    
+        instances = []
+        for phrase_text in phrases:
+
+            # exclude loc_\d+>
+            # need to get span if want to include category score
+            phrase_text_strip = re.sub(r'^loc_\d+>', '', phrase_text, count=1)
+
+            # phrase = phrase.replace('<poly>', '')
+            # phrase = phrase.replace('poly>', '')
+
+            if phrase_text_strip == '' and not allow_empty_phrase:
+                continue
+
+
+            # parse phrase, get string 
+            phrase = re.search(phrase_string_pattern, phrase_text_strip)
+            if phrase is None:
+                continue
+            phrase = phrase.group()
+            # remove leading and trailing spaces
+            phrase = phrase.strip()
+
+            # parse bboxes by box_pattern
+
+            # split by polygon_start_token and polygon_end_token first using polygons_instance_pattern
+            if polygon_start_token in phrase_text and polygon_end_token in phrase_text:
+                polygons_instances_parsed = list(re.finditer(polygons_instance_pattern, phrase_text))
+            else:
+                polygons_instances_parsed = [phrase_text]
+
+            for _polygons_instances_parsed in polygons_instances_parsed:
+                # Prepare instance.
+                instance = {}
+
+                # polygons_parsed= list(re.finditer(box_pattern, phrase_text))
+                if isinstance(_polygons_instances_parsed, str): 
+                    polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed))
+                else:
+                    polygons_parsed= list(re.finditer(box_pattern, _polygons_instances_parsed.group(1)))
+                if len(polygons_parsed) == 0:
+                    continue
+
+                # a list of list (polygon)
+                bbox = []
+                polygons = []
+                for _polygon_parsed in polygons_parsed:
+                    # group 1: whole <loc_\d+>...</loc_\d+>
+                    _polygon = _polygon_parsed.group(1)
+                    # parse into list of int
+                    _polygon = [int(_loc_parsed.group(1)) for _loc_parsed in re.finditer(r'<loc_(\d+)>', _polygon)]
+                    if with_box_at_start and len(bbox) == 0:
+                        if len(_polygon) > 4:
+                            # no valid bbox prediction
+                            bbox = _polygon[:4]
+                            _polygon = _polygon[4:]
+                        else:
+                            bbox = [0, 0, 0, 0]
+                    # abandon last element if is not paired 
+                    if len(_polygon) % 2 == 1:
+                        _polygon = _polygon[:-1]
+                    
+                    # reshape into (n, 2)
+                    _polygon = self.coordinates_quantizer.dequantize(
+                        torch.tensor(np.array(_polygon).reshape(-1, 2)),
+                        size=image_size
+                    ).reshape(-1).tolist()
+                    # reshape back
+                    polygons.append(_polygon)
+
+                instance['cat_name'] = phrase
+                instance['polygons'] = polygons
+                if len(bbox) != 0:
+                    instance['bbox'] = self.box_quantizer.dequantize(
+                        boxes=torch.tensor([bbox]),
+                        size=image_size
+                    ).tolist()[0]  
+
+                instances.append(instance)
+
+        return instances
+
+    def __call__(
+        self,
+        text=None,
+        sequence=None,
+        transition_beam_score=None,
+        image_size=None,
+        parse_tasks=None,
+    ):
+        """
+        Args:
+            text: model outputs
+            image_size: (width, height)
+            parse_tasks: a list of tasks to parse, if None, parse all tasks.
+
+        """
+        if parse_tasks is not None:
+            if isinstance(parse_tasks, str):
+                parse_tasks = [parse_tasks]
+            for _parse_task in parse_tasks:
+                assert _parse_task in self.parse_tasks, f'parse task {_parse_task} not supported'
+        
+        # sequence or text should be provided 
+        assert sequence is not None or text is not None, 'sequence or text should be provided'
+        assert sequence is None or text is None, 'only one of sequence and text should be provided'
+
+        if sequence is not None:
+            sequence = sequence.tolist()[1:]
+            text, spans = self.decode_with_spans(self.tokenizer, sequence)
+            if transition_beam_score is not None:
+                transition_beam_score = transition_beam_score.tolist()
+                assert len(sequence) == len(transition_beam_score)
+        else:
+            spans = None
+            transition_beam_score = None
+
+        parsed_dict = {
+            'text': text
+        }
+
+        for task in self.parse_tasks:
+            if parse_tasks is not None and task not in parse_tasks:
+                continue
+
+            pattern = self.parse_tasks_configs[task].get('PATTERN', None)
+            score_mode = self.parse_tasks_configs[task].get('SCORE_MODE', None)
+
+            if task == 'ocr':
+                instances = self.parse_ocr_from_text_and_spans(
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    area_threshold=self.parse_tasks_configs[task].get('AREA_THRESHOLD', 0.0),
+                )
+                parsed_dict['ocr'] = instances
+            elif task == 'phrase_grounding':
+                instances = self.parse_phrase_grounding_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['phrase_grounding'] = instances
+            elif task == 'pure_text':
+                parsed_dict['pure_text'] = text 
+            elif task == 'description_with_bboxes':
+                instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                    text,
+                    spans=spans,
+                    scores=transition_beam_score,
+                    score_mode=score_mode,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['description_with_bboxes'] = instances
+            elif task == 'description_with_polygons':
+                instances = self.parse_description_with_polygons_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                )
+                parsed_dict['description_with_polygons'] = instances
+            elif task == 'polygons':
+                instances = self.parse_description_with_polygons_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    allow_empty_phrase=True,
+                )
+                parsed_dict['polygons'] = instances
+            elif task == 'bboxes':
+                instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                    text,
+                    pattern=pattern,
+                    image_size=image_size,
+                    allow_empty_phrase=True,
+                )
+                parsed_dict['bboxes'] = instances
+            elif task == 'description_with_bboxes_or_polygons':
+                if '<poly>' in text:
+                    # only support either polygons or bboxes, not both at the same time
+                    instances = self.parse_description_with_polygons_from_text_and_spans( 
+                        text,
+                        pattern=pattern,
+                        image_size=image_size,
+                    )
+                else:
+                    instances = self.parse_description_with_bboxes_from_text_and_spans( 
+                        text,
+                        pattern=pattern,
+                        image_size=image_size,
+                    )
+                parsed_dict['description_with_bboxes_or_polygons'] = instances
+            else:
+                raise ValueError("task {} is not supported".format(task))
+
+        return parsed_dict

sd_config.py

@@ -0,0 +1,53 @@
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+class EmptyClass(PretrainedConfig):
+    def __init__(self):
+        pass
+class SDConfig(PretrainedConfig):
+
+    def __init__(self,
+                    override_total_steps = -1,
+                    freeze_vae = True,
+                    use_flash = False,
+                    adapt_topk = -1,
+                    loss = 'mse',
+                    mean = [0.485, 0.456, 0.406],
+                    std = [0.229, 0.224, 0.225],
+                    use_same_noise_among_timesteps = False,
+                    random_timestep_per_iteration = True, 
+                    rand_timestep_equal_int = False,
+                    output_dir = './outputs/First_Start',
+                    do_center_crop_size = 384,
+                    architectures = None,
+                    input = None,
+                    model = None,
+                    tta = None,
+                    **kwargs
+                    
+
+    ):
+        super().__init__()
+        self.model = EmptyClass()
+        self.model.override_total_steps = override_total_steps
+        self.model.freeze_vae = freeze_vae
+        self.model.use_flash = use_flash
+        self.tta = EmptyClass()
+        self.tta.gradient_descent = EmptyClass()
+        self.tta.adapt_topk = adapt_topk
+        self.tta.loss = loss
+        self.tta.use_same_noise_among_timesteps = use_same_noise_among_timesteps
+        self.tta.random_timestep_per_iteration = random_timestep_per_iteration 
+        self.tta.rand_timestep_equal_int = rand_timestep_equal_int
+        self.input = EmptyClass()
+        self.input.mean = mean
+        self.input.std = std
+        self.output_dir = output_dir
+        self.do_center_crop_size = do_center_crop_size
+        self.architectures = architectures
+        for k, v in kwargs.items():
+            setattr(self, k, v)
+if __name__ =='__main__':
+    SDConfig()

utils.py

@@ -0,0 +1,157 @@
+"""Utility functions"""
+import importlib
+import random
+
+import torch
+import numpy as np
+from PIL import Image
+
+
+
+class UnNormalize(object):
+    """Unformalize image as: image = (image * std) + mean
+    """
+    def __init__(self, mean, std):
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+
+    def __call__(self, tensor):
+        """
+        Args:
+            tensor: A tensor of shape [C, H, W] or [N, C, H, W]
+
+        Returns:
+            tensor: A tensor of shape [C, H, W] or [N, C, H, W]
+        """
+
+        std = self.std.to(tensor.device)
+        mean = self.mean.to(tensor.device)
+        if tensor.ndim == 3:
+            std, mean = std.view(-1, 1, 1), mean.view(-1, 1, 1)
+        elif tensor.ndim == 4:
+            std, mean = std.view(1, -1, 1, 1), mean.view(1, -1, 1, 1)
+        tensor = (tensor * std) + mean
+        return tensor
+
+
+class VQVAEUnNormalize(UnNormalize):
+    """Unformalize image as:
+    First: image = (image * std) + mean
+    Second: image = (image * 2) - 1
+    """
+    def __call__(self, tensor):
+        """
+        Args:
+            tensor (Tensor): Tensor image of size (C, H, W) or (N, C, H, W)
+                             to be unnormalized.
+        Returns:
+            Tensor: UnNormalized image.
+        """
+        tensor = super().__call__(tensor)
+        tensor = 2 * tensor - 1
+        return tensor
+    
+def normalize(image,rescale=True):
+    
+    if rescale:
+        image = image.float() / 255.0  # Convert to float and rescale to [0, 1]
+    normalize_image = 2*image-1 # normalize to [-1, 1]
+
+    return normalize_image
+
+# train_transforms = transforms.Compose(
+#     [
+#         transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+#         transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
+#         transforms.RandomHorizontalFlip() if args.random_flip else transforms.Lambda(lambda x: x),
+#         transforms.ToTensor(),
+#         transforms.Normalize([0.5], [0.5]),
+#     ]
+# )
+
+
+def mean_list(l):
+    l = [int(_l) for _l in l]
+    return float(sum(l)) / len(l)
+
+
+def segment_mean(x, index):
+    """Function as tf.segment_mean.
+    """
+    x = x.view(-1, x.shape[-1])
+    index = index.view(-1)
+
+    max_index = index.max() + 1
+    sum_x = torch.zeros((max_index, x.shape[-1]),
+                        dtype=x.dtype,
+                        device=x.device)
+    num_index = torch.zeros((max_index,),
+                            dtype=x.dtype,
+                            device=x.device)
+
+    num_index = num_index.scatter_add_(
+        0, index, torch.ones_like(index, dtype=x.dtype))
+    num_index = torch.where(torch.eq(num_index, 0),
+                            torch.ones_like(num_index, dtype=x.dtype),
+                            num_index)
+
+    index_2d = index.view(-1, 1).expand(-1, x.shape[-1])
+    sum_x = sum_x.scatter_add_(0, index_2d, x)
+    mean_x = sum_x.div_(num_index.view(-1, 1))
+
+    return mean_x
+
+
+
+
+
+def initiate_time_steps(step, total_timestep, batch_size, config):
+    """A helper function to initiate time steps for the diffusion model.
+
+    Args:
+        step: An integer of the constant step
+        total_timestep: An integer of the total timesteps of the diffusion model
+        batch_size: An integer of the batch size
+        config: A config object
+
+    Returns:
+        timesteps: A tensor of shape [batch_size,] of the time steps
+    """
+    if config.tta.rand_timestep_equal_int:
+        # the same timestep for each image in the batch
+        interval_val = total_timestep // batch_size
+        start_point = random.randint(0, interval_val - 1)
+        timesteps = torch.tensor(
+            list(range(start_point, total_timestep, interval_val))
+        ).long()
+        return timesteps
+    elif config.tta.random_timestep_per_iteration:
+        # random timestep for each image in the batch
+        return torch.randint(0, total_timestep, (batch_size,)).long()          #default
+    else:
+        # why we need to do this?
+        return torch.tensor([step] * batch_size).long()
+
+
+def instantiate_from_config(config):
+    """A helper function to instantiate a class from a config object.
+    See https://github.com/CompVis/stable-diffusion/blob/main/ldm/util.py
+    """
+    if not "target" in config:
+        if config == '__is_first_stage__':
+            return None
+        elif config == "__is_unconditional__":
+            return None
+        raise KeyError("Expected key `target` to instantiate.")
+    return get_obj_from_str(config["target"])(**config.get("params", dict()))
+
+
+def get_obj_from_str(string, reload=False):
+    """A helper function to instantiate a class from a config object.
+    See https://github.com/CompVis/stable-diffusion/blob/main/ldm/util.py
+    """
+    module, cls = string.rsplit(".", 1)
+    if reload:
+        module_imp = importlib.import_module(module)
+        importlib.reload(module_imp)
+    return getattr(importlib.import_module(module, package=None), cls)