Update model.py
Browse files
model.py
CHANGED
@@ -1,8 +1,8 @@
|
|
1 |
"""
|
2 |
-
Age-Gender Prediction Model
|
3 |
-
|
4 |
|
5 |
-
Usage:
|
6 |
from transformers import pipeline
|
7 |
classifier = pipeline("image-classification", model="abhilash88/age-gender-prediction", trust_remote_code=True)
|
8 |
result = classifier("image.jpg")
|
@@ -14,131 +14,91 @@ import torch.nn as nn
|
|
14 |
from transformers import (
|
15 |
ViTModel,
|
16 |
ViTImageProcessor,
|
17 |
-
|
18 |
-
|
19 |
-
ImageClassificationPipeline
|
20 |
-
Pipeline
|
21 |
)
|
22 |
from PIL import Image
|
23 |
import numpy as np
|
24 |
|
25 |
|
26 |
-
class
|
27 |
-
"""
|
28 |
-
model_type = "age-gender-vit"
|
29 |
|
30 |
-
def __init__(self,
|
31 |
-
super().__init__(**kwargs)
|
32 |
-
self.vit_model_name = kwargs.get("vit_model_name", "google/vit-base-patch16-224")
|
33 |
-
self.hidden_size = kwargs.get("hidden_size", 768)
|
34 |
-
self.intermediate_size = kwargs.get("intermediate_size", 256)
|
35 |
-
self.final_size = kwargs.get("final_size", 64)
|
36 |
-
self.dropout_rate = kwargs.get("dropout_rate", 0.1)
|
37 |
-
|
38 |
-
|
39 |
-
class AgeGenderViTModel(PreTrainedModel):
|
40 |
-
"""Age-Gender ViT Model following HuggingFace standards"""
|
41 |
-
config_class = AgeGenderConfig
|
42 |
-
|
43 |
-
def __init__(self, config=None):
|
44 |
-
if config is None:
|
45 |
-
config = AgeGenderConfig()
|
46 |
super().__init__(config)
|
47 |
|
48 |
-
self.vit = ViTModel
|
49 |
|
50 |
# Age head: 768 → 256 → 64 → 1
|
51 |
self.age_head = nn.Sequential(
|
52 |
-
nn.Linear(
|
53 |
-
nn.
|
|
|
|
|
|
|
|
|
54 |
nn.Linear(64, 1)
|
55 |
)
|
56 |
|
57 |
# Gender head: 768 → 256 → 64 → 1
|
58 |
self.gender_head = nn.Sequential(
|
59 |
-
nn.Linear(
|
60 |
-
nn.
|
61 |
-
nn.
|
|
|
|
|
|
|
|
|
|
|
62 |
)
|
63 |
|
64 |
-
#
|
65 |
-
self.classifier = nn.Linear(
|
|
|
|
|
|
|
66 |
|
67 |
-
def forward(self, pixel_values, **kwargs):
|
68 |
"""Forward pass returning logits for pipeline"""
|
69 |
-
vit_outputs = self.vit(pixel_values=pixel_values)
|
70 |
-
pooled_output = vit_outputs.pooler_output
|
71 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
age_output = self.age_head(pooled_output)
|
73 |
gender_output = self.gender_head(pooled_output)
|
74 |
|
75 |
-
# Create
|
76 |
logits = torch.cat([age_output, gender_output], dim=1)
|
|
|
77 |
return {"logits": logits}
|
78 |
|
79 |
|
80 |
class AgeGenderImageClassificationPipeline(ImageClassificationPipeline):
|
81 |
-
"""
|
82 |
-
Custom pipeline following HuggingFace documentation standards
|
83 |
-
Reference: https://huggingface.co/docs/transformers/add_new_pipeline
|
84 |
-
"""
|
85 |
-
|
86 |
-
def _sanitize_parameters(self, **kwargs):
|
87 |
-
"""Sanitize parameters following HF guidelines"""
|
88 |
-
preprocess_kwargs = {}
|
89 |
-
postprocess_kwargs = {}
|
90 |
-
|
91 |
-
# Handle any custom parameters here if needed
|
92 |
-
if "top_k" in kwargs:
|
93 |
-
postprocess_kwargs["top_k"] = kwargs["top_k"]
|
94 |
-
|
95 |
-
return preprocess_kwargs, {}, postprocess_kwargs
|
96 |
-
|
97 |
-
def preprocess(self, inputs, **kwargs):
|
98 |
-
"""Preprocess inputs following HF guidelines"""
|
99 |
-
# Handle different input types
|
100 |
-
if isinstance(inputs, str):
|
101 |
-
if inputs.startswith(('http://', 'https://')):
|
102 |
-
import requests
|
103 |
-
from io import BytesIO
|
104 |
-
response = requests.get(inputs)
|
105 |
-
inputs = Image.open(BytesIO(response.content)).convert('RGB')
|
106 |
-
else:
|
107 |
-
inputs = Image.open(inputs).convert('RGB')
|
108 |
-
elif isinstance(inputs, np.ndarray):
|
109 |
-
inputs = Image.fromarray(inputs).convert('RGB')
|
110 |
-
elif not isinstance(inputs, Image.Image):
|
111 |
-
inputs = inputs.convert('RGB')
|
112 |
-
|
113 |
-
# Use the model's image processor
|
114 |
-
return super().preprocess(inputs, **kwargs)
|
115 |
-
|
116 |
-
def _forward(self, model_inputs, **kwargs):
|
117 |
-
"""Forward pass following HF guidelines"""
|
118 |
-
return self.model(**model_inputs)
|
119 |
|
120 |
def postprocess(self, model_outputs, top_k=1, **kwargs):
|
121 |
-
"""
|
122 |
-
Postprocess model outputs to age/gender format
|
123 |
-
This is where LABEL_0/LABEL_1 gets converted to real predictions
|
124 |
-
"""
|
125 |
-
# Extract logits from model output
|
126 |
-
logits = model_outputs["logits"]
|
127 |
|
128 |
-
#
|
129 |
-
|
130 |
-
|
|
|
131 |
|
132 |
-
# Apply
|
133 |
-
# age_raw ~0.7 maps to realistic ages using this formula:
|
134 |
age = int(max(18, min(70, ((age_raw - 1.5) / 1.0) * 50 + 20)))
|
135 |
|
136 |
-
# Process gender
|
137 |
gender_prob = gender_raw
|
138 |
gender = "Female" if gender_prob > 0.5 else "Male"
|
139 |
confidence = gender_prob if gender_prob > 0.5 else 1 - gender_prob
|
140 |
|
141 |
-
# Return
|
142 |
return [{
|
143 |
"label": f"{age} years, {gender}",
|
144 |
"score": confidence,
|
@@ -150,13 +110,13 @@ class AgeGenderImageClassificationPipeline(ImageClassificationPipeline):
|
|
150 |
}]
|
151 |
|
152 |
|
153 |
-
#
|
154 |
-
def
|
155 |
-
"""Manual prediction
|
156 |
import torch.nn as nn
|
157 |
from transformers import ViTImageProcessor, ViTModel
|
158 |
|
159 |
-
class
|
160 |
def __init__(self):
|
161 |
super().__init__()
|
162 |
self.vit = ViTModel.from_pretrained('google/vit-base-patch16-224')
|
@@ -181,7 +141,7 @@ def predict_age_gender(image_path):
|
|
181 |
return age_output, gender_output
|
182 |
|
183 |
# Load model
|
184 |
-
model =
|
185 |
model_url = "https://huggingface.co/abhilash88/age-gender-prediction/resolve/main/pytorch_model.bin"
|
186 |
weights = torch.hub.load_state_dict_from_url(model_url, map_location='cpu')
|
187 |
filtered_weights = {k: v for k, v in weights.items() if not k.startswith('classifier.')}
|
@@ -204,7 +164,7 @@ def predict_age_gender(image_path):
|
|
204 |
with torch.no_grad():
|
205 |
age_raw, gender_raw = model(inputs["pixel_values"])
|
206 |
|
207 |
-
# Apply
|
208 |
age_val = age_raw.item()
|
209 |
age = int(max(18, min(70, ((age_val - 1.5) / 1.0) * 50 + 20)))
|
210 |
|
@@ -220,29 +180,22 @@ def predict_age_gender(image_path):
|
|
220 |
}
|
221 |
|
222 |
|
223 |
-
# Test function
|
224 |
if __name__ == "__main__":
|
225 |
-
print("🧪 Testing Age-Gender
|
226 |
|
227 |
try:
|
228 |
-
# Test the one-liner pipeline
|
229 |
from transformers import pipeline
|
230 |
|
|
|
231 |
classifier = pipeline("image-classification", model="abhilash88/age-gender-prediction", trust_remote_code=True)
|
232 |
-
|
233 |
-
# Test with a sample URL
|
234 |
test_url = "https://images.unsplash.com/photo-1507003211169-0a1dd7228f2d?w=300"
|
235 |
result = classifier(test_url)
|
236 |
|
237 |
-
print(f"✅ Pipeline result
|
238 |
-
print(f"✅ Age: {result[0]['age']}, Gender: {result[0]['gender']}")
|
239 |
-
|
240 |
-
# Test manual approach
|
241 |
-
manual_result = predict_age_gender(test_url)
|
242 |
-
print(f"✅ Manual result: {manual_result['summary']}")
|
243 |
|
244 |
-
|
|
|
|
|
245 |
|
246 |
except Exception as e:
|
247 |
-
print(f"❌ Error: {e}")
|
248 |
-
print("Upload the corrected files to enable the one-liner")
|
|
|
1 |
"""
|
2 |
+
Age-Gender Prediction Model - Simplified Working Version
|
3 |
+
Uses standard ViT model_type to avoid CONFIG_MAPPING issues
|
4 |
|
5 |
+
EXACT Usage:
|
6 |
from transformers import pipeline
|
7 |
classifier = pipeline("image-classification", model="abhilash88/age-gender-prediction", trust_remote_code=True)
|
8 |
result = classifier("image.jpg")
|
|
|
14 |
from transformers import (
|
15 |
ViTModel,
|
16 |
ViTImageProcessor,
|
17 |
+
ViTPreTrainedModel,
|
18 |
+
ViTConfig,
|
19 |
+
ImageClassificationPipeline
|
|
|
20 |
)
|
21 |
from PIL import Image
|
22 |
import numpy as np
|
23 |
|
24 |
|
25 |
+
class AgeGenderViTModel(ViTPreTrainedModel):
|
26 |
+
"""Age-Gender ViT Model using standard ViT architecture"""
|
|
|
27 |
|
28 |
+
def __init__(self, config):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
super().__init__(config)
|
30 |
|
31 |
+
self.vit = ViTModel(config, add_pooling_layer=False)
|
32 |
|
33 |
# Age head: 768 → 256 → 64 → 1
|
34 |
self.age_head = nn.Sequential(
|
35 |
+
nn.Linear(config.hidden_size, 256),
|
36 |
+
nn.ReLU(),
|
37 |
+
nn.Dropout(0.1),
|
38 |
+
nn.Linear(256, 64),
|
39 |
+
nn.ReLU(),
|
40 |
+
nn.Dropout(0.1),
|
41 |
nn.Linear(64, 1)
|
42 |
)
|
43 |
|
44 |
# Gender head: 768 → 256 → 64 → 1
|
45 |
self.gender_head = nn.Sequential(
|
46 |
+
nn.Linear(config.hidden_size, 256),
|
47 |
+
nn.ReLU(),
|
48 |
+
nn.Dropout(0.1),
|
49 |
+
nn.Linear(256, 64),
|
50 |
+
nn.ReLU(),
|
51 |
+
nn.Dropout(0.1),
|
52 |
+
nn.Linear(64, 1),
|
53 |
+
nn.Sigmoid()
|
54 |
)
|
55 |
|
56 |
+
# Standard classifier for compatibility
|
57 |
+
self.classifier = nn.Linear(config.hidden_size, 2)
|
58 |
+
|
59 |
+
# Initialize weights
|
60 |
+
self.post_init()
|
61 |
|
62 |
+
def forward(self, pixel_values=None, **kwargs):
|
63 |
"""Forward pass returning logits for pipeline"""
|
|
|
|
|
64 |
|
65 |
+
# Get ViT outputs
|
66 |
+
outputs = self.vit(pixel_values=pixel_values, **kwargs)
|
67 |
+
|
68 |
+
# Use the last hidden state and pool it
|
69 |
+
sequence_output = outputs[0]
|
70 |
+
pooled_output = sequence_output[:, 0] # Use [CLS] token
|
71 |
+
|
72 |
+
# Get age and gender predictions
|
73 |
age_output = self.age_head(pooled_output)
|
74 |
gender_output = self.gender_head(pooled_output)
|
75 |
|
76 |
+
# Create logits for pipeline
|
77 |
logits = torch.cat([age_output, gender_output], dim=1)
|
78 |
+
|
79 |
return {"logits": logits}
|
80 |
|
81 |
|
82 |
class AgeGenderImageClassificationPipeline(ImageClassificationPipeline):
|
83 |
+
"""Custom pipeline that converts model outputs to age/gender"""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
|
85 |
def postprocess(self, model_outputs, top_k=1, **kwargs):
|
86 |
+
"""Convert logits to age/gender predictions"""
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
+
# Extract logits
|
89 |
+
logits = model_outputs["logits"]
|
90 |
+
age_raw = logits[0, 0].item()
|
91 |
+
gender_raw = logits[0, 1].item()
|
92 |
|
93 |
+
# Apply scaling discovered through testing
|
|
|
94 |
age = int(max(18, min(70, ((age_raw - 1.5) / 1.0) * 50 + 20)))
|
95 |
|
96 |
+
# Process gender
|
97 |
gender_prob = gender_raw
|
98 |
gender = "Female" if gender_prob > 0.5 else "Male"
|
99 |
confidence = gender_prob if gender_prob > 0.5 else 1 - gender_prob
|
100 |
|
101 |
+
# Return standard pipeline format
|
102 |
return [{
|
103 |
"label": f"{age} years, {gender}",
|
104 |
"score": confidence,
|
|
|
110 |
}]
|
111 |
|
112 |
|
113 |
+
# Helper function for manual usage
|
114 |
+
def predict_age_gender_manual(image_path):
|
115 |
+
"""Manual prediction without pipeline"""
|
116 |
import torch.nn as nn
|
117 |
from transformers import ViTImageProcessor, ViTModel
|
118 |
|
119 |
+
class SimpleModel(nn.Module):
|
120 |
def __init__(self):
|
121 |
super().__init__()
|
122 |
self.vit = ViTModel.from_pretrained('google/vit-base-patch16-224')
|
|
|
141 |
return age_output, gender_output
|
142 |
|
143 |
# Load model
|
144 |
+
model = SimpleModel()
|
145 |
model_url = "https://huggingface.co/abhilash88/age-gender-prediction/resolve/main/pytorch_model.bin"
|
146 |
weights = torch.hub.load_state_dict_from_url(model_url, map_location='cpu')
|
147 |
filtered_weights = {k: v for k, v in weights.items() if not k.startswith('classifier.')}
|
|
|
164 |
with torch.no_grad():
|
165 |
age_raw, gender_raw = model(inputs["pixel_values"])
|
166 |
|
167 |
+
# Apply scaling
|
168 |
age_val = age_raw.item()
|
169 |
age = int(max(18, min(70, ((age_val - 1.5) / 1.0) * 50 + 20)))
|
170 |
|
|
|
180 |
}
|
181 |
|
182 |
|
|
|
183 |
if __name__ == "__main__":
|
184 |
+
print("🧪 Testing simplified Age-Gender model...")
|
185 |
|
186 |
try:
|
|
|
187 |
from transformers import pipeline
|
188 |
|
189 |
+
# Test pipeline
|
190 |
classifier = pipeline("image-classification", model="abhilash88/age-gender-prediction", trust_remote_code=True)
|
|
|
|
|
191 |
test_url = "https://images.unsplash.com/photo-1507003211169-0a1dd7228f2d?w=300"
|
192 |
result = classifier(test_url)
|
193 |
|
194 |
+
print(f"✅ Pipeline: Age {result[0]['age']}, Gender {result[0]['gender']}")
|
|
|
|
|
|
|
|
|
|
|
195 |
|
196 |
+
# Test manual
|
197 |
+
manual_result = predict_age_gender_manual(test_url)
|
198 |
+
print(f"✅ Manual: {manual_result['summary']}")
|
199 |
|
200 |
except Exception as e:
|
201 |
+
print(f"❌ Error: {e}")
|
|