Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,96 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
---
|
4 |
+
|
5 |
+
## LoRA Training Config
|
6 |
+
An example (fetal abdomen) of LoRA training configuration.
|
7 |
+
```
|
8 |
+
{
|
9 |
+
"pretrained_model_name_or_path": "/root/autodl-tmp/sd_ckpt/v1-5-pruned.safetensors",
|
10 |
+
"v2": false,
|
11 |
+
"v_parameterization": false,
|
12 |
+
"logging_dir": "/root/autodl-tmp/fetal_us/datasets/log",
|
13 |
+
"train_data_dir": "/root/autodl-tmp/fetal_us/datasets/img",
|
14 |
+
"reg_data_dir": "",
|
15 |
+
"output_dir": "/root/autodl-tmp/fetal_us/datasets/output",
|
16 |
+
"max_resolution": "512,512",
|
17 |
+
"learning_rate": "0.0001",
|
18 |
+
"lr_scheduler": "constant",
|
19 |
+
"lr_warmup": "0",
|
20 |
+
"train_batch_size": 1,
|
21 |
+
"epoch": "1",
|
22 |
+
"save_every_n_epochs": "1",
|
23 |
+
"mixed_precision": "fp16",
|
24 |
+
"save_precision": "fp16",
|
25 |
+
"seed": "1234",
|
26 |
+
"num_cpu_threads_per_process": 2,
|
27 |
+
"cache_latents": true,
|
28 |
+
"caption_extension": ".txt",
|
29 |
+
"enable_bucket": true,
|
30 |
+
"gradient_checkpointing": false,
|
31 |
+
"full_fp16": false,
|
32 |
+
"no_token_padding": false,
|
33 |
+
"stop_text_encoder_training": 0,
|
34 |
+
"xformers": false,
|
35 |
+
"save_model_as": "safetensors",
|
36 |
+
"shuffle_caption": false,
|
37 |
+
"save_state": true,
|
38 |
+
"resume": "",
|
39 |
+
"prior_loss_weight": 1.0,
|
40 |
+
"text_encoder_lr": "5e-5",
|
41 |
+
"unet_lr": "0.0001",
|
42 |
+
"network_dim": 128,
|
43 |
+
"lora_network_weights": "",
|
44 |
+
"color_aug": false,
|
45 |
+
"flip_aug": false,
|
46 |
+
"clip_skip": 2,
|
47 |
+
"gradient_accumulation_steps": 1.0,
|
48 |
+
"mem_eff_attn": false,
|
49 |
+
"output_name": "fetal_ultrasound_v1.0",
|
50 |
+
"model_list": "custom",
|
51 |
+
"max_token_length": "75",
|
52 |
+
"max_train_epochs": "1",
|
53 |
+
"max_data_loader_n_workers": "1",
|
54 |
+
"network_alpha": 128,
|
55 |
+
"training_comment": "",
|
56 |
+
"keep_tokens": "0",
|
57 |
+
"lr_scheduler_num_cycles": "",
|
58 |
+
"lr_scheduler_power": "",
|
59 |
+
"persistent_data_loader_workers": false,
|
60 |
+
"bucket_no_upscale": true,
|
61 |
+
"random_crop": false,
|
62 |
+
"bucket_reso_steps": 64.0,
|
63 |
+
"caption_dropout_every_n_epochs": 0.0,
|
64 |
+
"caption_dropout_rate": 0,
|
65 |
+
"optimizer": "AdamW8bit",
|
66 |
+
"optimizer_args": "",
|
67 |
+
"noise_offset": "",
|
68 |
+
"LoRA_type": "Standard",
|
69 |
+
"conv_dim": 1,
|
70 |
+
"conv_alpha": 1
|
71 |
+
}
|
72 |
+
```
|
73 |
+
|
74 |
+
## FU-LoRA: LoRA Models
|
75 |
+
All LoRA models are available on [HaggingFace](https://huggingface.co/fangyijie/fu-lora)
|
76 |
+
| LoRA Model | Rank |
|
77 |
+
| - | - |
|
78 |
+
| fetal_ultrasound_v1.0.safetensors | 128 |
|
79 |
+
| fetal_ultrasound_v2.0.safetensors | 32 |
|
80 |
+
| fetal_ultrasound_v3.0.safetensors | 8 |
|
81 |
+
|
82 |
+
## Paper
|
83 |
+
URL: https://arxiv.org/abs/2407.20072
|
84 |
+
|
85 |
+
## Citation
|
86 |
+
```
|
87 |
+
@misc{wang2024generative,
|
88 |
+
title={Generative Diffusion Model Bootstraps Zero-shot Classification of Fetal Ultrasound Images In Underrepresented African Populations},
|
89 |
+
author={Fangyijie Wang and Kevin Whelan and Guénolé Silvestre and Kathleen M. Curran},
|
90 |
+
year={2024},
|
91 |
+
eprint={2407.20072},
|
92 |
+
archivePrefix={arXiv},
|
93 |
+
primaryClass={eess.IV},
|
94 |
+
url={https://arxiv.org/abs/2407.20072},
|
95 |
+
}
|
96 |
+
```
|