czg1225 / slimsam Goto Github PK

View Code? Open in Web Editor NEW

236.0 6.0 13.0 36.8 MB

SlimSAM: 0.1% Data Makes Segment Anything Slim

License: Apache License 2.0

Python 100.00%

knowledge-distillation model-compression model-pruning segment-anything-model

slimsam's People

Contributors

Stargazers

Watchers

Forkers

liukang1811 hiyyg camenduru paperwave eltociear songyang86 liuguoyou jie311 aiyou9 xenova techthiyanes mali-farooq pengxinnn

slimsam's Issues

AttributeError: 'GELU' object has no attribute 'approximate'

Hi 👋 thanks for the project ❤ I got an error when I tried running inference.py (torch 2.1.0+cu118)

CUDA visible devices: 1
CUDA Device Name: Tesla T4
model_path: checkpoints/SlimSAM-77.pth
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
[<ipython-input-19-14955c5395e4>](https://localhost:8080/#) in <cell line: 165>()
    164 
    165 if __name__ == '__main__':
--> 166     test_model()

15 frames
[/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py](https://localhost:8080/#) in __getattr__(self, name)
   1693             if name in modules:
   1694                 return modules[name]
-> 1695         raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
   1696 
   1697     def __setattr__(self, name: str, value: Union[Tensor, 'Module']) -> None:

AttributeError: 'GELU' object has no attribute 'approximate'

Is there a way to fix it?

Prune ratios for each step

Thank you for your great work! Would you mind sharing the prune ratios for each step? I don't find them in the paper or code. Thanks!

Low IOU when trained with self captured data.

Thanks for your great work. But when I try to train with my own data. The IOU score seems too low. Here is the log.

CUDA visible devices: 1                                                         
CUDA Device Name: NVIDIA GeForce RTX 4090                                       
===========================Parameter Settings===========================                                                                                        
Pruning Ratio: 0.5                                                              
VIT num_heads: 12                       
norm_type: mean                                                                 
imptype: Disturb                        
global: False                                                                   
learning rate: 0.0001
global: False                                                                                                                                           [0/1867]
learning rate: 0.0001                   
a_weight: 0.5                           
round_to 12                             
TRAIN_SIZE 7825 VAL_SIZE 200 GRAD_SIZE 1000 Epochs 20                           
===========================Pruning Start===========================             
/home/user/workspace/SlimSAM-master/torch_pruning/dependency.py:639: UserWarning: Unwrapped parameters detected: ['neck.3.bias', 'neck.1.bias', 'pos_embed', '
neck.1.weight', 'neck.3.weight'].       
 Torch-Pruning will prune the last non-singleton dimension of a parameter. If you wish to customize this behavior, please provide an unwrapped_parameters argume
nt.                                     
  warnings.warn(warning_str)            
vit_b Pruning:                          
  Params: 89578240 => 45116800          
  Macs: 368858711040.0 => 185712844800.0                                        
  Output:                               
torch.Size([1, 256, 64, 64])            
torch.Size([600, 14, 14, 768])                                                  
torch.Size([12, 64, 64, 768])           
torch.Size([12, 64, 64, 3072])                                                  
torch.Size([25, 64, 64, 384])                                                   
------------------------------------------------------                          
                                        
save checkpoint                         
epoch: 0                                                                                                                                                        
IOU: 0.00037980064841486576 Best IOU 0.00037980064841486576                     
epoch: 1                                                                                                                                                        
IOU: 0.0003555969132073369 Best IOU 0.00037980064841486576                      
save checkpoint                                                                                                                                                 
epoch: 2                                                                        
IOU: 0.0004798856262954162 Best IOU 0.0004798856262954162                       
epoch: 3                                                                        
IOU: 0.00038100686134219785 Best IOU 0.0004798856262954162                      
epoch: 4                                                                        
IOU: 0.00033190775964380326 Best IOU 0.0004798856262954162                                                                                                      
epoch: 5                                                                                                                                                        
IOU: 0.00034291165492228654 Best IOU 0.0004798856262954162                      
Epoch 00007: reducing learning rate of group 0 to 5.0000e-05.                   
epoch: 6                                                                                                                                                        
IOU: 0.00033288924349753746 Best IOU 0.0004798856262954162

By the way, my mask files are transformed to COCO RLE format from the origin labeled files which only include polygons. I transformed them first to the binary masks and then to the COCO RLE format. So the transformed JSON files don't have point_coords in the dict as the SA-1B dataset. So I altered the code in the prune_distill_step1.py. Is that the key reason?

                                for example in dict_data:

                                    sub_count += 1

                                    # input_point = np.array(example['point_coords'])
                                    # input_label = np.array([1])

                                    mask = mask_utils.decode(example["segmentation"])
                                    
                                    # point_coords = transform.apply_coords(input_point, original_image_size)
                                    # coords_torch = torch.as_tensor(point_coords, dtype=torch.float, device=device)
                                    # labels_torch = torch.as_tensor(input_label, dtype=torch.int, device=device)
                                    # coords_torch, labels_torch = coords_torch[None, :, :], labels_torch[None, :]
                                    # points = (coords_torch, labels_torch)

                                    # Model inference
                                    image_embedding,_,_,_,_ = model.image_encoder(input_image)
                                    sparse_embeddings, dense_embeddings = model.prompt_encoder(
                                        points=None, # points,
                                        boxes=None,
                                        masks=None,
                                    )

Import SAMModel,SAMProcessor from hugging face

Hi,

Can someone please tell the syntax to load SAMModel,SAMProcessor from hugging face, the library importing step is missing in documentation ?

list index out of range

I'm using the dataset from roboflow. Can you help me?

!CUDA_VISIBLE_DEVICES=0 python prune_distill_step1.py --traindata_path "/kaggle/working/Crop-Fields-LOD-13-14-15-4/train/_annotations.coco.json" --valdata_path "/kaggle/working/Crop-Fields-LOD-13-14-15-4/valid/_annotations.coco.json" --trainsize 480 --valsize 126 --prune_ratio 0.3 --epochs 20

===========================Parameter Settings===========================
Pruning Ratio: 0.3
VIT num_heads: 12
norm_type: mean
imptype: Disturb
global: False
learning rate: 0.0001
a_weight: 0.5
round_to 12
TRAIN_SIZE 480 VAL_SIZE 126 GRAD_SIZE 1000 Epochs 20
Traceback (most recent call last):
File "/kaggle/working/SlimSAM/prune_distill_step1.py", line 295, in
train_model()
File "/kaggle/working/SlimSAM/prune_distill_step1.py", line 118, in train_model
batch = next(grad_iter)
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 634, in next
data = self._next_data()
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1346, in _next_data
return self._process_data(data)
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/dataloader.py", line 1372, in _process_data
data.reraise()
File "/opt/conda/lib/python3.10/site-packages/torch/_utils.py", line 644, in reraise
raise exception
IndexError: Caught IndexError in DataLoader worker process 0.
Original Traceback (most recent call last):
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/_utils/worker.py", line 308, in _worker_loop
data = fetcher.fetch(index)
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 51, in fetch
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/opt/conda/lib/python3.10/site-packages/torch/utils/data/_utils/fetch.py", line 51, in
data = [self.dataset[idx] for idx in possibly_batched_index]
File "/kaggle/working/SlimSAM/load_sam_json.py", line 39, in getitem
image = cv2.imread(self._image_paths[index])
IndexError: list index out of range

https://universe.roboflow.com/cropfields/crop-fields-lod-13-14-15/dataset/4

SlimSAM in transformers.js

Hello 👋 wanted to ping you about recent transformers.js release. @xenova recently ported SlimSAM to run in browser and the demo can be found here in case you'd like check it out. https://huggingface.co/spaces/Xenova/segment-anything-web

Is SlimSAM compatible with SAM?

Hi,

Can we plug in the SlimSAM weights into SAM (by recombining the q, k, v weights into a single matrix per layer)?

If yes, then SlimSAM could be ported easily to the 🤗 hub. Currently I'm getting errors like:

size mismatch for vision_encoder.layers.7.mlp.lin2.bias: copying a param with shape torch.Size([168]) from checkpoint, the shape in current model is torch.Size([768]).
        size mismatch for vision_encoder.layers.8.layer_norm1.weight: copying a param with shape torch.Size([168]) from checkpoint, the shape in current model is torch.Size([768]).

It seems that the dimensions are different per layer based on the pruning. Any way to load such a state dict in PyTorch?

How long did the project take to train?

evaluate algorithms such as SAM, EfficientSAM, etc.

Hi, I found your work,it's great.

I am a newbie having some difficulties in reproducing your experiment and would like to seek your help.

I found that you did a comparison experiment in your paper, using the SA-1B dataset, and I successfully reproduced the training and validation results of SlimSAM in it.

However, since you didn't release the validation scripts for the time being, I had to follow your experimental setup（https://github.com/czg1225/SlimSAM/issues/5#issuecomment-1875436503） and write my own scripts to validate FastSAM, MobileSAM, EfficientSAM and other algorithms on SA-1B, but I couldn't reproduce the results listed in your table very well, especially the EfficientSAM（only 28%） and EdgeSAM algorithms.

I wonder if it would be convenient for you to put out the validation scripts for these comparison algorithms for learning purposes, I would appreciate it.Thank you very much!

Question about del_pos_init and get_pos_init

Hi! Could you please tell me what is the meaning of del_pos_init and get_pos_init? There is no such operations in the original torch_pruning when pruning ViT or Swin transformer. Why is positional embedding removed here and what is the effect of this on pruning?

License

Under what license are the model and code released?

Batch Prompting for Multiple Boxes

Hi @czg1225

Thanks for this repo!

Can you please tell how can I do batch prompting by giving multiple boxes as prompts to the model?

Right now, predict function takes one numpy array of length 4

Also can you please clarify that for custom dataset training, does each image file should have a corresponding json or a single json for the whole folder is enough since I am getting this error. I have se gradsize to be 100 and 1000. My training images are 1306 and validation are 293.

Nice Work ! Could you provide the configuration about the SA-1B dataset?

Hi, could you please describe in detail how the SA-1B dataset is split into the training and the validation set in this work?