nvlabs / gcvit Goto Github PK

Hi thanks for sharing the great work. Some problems encountered during installation. README.md said "This repository is compatible with NVIDIA PyTorch docker nvcr>=21.06", my confusion is whether “NVIDIA PyTorch docker” is necessary and whether it is possible to simply install the correct Pytorch instead of installing “NVIDIA PyTorch docker”.
Can you provide the complete requirements includes the version of Pytorch and other dependencies required for the project?
Looking forward to your reply, thank you!

Could you please publish semantic segmentation code?

It is not clear which parameters depend on pretrained image size and which on actual one.

E.g. in Swin transformer for classification there is no shift in the last (4-th) stage (BasicLayer) because of https://github.com/microsoft/Swin-Transformer/blob/main/models/swin_transformer.py#L192
But there is no such condition in semantic segmentation backbone https://github.com/SwinTransformer/Swin-Transformer-Semantic-Segmentation/blob/main/mmseg/models/backbones/swin_transformer.py#L182

So, it is unclear how correctly apply GCViT for segmentation on image size different from pretrained.

I am not using distributed training.

Here is the error:

My python command is:

python3 train.py --model gcvit_xxtiny --pretrained --data_dir /gcvit/datasets --dataset May-25-2023 --num-classes 3 --batch-size 32 -c configs/gc_vit_xxtiny.yml --crop-pct 1.0

My dataset is an ImageNet style dataset with "train" and "validation" folders both containing 3 folders for each class (hence num-classes 3).

There are also a couple of other errors that I had to get around (but it worked) - I'll put in a pull request in the coming days to fix the train.py file

Lastly, curiously, changing the num-classes to 4, started the training without any errors.

PS. I'm using --crop-pct 1.0 because I don't want to lose any of my image information during training or validation. I'm curious as to why a 0.875 crop is used for validation - why would you want to lose information?

Thank you,
Rohin Nanavati

hello, i validate the GCViT using your public checkpoint and the accuracy for tiny, xtiny, xxtiny is very low(top1_error 99.9, top5_error 99.75)
I wonder either you public the wrong checkpoints or my validation bash has something wrong?

 python validate.py \
 --model gc_vit_xxtiny \
 --checkpoint ckpts/gcvit_xxtiny_best_1k.pth.tar \
 --data_dir data/imagenet/ \
 --batch-size 32

Will you provide example to downstream tasks like object detection?

If click the model link, the connection is not possible due to the wrong address. can you please check it?

Hello,

I am the developer of flaim, a library of pre-trained vision models for the JAX/Flax ecosystem, which includes GC ViT as one of its available architectures. I would appreciate it if you would consider mentioning flaim in the third-pary implementations section of the README.

Best,
Borna

Hi,

1. How can I train a semantic segmentation model using a custom dataset by fine-tuning a pretrained segmentation model?
2. As a follow-up question, are there pretrained segmentation models available to fine-tune it using a custom dataset?

Thank you.

Here https://github.com/NVlabs/GCVit/blob/main/models/gc_vit.py#L382 "x" is features in channel-last format [batch, height, width, channels]
You passing it thru some Conv2D layers (FeatExtract) which work with features in channel-first format ([batch, height, width, channels]). But instead of transpose operation you are making reshape with mixes channels and spatial axis.

I believe most architectural ideas in GCViT will boost performance, but i'd like to know reliable differences in metrices.
At the moment it is not possible due to all models trained with unusual window size.

More precisely Swin V1 with 224x222 input trained with window_size=7 for all stages.
GCViT at the same time uses [7, 7, 14, 7] window sizes. So at the third stage it has twice larger receptive field.

Due to this code branch https://github.com/NVlabs/GCVit/blob/main/models/gc_vit.py#L503 i believe you made some experiments.
Could you please publish their results, so we can understand what part of performance gained by architectural changes and what by receptive field expansion?

Dear all,
I have completed a new re-implementation of your work in TensorFlow2/Keras, including used timm modules : GCViT-TensorFlow.
It would be a pleasure to be included in "Third-party Implementations and Resources" section of your repo.
Anyway, congratulations for your work.

Thank you.
Best regards

I've attempting to train the 'tiny' model from scratch for verification but running into some problems. I hit an accuracy wall of sorts approaching 80% and the model should well pass that based on your results.

I am (obvioulsy) familiar with the train scripts, hparams in general, have trained quite a few related models to expected accuracy.

A few questions

• What is your train batch size? the paper mentions 'total batch size' is 128, that seems very low, when I read total I assume it means global (the sum across all train nodes). You also use 1e-3 learning rate which is very very high for 128 so perhaps you meant 128 is the local (per GPU) batch size?
• If 128 is the local batch size, what is the global? How many GPUS did you train with (N * 128)?
• Are there any sensitivies, epsilons, etc you ran into training that I may not have noticed a deviation from timm defaults?
• Your train command indicates you used EMA weight averaging, did you compare without?
• Your train command suggests you did NOT use AMP? Is that true, the accuracy / loss wall I'm hitting looks like it could be a floating point precision issue

what is the input shape ,B H W C or B C H W?

dataloader seems using only single process to handle image processing, num_workers waiting 1 process. gpu is fast ,but the cpu is slow. why is that? how to make the training faster?

Dear all,
I am training a model for emotion recognition and figure out this error in line line 521.
RuntimeError: shape '[32, 1, 49, 3, 32]' is invalid for input of size 49152.
The batch size = 32 and the model is GCVit-S
Can someone help to solve this issue?
Best regards

Hi!
I'm probably doing something fundamentally wrong when I try to train the model in a slightly different input size.
For example

python train.py --config configs/gc_vit_large.yml --data_dir '/mnt/…/my/custom/dataset/' --num-classes 17 --no-aug --crop-pct 1 --input-size 3 384 384 --experiment dif_InputSize

gives me the following error:

Traceback (most recent call last):
File "train.py", line 864, in
main()
File "train.py", line 648, in main
train_metrics = train_one_epoch(
File "train.py", line 722, in train_one_epoch
output = model(input)
File "/home/tan/env-GCVit/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/tan/GCVit/models/gc_vit.py", line 705, in forward
x = self.forward_features(x)
File "/home/tan/GCVit/models/gc_vit.py", line 696, in forward_features
x = level(x)
File "/home/tan/env-GCVit/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/tan/GCVit/models/gc_vit.py", line 600, in forward
q_global = self.q_global_gen(_to_channel_first(x))
File "/home/tan/env-GCVit/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1194, in _call_impl
return forward_call(*input, **kwargs)
File "/home/tan/GCVit/models/gc_vit.py", line 536, in forward
x = x.reshape(B, 1, self.N, self.num_heads, self.dim_head).permute(0, 1, 3, 2, 4)
RuntimeError: shape '[32, 1, 49, 6, 32]' is invalid for input of size 884736

If you could explain what I'm doing wrong here and how can I fix it, that would be great.

Is there a way to use this with arbitrary resolution?

model = timm.create_model('gcvit_tiny', pretrained=True)
sample = model(torch.randn(2, 3, 640, 640))

Error

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/tmp/ipykernel_9221/897640580.py in <module>
----> 1 sample = model(torch.randn(2, 3, 640, 640))

~/anaconda3/envs/yl/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1100         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102             return forward_call(*input, **kwargs)
   1103         # Do not call functions when jit is used
   1104         full_backward_hooks, non_full_backward_hooks = [], []

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in forward(self, x)
    524 
    525     def forward(self, x: torch.Tensor) -> torch.Tensor:
--> 526         x = self.forward_features(x)
    527         x = self.forward_head(x)
    528         return x

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in forward_features(self, x)
    517     def forward_features(self, x: torch.Tensor) -> torch.Tensor:
    518         x = self.stem(x)
--> 519         x = self.stages(x)
    520         return x
    521 

~/anaconda3/envs/yl/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1100         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102             return forward_call(*input, **kwargs)
   1103         # Do not call functions when jit is used
   1104         full_backward_hooks, non_full_backward_hooks = [], []

~/anaconda3/envs/yl/lib/python3.7/site-packages/torch/nn/modules/container.py in forward(self, input)
    139     def forward(self, input):
    140         for module in self:
--> 141             input = module(input)
    142         return input
    143 

~/anaconda3/envs/yl/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1100         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102             return forward_call(*input, **kwargs)
   1103         # Do not call functions when jit is used
   1104         full_backward_hooks, non_full_backward_hooks = [], []

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in forward(self, x)
    388                 x = checkpoint.checkpoint(blk, x)
    389             else:
--> 390                 x = blk(x, global_query)
    391         x = self.norm(x)
    392         x = x.permute(0, 3, 1, 2).contiguous()  # back to NCHW

~/anaconda3/envs/yl/lib/python3.7/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
   1100         if not (self._backward_hooks or self._forward_hooks or self._forward_pre_hooks or _global_backward_hooks
   1101                 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1102             return forward_call(*input, **kwargs)
   1103         # Do not call functions when jit is used
   1104         full_backward_hooks, non_full_backward_hooks = [], []

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in forward(self, x, q_global)
    310 
    311     def forward(self, x, q_global: Optional[torch.Tensor] = None):
--> 312         x = x + self.drop_path1(self.ls1(self._window_attn(self.norm1(x), q_global)))
    313         x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
    314         return x

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in _window_attn(self, x, q_global)
    303     def _window_attn(self, x, q_global: Optional[torch.Tensor] = None):
    304         B, H, W, C = x.shape
--> 305         x_win = window_partition(x, self.window_size)
    306         x_win = x_win.view(-1, self.window_size[0] * self.window_size[1], C)
    307         attn_win = self.attn(x_win, q_global)

~/anaconda3/envs/yl/lib/python3.7/site-packages/timm/models/gcvit.py in window_partition(x, window_size)
    235 def window_partition(x, window_size: Tuple[int, int]):
    236     B, H, W, C = x.shape
--> 237     x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
    238     windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
    239     return windows

RuntimeError: shape '[2, 22, 7, 22, 7, 64]' is invalid for input of size 3276800

Hi,

When I try to train the model by myself, I have several questions about the ImageNet dataset. For ImageNet-1k, the validation data from official website is not a format of:

  imagenet
  ├── val
      ├── class1
      │   ├── img4.jpeg
      │   ├── img5.jpeg
      │   └── ...
      ├── class2
      │   ├── img6.jpeg
      │   └── ...
      └── ...
 

but the validation tar is formatted as:

  imagenet
  ├── val
      ├──  img1.jpeg
      ├──  img2.jpeg
      ├──  img3.jpeg
      └── ...

How to prepare the validation set into the desired format?


In addition, if the train set contains class0~class900, val set contains class900~1000. Should I create empty folders for train directory and test directory (so as to make each of them has 1000 folders) ?

Thanks a lot!

Model code has an issue in adapting global query for attention estimation: https://github.com/NVlabs/GCVit/blob/main/models/gc_vit.py#L223
Repeating along batch dimension results not in [1, 1, 1, 2, 2, 2, 3, 3, 3] bached element order, but in [1, 2, 3, 1, 2, 3, 1, 2, 3].
In the same time after window_partition key and value "windows" are odered as [1, 1, 1, 2, 2, 2, 3, 3, 3].

You can verify this by evaluating imagenet with batch size 1 and 16. Results would be different.
In my tests results for batch size 1 are always equals. And results for batch size 16 will be always different if we add stohasticy (shuffle batch items and labels simultaneously).

During training this bug leads to partly wrong attention matrix estimation (because around "B_//B - 1" portion of channels will try to attend to wrong batch elements).
I suppose that fixing this bug and finetuning existing checkpoints should get even better metrics.

Hi!
Thank you for your great work.

I am getting an error that the output shape for each stage in GC ViT is incorrect. I am not sure where the error is coming from, but I would appreciate your help in debugging it.

I wrote a print statement to print out the shape of x in the forward_features method of the GCViT class. The code is as follows:
def forward_features(self, x):
x = self.patch_embed(x)
x = self.pos_drop(x)
for level in self.levels:
x = level(x)
print(x.shape)
......

Then I run the following code:
model=gc_vit_xxtiny()
inp = torch.rand(1,3, 224, 224)
res=model(inp)

The results are as follows:
torch.Size([1, 28, 28, 128])
torch.Size([1, 14, 14, 256])
torch.Size([1, 7, 7, 512])
torch.Size([1, 7, 7, 512])

I am not sure if I made a mistake, but the results I am getting are different from what you mentioned in the paper.
Thanks

I was trying to implement GCVit as backbone for YOLO one stage detector. However, one thing I noticed is that YOLO architectures uses only three feature maps from its backbone before it passes it to the neck architecture. Howevr, ViT like GCVit and Swin-T have 4 feature maps. How do I properly use these feature maps without loss of vital informations?

Hi thanks for sharing the great work.

I am kindly wondering whether you are aware of our recent work that also proposes to use serial global-local attention? I am grateful if you can add a discussion if you happen to update it: MaxViT: Multi-Axis Vision Transformer

Hello @ahatamiz
Thank you very much for your work.
I find that your visualization of global attention is very attractive in understanding your work. I would like to ask about the details of your visualization of global attention.

Thanks in advance!

Hi ,
Could you please share the pretrained model of GC_ViT-B_384 in IN1-K ?

Hi!
Thank you for your great work.

I am getting an error that the output shape for each stage in GC ViT is incorrect. I am not sure where the error is coming from, but I would appreciate your help in debugging it.

I wrote a print statement to print out the shape of x in the forward_features method of the GCViT class. The code is as follows:
def forward_features(self, x):
x = self.patch_embed(x)
x = self.pos_drop(x)
for level in self.levels:
x = level(x)
print(x.shape)
......

Then I run the following code:
model=gc_vit_xxtiny()
inp = torch.rand(1,3, 224, 224)
res=model(inp)

The results are as follows:
torch.Size([1, 28, 28, 128])
torch.Size([1, 14, 14, 256])
torch.Size([1, 7, 7, 512])
torch.Size([1, 7, 7, 512])

I am not sure if I made a mistake, but the results I am getting are different from what you mentioned in the paper.
Thanks

Model code has an issue in adapting global query for attention estimation: https://github.com/NVlabs/GCVit/blob/main/models/gc_vit.py#L223
Heads in keys and values and local query produced by splitting channels and head dimension transposition to "batch" dimension (in terms of matmul).
Heads in global query produced by shifting spatial size https://github.com/NVlabs/GCVit/blob/main/models/gc_vit.py#L224 because of missing transpose op that moves head dimension.

As far as i understand this results in not full query-to-key spatial interactions and adds some query-key channel shifting.