ViT-pytorch

ViT-pytorch copied to clipboard

--gradient_accumulation_steps

2

Hello, I would like to ask you about this step in README, The default batch size is 512. When GPU memory is insufficient, you can proceed with training by adjusting...

zhangchao77

Can you update how to use imagenet21k_R50+ViT-L_32.npz?

3

Hi, @jeonsworld The imagenet21k_R50+ViT-L_32.npz model has been released. Can you update how to use imagenet21k_R50+ViT-L_32.npz? I tried it on current your source, but I kept having problems. thanks

peternara

I added data enhancement methods such as translation, rotation, and scaling to the test data sample, hoping to use the inductive bias of CNN, but R50+ViT did not achieve the...

haodonga

alirezaabdollahpour

Which GPU did you use?

5

Sorry, there is training-time show in your experiment. I wonder which GPU did you use, and how many of them?

nessessence

Why the kernel is normalized in StdConv2d?

1

I noticed that you used class StdConv2d(nn.Conv2d): def forward(self, x): w = self.weight v, m = torch.var_mean(w, dim=[1, 2, 3], keepdim=True, unbiased=False) w = (w - m) / torch.sqrt(v +...

xychenunc

When use my custom dataset, which contains 6 classes, so I modified the data_utils.py, and change the 'num_classes = 6' in train.py. But I got these errors: Training (X /...

superxiaoying

vit = VisionTransformer(CONFIGS['R50-ViT-B_16'], zero_head=False, img_size=200) leads to "float division by zero" exception: --------------------------------------------------------------------------- ZeroDivisionError Traceback (most recent call last) in ----> 1 vit = VisionTransformer(CONFIGS['R50-ViT-B_16'], zero_head=False, img_size=200) ViT-pytorch/models/modeling.py in __init__(self,...

pgagarinov

train set, val set and test set

1

Hi, perhaps, I didn't understand your code clearly, but I think you didn't split the dataset into train set, val set and test set. You straightly validate and test your...

Anonymity111222