Hello! I have some questions when I test you code. The error is that provided spynet_sintel_final-3d2a1287.pth model does not match the network. Can you provide me the true model file? Thanks ！！！ I will be appreciated if you can reply to me as soon as you can. Thanks a lot.

Could you please open source the final trained model files? Since 1200k iter training is so resource comsuming, it is not so possible for me to train the model starting from scratch

A bug hard to notice: in 'basicsr/models/crop_validation.py', the beginning of method 'forward_crop', two assert statements are used: assert lq_size == 64 or 48, "Default patch size of LR images during training and validation should be {}.".format(lq_size) assert overlap == 16 or 12, "Default overlap of patches during validation should be {}.".format(overlap) you should insteadly use assert lq_size == 64 or lq_size ==48, "Default patch size of LR images during training and validation should be {}.".format(lq_size) assert overlap == 16 or lq_size ==12, "Default overlap of patches during validation should be {}.".format(overlap) because origin statement will be explained as assert (lq_size==64) or (48), which is always True

Hello, thank you for contributing this code and a method of applying transformer in the field of video. I have some problems with your paper and code. I'm just getting started, so I have more doubts

1. I see that your optical flow distortion is estimated by using the input original image to distort the features, which is distorted five times, and the optical flow is not specially supervised. Why use optical flow in feedforward? Isn't self attention a good fusion of features, and isn't there an error in optical flow? Will it affect performance

2. For patch and window size, you set patch to 8 in the project × 8. Is this patch the bigger the better or the smaller the better? If the patch is too large, there will be less local information. And whether it's better to set the window larger. You set the window to 64 mainly for the trade-off of calculation, right？

3. As for the number of layers of transformer, is the more layers the better? Is 5 layers the best choice?

Some questions ask you mainly because the laboratory computing resources are limited and there are only two 3090 cards, so I can't verify them one by one because the training speed is too slow.

I look forward to your reply very much. Thank you very much

Hello! It is really a god job. I have a question about the input size of network in training and testing. I noticed that the network used nn.LayerNorm([num_feat, feat_size, feat_size]) in the transformer block, and the (h,w) of input is (64,64) in training. Does it mean that the input of (h,w) should be (64,64) in testing as well? As the nn.LayerNorm([num_feat, 64, 64]) in saved model in training is fixed. Maybe it's an easy question, it makes me confused. I will be appreciated if you can reply to me as soon as you can. Thanks a lot.

Hello! I have some questions when I watch your code. I find that you use the same result of optical flows to warp those feature map for five times. Is that your original idea or just a mistake? For all the layers, flows are the same.

for attn, ff in self.layers: x = attn(x) x = ff(x, lrs=lrs, flows=flows) return x

in vsrTransformer_arch.py / class Transformer / function forward

Here is another problem. No matter "lq_size" is equl to 64 or others, these "assert" will always be True.

assert lq_size == 64 or 48, "Default patch size of LR images during training and validation should be {}.".format(lq_size) assert overlap == 16 or 12, "Default overlap of patches during validation should be {}.".format(overlap)

in crop_validation.py / function forward_crop

I will be appreciated if you can reply to me as soon as you can. Thanks a lot.