It looks like there are some spelling errors in your command line args. Please can you send it again with code formatting please so we can read it better.

sorry Let me rephrase
I trained this network on the CoNSeP dataset and got the training weight of each epoch (net_epoch=5.tar...(As shown above)) ,How should I use these weights in run_infer.py. I tried to run the run_infer.py file like this: python run_infer.py --gpu="0,1" --nr_types=6 --type_info_path="type_info.json" --batch_size=64 --model_mode=oright --model_path="./logs/01/net_epoch=50. tar "--nr_inference_workers=8 --nr_post_proc_workers=16 wsi- input_dir = ". / the Dataset/CoNSeP/Test/Images/" - output_dir = ". / the Dataset/sample_tiles/Mr Pred/" - save_thumb - save_mask. However, the following error is displayed
/home/dc2-user/anaconda3/envs/pytorch160/bin/python /data0/hover_net-master/run_infer.py --gpu=0,1 --nr_types=6 --type_info_path=type_info.json --batch_size=64 --model_mode=oright --model_path=./logs/01/net_epoch=50.tar --nr_inference_workers=8 --nr_post_proc_workers=16 wsi --input_dir=./Dataset/CoNSeP/Test/Images/ --output_dir=./Dataset/sample_tiles/pred/ --save_thumb --save_mask
|2021-01-19|08:35:52.802| [INFO] .... Detect #GPUS: 2
WARNING: Detect checkpoint saved in data-parallel mode. Converting saved model to single GPU mode.
Traceback (most recent call last):
File "/data0/hover_net-master/run_infer.py", line 181, in
infer = InferManager(**method_args)
File "/data0/hover_net-master/infer/base.py", line 27, in init
self.__load_model()
File "/data0/hover_net-master/infer/base.py", line 68, in __load_model
net.load_state_dict(saved_state_dict, strict=True)
File "/home/dc2-user/anaconda3/envs/pytorch160/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1045, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for HoVerNet:
size mismatch for decoder.tp.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).
size mismatch for decoder.tp.u0.conv.weight: copying a param with shape torch.Size([5, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([6, 64, 1, 1]).
size mismatch for decoder.tp.u0.conv.bias: copying a param with shape torch.Size([5]) from checkpoint, the shape in current model is torch.Size([6]).
size mismatch for decoder.np.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.np.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.np.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).
size mismatch for decoder.hv.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).

Process finished with exit code 1

How should I set the path to --model_path? Or I should have made those arguments to run_infer.py？？？

I am copying your command line arguments as it makes it easier to interpret and refer to:

python run_infer.py --gpu="0,1" --nr_types=6 --type_info_path="type_info.json" --batch_size=64 --model_mode=oright --model_path="./logs/01/net_epoch=50. tar "--nr_inference_workers=8 --nr_post_proc_workers=16 wsi- input_dir = ". / the Dataset/CoNSeP/Test/Images/" - output_dir = ". / the Dataset/sample_tiles/Mr Pred/" - save_thumb - save_mask

Your input is the CoNSeP dataset which are standard image tiles, not WSIs. Therefore you should not be using WSI mode. Also, you have selected an invalid model mode - this should be fast or original. You have used oright. @vqdang has added an assert in this commit to catch this now.

Please refer to this as an example on how to use the command line arguments with tile inference mode. You will also see that you have not got the correct number of nuclei types for the CoNSeP dataset - this should be 5.

I retrained the pytroch version network to run the same process as the original TensorFlow version, In addition, I also changed the way of extracting slices from win_size = [540,540] step_size = [164, 164] to [540,540] [80,80] during the training, and the weights I obtained for each round of training are shown in the following figure，so I am not using WSI mode., I used the tile model.

I used invalid model mode-oright, but this error is due to a network structure mismatch. I used the same network for training and testing, but the weight of the model used in the testing stage was the weight saved in the training stage. Why did it prompt me about the problem of network structure mismatch.
python run_infer.py --gpu="0,1" --nr_types=6 --type_info_path="type_info.json" --batch_size=64 --model_mode=oright --model_path="./logs/01/net_epoch=net_epoch=5. tar "--nr_inference_workers=8 --nr_post_proc_workers=16 wsi- input_dir = ". / the Dataset/CoNSeP/Test/Images/" - output_dir = ". / the Dataset/sample_tiles/Mr Pred/" - save_thumb - save_mask

In the test phase, should the path of this model point to the weights of a certain round that have been saved during the training phase, such as the weights of round 5 in the figure above.

2021-01-19|08:35:52.802| [INFO] .... Detect #GPUS: 2
WARNING: Detect checkpoint saved in data-parallel mode. Converting saved model to single GPU mode.
Traceback (most recent call last):
File "/data0/hover_net-master/run_infer.py", line 181, in
infer = InferManager(**method_args)
File "/data0/hover_net-master/infer/base.py", line 27, in init
self.__load_model()
File "/data0/hover_net-master/infer/base.py", line 68, in __load_model
net.load_state_dict(saved_state_dict, strict=True)
File "/home/dc2-user/anaconda3/envs/pytorch160/lib/python3.6/site-packages/torch/nn/modules/module.py", line 1045, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for HoVerNet:
size mismatch for decoder.tp.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.tp.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).
size mismatch for decoder.tp.u0.conv.weight: copying a param with shape torch.Size([5, 64, 1, 1]) from checkpoint, the shape in current model is torch.Size([6, 64, 1, 1]).
size mismatch for decoder.tp.u0.conv.bias: copying a param with shape torch.Size([5]) from checkpoint, the shape in current model is torch.Size([6]).
size mismatch for decoder.np.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.np.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.np.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.np.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).
size mismatch for decoder.hv.u3.conva.weight: copying a param with shape torch.Size([256, 1024, 5, 5]) from checkpoint, the shape in current model is torch.Size([256, 1024, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.4.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.5.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.6.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u3.dense.units.7.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.conva.weight: copying a param with shape torch.Size([128, 512, 5, 5]) from checkpoint, the shape in current model is torch.Size([128, 512, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.0.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.1.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.2.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u2.dense.units.3.conv2.weight: copying a param with shape torch.Size([32, 32, 5, 5]) from checkpoint, the shape in current model is torch.Size([32, 32, 3, 3]).
size mismatch for decoder.hv.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).

@zerodohero For everyone sake, please please read what you input as the command

python run_infer.py --gpu="0,1" --nr_types=6 --type_info_path="type_info.json" --batch_size=64 --model_mode=oright
-------------------------------------------------------------------------------------------------------------------------------^^^

Also, the internal pytorch error clearly indicate what wrong with it, please read it try to understand it.

size mismatch for decoder.hv.u1.conva.weight: copying a param with shape torch.Size([64, 256, 5, 5]) from checkpoint, the shape in current model is torch.Size([64, 256, 3, 3]).

Only original use 5x5 kernel, fast use 3x3. You are loading a orignal checkpoint into fast model because of wrong input in the command.

I close the thread as this doesnt seem to go anywhere when you can't seem to understand what we said. As far as the problem goes, the cause is in the input format and we already pointed it out.