About 'MSR-VTT' dataset.

[q-rong]

Hi,

When I was training the model with the MSR-VTT dataset, I encountered the following error. I lack the total.json file, can you provide it?

IOError: [Errno 2] No such file or directory: 'data/MSR-VTT/metadata/total.json'

Thank you very much.

Best wish.

[hobincar]

Hi,

I uploaded the file at 'data/MSR-VTT/metadata/total.json'. Thank you for reporting the missing file :)

[q-rong]

Hi,

I noticed that the MSR-VTT dataset you used was only 7,927 videos. Among them, the train set was 5,175, the val set was 398, and the test set was 2,354.

However, there are 10,000 MSR-VTT-10K datasets mentioned in this paper. Among them, there were 6513 train set, 497 val set and 2990 test set.

This seems inconsistent. Is it because of inconsistent data sets that the results of the model are lower?

Why don't you use the MSR-VTT-10K dataset?

[hobincar]

You might have seen the lines that specify the number of a train, a validation, and a test dataset, respectively. Those numbers were true when I manually download the MSR-VTT dataset with a script, and some videos were missing as they were not available anymore. However, after I found this link that provides the full MSR-VTT dataset I could train the RecNet with all videos.

Those (legacy) numbers do not affect the operation of the model and just there only for giving us information. As it is no longer true, I'll remove it to avoid confusion.

[q-rong]

Hi,

I noticed that when training the model on the MSVD dataset, the learning rate was allowed to change after the 20th epochs.

However, the MSVD dataset was set to train 50 epochs, while the MSR-VTT-10K dataset only trained 30 epochs.

May I ask, when training the model on the MSR-VTT-10K dataset, is the learning rate change also allowed at the beginning of the 20th epochs?

[hobincar]

Yes, that's what I intended. However, you can find a better learning rate policy since I did not carefully choose it. The 20th is what I found to be good for the MSVD dataset, and I just applied the same to the MSR-VTT dataset.

[maoao686868]

@q-rong Hi ,when you test run.py with the inception.v4 features from MSVD dataset, did you meet this problem？
Traceback (most recent call last):
File "run.py", line 96, in
run('/root/Workspace/rn/checkpoint/RecNet-global_MSVD(1).ckpt')
File "run.py", line 44, in run
decoder.load_state_dict(checkpoint['decoder'])
File "/root/anaconda3/envs/rn/lib/python3.6/site-packages/torch/nn/modules/module.py", line 777, in load_state_dict
self.class.name, "\n\t".join(error_msgs)))
RuntimeError: Error(s) in loading state_dict for Decoder:
size mismatch for embedding.weight: copying a param with shape torch.Size([13380, 468]) from checkpoint, the shape in current model is torch.Size([13475, 468]).
size mismatch for out.weight: copying a param with shape torch.Size([13380, 512]) from checkpoint, the shape in current model is torch.Size([13475, 512]).
size mismatch for out.bias: copying a param with shape torch.Size([13380]) from checkpoint, the shape in current model is torch.Size([13475]).

[hobincar]

Seems to be a duplicate of #32.