Closed Kerry-zzx closed 1 year ago
The newest result after training 1300 epochs:
Hello @Kerry-zzx! After your issue, I cloned the repo and started training the network from scratch. The experiments are going correctly. Therefore, I think there isn't a bug in the code.
This is a qualitative result after 300 epochs:
700 epochs:
1000 epochs:
1300 epochs:
1600 epochs:
This model is trained on an NVIDIA 2080 Ti for 19 hours.
Please double-check that you downloaded the correct files and there are no errors. This is the output of the sha1sum
command.
04de5a8c292ae3fe8e4911e78624e1c5767b9aa1 IAM-32.pickle
c884705b413c9d6cd415b80b0ae2ae43c495385b resnet_18_pretrained.pth
17ec41d538dada983c46ce65bdb013f799e61da3 unifont.pickle
20c50bfb8da324bc8aaf7d84714e7b40d46b8daa english_words.txt
I remind you that for training, you have to use the pretrained checkpoint of the feature encoder (default):
python train.py --feat_model_path files/resnet_18_pretrained.pth
Hi,
I am glad to hear that you have been able to successfully train the network from scratch without encountering any issues. This suggests that there may not be a bug in the code itself, as you mentioned.
In light of your successful training results, I took your advice and retrained the experiment on my end as well. I re-downloaded all the pre-trained files. I am pleased to inform you that this time, the results have shown considerable improvement. It appears that the previous discrepancies were likely due to the download files.
I apologize for any confusion or inconvenience caused by my previous observations. It seems that the issue has been resolved, and the code is functioning as expected. I am grateful for your guidance and assistance throughout this process.
Hi, Vittorio Pippi,
After implementing the revisions you made to the code, I attempted to train the model using the updated version. However, I regret to inform you that I encountered difficulties and the training process did not yield the expected results.
This is the result after training 300 epochs:
Despite carefully incorporating the changes you suggested, the model's performance during training remained unsatisfactory. I wanted to reach out to you once again to request your assistance in troubleshooting this issue. I am eager to understand the root cause of the problem and identify any potential steps or adjustments necessary to rectify it.