Add Support for Final Validation After Fine-Tuning #116
Conversation
…roved compatibility
…bility with PyTorch
…tocast usage in trainer
…patibility with PyTorch
…bility with PyTorch
…ng in strip_optimizer
…utocast usage in training loop
…zation before activation
…patibility with PyTorch
…bility with PyTorch
|
@Geo99pro Hello! Thanks for your contribution, I will review your PR and give responses within this week (sorry for that cause I am kind of busy) |
Hi sir @JulioZhao97, There is no problem, i can understand it. I make myself available anytime. |
…rity and compatibility
…atibility with PyTorch
…compatibility with PyTorch
| if hasattr(self.dcv, "bn") and self.dcv.bn is not None: | ||
| bn = self.dcv.bn | ||
| x = bn(F.conv2d(x, weight, stride=1, padding=padding, dilation=dilation)) | ||
| else: | ||
| x = F.conv2d(x, weight, stride=1, padding=padding, dilation=dilation) | ||
|
|
||
| return act(x) # Apply activation function |
There was a problem hiding this comment.
Sir @JulioZhao97 I did this because while debugging, I noticed that in: https://github.com/opendatalab/DocLayout-YOLO/blob/main/doclayout_yolo/nn/tasks.py more specifically in the fuse function (line 180) you merge the Conv2d() and the BatchNorm2d() whose purpose is to speed up inference. Since you delete after the bn, the model undergoes this alteration. I could be wrong, but it will be a pleasure to learn from you.
…line Python installation
…ile by simplifying the COPY and WORKDIR instructions. Enable the use of requirements file to install only the necessary library instead of the biblio.
…ithub.com/Geo99pro/DocLayout-YOLO into feature/final-validation-after-finetuning
There was a problem hiding this comment.
Hello sir @JulioZhao97, I've added a requirements file to better manage the dependencies that will make it easy to use docker.
2 participants
Hi everyone,
I hope you're all doing well!
First of all, I would like to congratulate you all on the amazing work you've done and made available to the community. It’s truly impressive!
@JulioZhao97, I encountered an issue when trying to fine-tune the model on custom data. Specifically, the model was unable to perform the final validation step after training.
This issue seems related to previous reports from other users. For example, the issue mentioned in #72.
After some debugging sessions, I made updates to the code to address this problem.
🆕 I've also added a Dockerfile, in case anyone wants to use it.
I hope my contribution is helpful and adds value to your excellent project.
Looking forward to your feedback!
Best regards,