added sample code for fasterrcnn_resnet50_fpn (optional) #796
Conversation
|
Deploy preview for pytorch-tutorials-preview ready! Built with commit fe6080f https://deploy-preview-796--pytorch-tutorials-preview.netlify.com |
| --------------------------- | ||
|
|
||
| Before iterating over the dataset, it's always good to see what the model | ||
| expects during training and inference time with random tensors. |
There was a problem hiding this comment.
This is something we should be careful about: if we have batch normalization layers in the model (which is not the case for detection models because we freeze the batch norm layers, but happens in classification), feeding randomly-initialized data will affect the batch norm statistics, leading to potentially unexpected results
There was a problem hiding this comment.
I've now changed the code from using random tensors to sampled data from the actual dataset. This will fix the batch norm stats issue. This is not what I had intended earlier, I just wanted to show what the model actually requires and how one can reproduce the output from model by passing random data because most users would not peek inside engine.py to learn what's happening initially. But you're right, batch norm stats would be calculated incorrectly as per previous commit. I guess this will be fine.
| for i in range(len(images)): | ||
| d = {} | ||
| d['boxes'] = boxes[i] | ||
| d['labels'] = labels[i].type(torch.int64) |
There was a problem hiding this comment.
The use of .type is not encouraged (it's legacy), and instead we should use .to. See my fixes (including lint fixes) in pytorch/vision#1713
There was a problem hiding this comment.
Thanks for pointing this out. I've fixed it.
|
Let me know, if this seems fine. |
|
Thanks, how can I fix the tests ? |
|
@brianjo Can you please merge it ? |
Hi @prajjwal1, @yf225 and I are working on fixing the build right now and I'll rebase and merge this one when we get that going. Thanks! |
|
@yf225 , Can you please have a look at this ? Can this be merged now. I'll close the issue in torchvision if it's merged. |
|
@prajjwal1 I believe all build jobs passed on this PR. @brianjo Do you think we can merge this PR? |
Will do. Still working through a Netlify issue, but this looks good. I'll rebase it first, but I'll get it in when it passes again. |
|
Thanks ! |
added sample code for fasterrcnn_resnet50_fpn (optional)
Discussed here.