Movatterモバイル変換

[0]ホーム
URL:

Skip to content

Navigation Menu

Sign in
Appearance settings

Search code, repositories, users, issues, pull requests...

Provide feedback

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly

 Sign up
Appearance settings

Warmup on uneven last-batch-size invalidate.py#2243

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to ourterms of service andprivacy statement. We’ll occasionally send you account related emails.

Already on GitHub?Sign in to your account

Open
eqy wants to merge1 commit intohuggingface:main
base:main
Choose a base branch
Loading
fromeqy:patch-1

Conversation

eqy
Copy link

@eqyeqy commentedJul 26, 2024

In the spirit of warming up for JIT compilation, add a warmup iteration in case the very last batch has a different size that may unwittingly trigger recompilation

@eqy
Copy link
Author

eqy commentedAug 29, 2024

This also has the effect of including autotuning time for the last batch if it is uneven astorch.backends.cudnn.benchmark = True is set

@rwightman
Copy link
Collaborator

@eqy there's a small issue here, which is a bit to explain which is why this sat...

Validation script should work in streaming mode without a defined length, I believe it used to work but I actually made it a bit too strict (I need to fix). So the partial batch check must catch the situation where dataset length isn't defined.

Comment lines 307/308 in reader_wds.py

        #if not self.num_samples:        #    raise RuntimeError(f'Invalid split definition, num_samples not specified.')

..and then below should work:

python validate.py --data-dir 'pipe:curl -s -H "Authorization: Bearer $HFT" -f -L https://huggingface.co/datasets/timm/imagenet-1k-wds/resolve/main/' --dataset wds/ --split 'imagenet1k-validation-{00..10}.tar'Validating in float32. AMP not enabled.Loading pretrained weights from Hugging Face hub (timm/dpn92.mx_in1k)[timm/dpn92.mx_in1k] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.Model dpn92 created, param count: 37668392Data processing configuration for current model + dataset:input_size: (3, 224, 224)interpolation: bicubicmean: (0.48627450980392156, 0.4588235294117647, 0.40784313725490196)std: (0.23482446870963955, 0.23482446870963955, 0.23482446870963955)crop_pct: 0.875crop_mode: centerTest: [   0/0]  Time: 2.018s (2.018s,  126.88/s)  Loss:  0.9466 (0.9466)  Acc@1:  77.734 ( 77.734)  Acc@5:  94.141 ( 94.141)Test: [  10/0]  Time: 0.329s (0.507s,  504.63/s)  Loss:  0.7221 (0.8023)  Acc@1:  80.469 ( 79.936)  Acc@5:  97.266 ( 95.241)Test: [  20/0]  Time: 0.330s (0.438s,  584.59/s)  Loss:  0.7316 (0.8055)  Acc@1:  82.812 ( 80.283)  Acc@5:  95.703 ( 94.754)Test: [  30/0]  Time: 0.328s (0.409s,  626.20/s)  Loss:  0.6582 (0.7941)  Acc@1:  83.203 ( 80.262)  Acc@5:  95.703 ( 95.030) * Acc@1 80.079 (19.921) Acc@5 95.031 (4.969)--result{    "model": "dpn92",    "top1": 80.0791,    "top1_err": 19.9209,    "top5": 95.0314,    "top5_err": 4.9686,    "param_count": 37.67,    "img_size": 224,    "crop_pct": 0.875,    "interpolation": "bicubic"}

Sign up for freeto join this conversation on GitHub. Already have an account?Sign in to comment
Reviewers
No reviews
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

2 participants
@eqy@rwightman
[8]ページ先頭
©2009-2025 Movatter.jp