-
|
In the paper the accuracy for Resnet50，Resnet101，and Resnet152 are 80.4，81.5 and 82.0 respectively. However their results reported in the file https://github.com/huggingface/pytorch-image-models/blob/main/results/results-imagenet.csv are 80.37，81.94 and 82.82. There are clear gaps of two results for Resnet101，and Resnet152. Why? Is there any change for A1 training procedure? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
@iminfine 101 and 152 in the results did use a different recipe yes, they used a variation with higher aug+reg. This is the weights used for the last results, note the 'a1h' in the name: https://github.com/huggingface/pytorch-image-models/blob/0.6.x/timm/models/resnet.py#L75-L84 In 0.9, all of the variants are present and almost done the run to update the results files (which will include the .tag names to differentiate) https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/resnet.py#L704-L737 |
Beta Was this translation helpful? Give feedback.
@iminfine 101 and 152 in the results did use a different recipe yes, they used a variation with higher aug+reg.
This is the weights used for the last results, note the 'a1h' in the name: https://github.com/huggingface/pytorch-image-models/blob/0.6.x/timm/models/resnet.py#L75-L84
In 0.9, all of the variants are present and almost done the run to update the results files (which will include the .tag names to differentiate)
https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/resnet.py#L704-L737