ResNets-ImageNet (Ported Weights)

All the weights available here have been ported either from timm or from torchvision, so thanks to @rwightman and @pytorch, respectively ⭐

📌 Since the architecture of the ResNets v1 is the same that the ones trained by @rwightman and @pytorch, the process of porting the weights has consisted just of renaming the layers that differ between both implementations defined in the state_dict file.

ResNets-CIFAR10

All the weights included in this release have been trained on a single GPU NVIDIA GeForce GTX 1070 using a batch size of 128, trained with the whole CIFAR10 train dataset, and tested with the whole test dataset. All the training steps defined in the original ResNet paper have been reproduced, besides the train/val split of 45k/5k, which has not been used.

For more information on the training process and the implementation please check README.md.

Architecture	Option	# Layers	# Parameters	Train Error	Test Error	Train Time	Size	W&B Run
ResNet20	A	20	0.269M	0.00856	0.0828	55m 4s	1.07MB	URL
ResNet20	B	22	0.272M	0.00646	0.0811	55m 9s	1.09MB	URL
ResNet32	A	32	0.464M	0.02876	0.0956	1h 9m 31s	1.84MB	URL
ResNet32	B	34	0.466M	0.00242	0.0725	1h 9m 14s	1.86MB	URL
ResNet44	A	44	0.658M	0.00151	0.0714	1h 24m 31s	2.61MB	URL
ResNet44	B	46	0.658M	0.0038	0.0888	1h 24m 39s	2.61MB	URL

For more information about the training process, please check the W&B Monitoring Results.