Visual-Transformers

Unofficial implimentation of Visual Transformers: Token-based Image Representation and Processing for Computer Vision paper.

Usage:

python main.py task_mode learning_mode data --model --weights, where:

• task_mode: classification or semantic_segmentation for corresponding task
• learning_mode: train to train --model from scratch, test to validate --model with --weights on validation data.
• data: path to dataset, in case of classification should be path to image net, in case of semantic segmentation to coco.
• --model:
  ○ classification: ResNet18 or VT_ResNet18 (will be used by default).
  ○ semantic segmentation: PanopticFPN or VT_FPN (will be used by default).
• --weights must be provided if learning_mode equals to test, won't be used in train mode.
• --from_pretrained uses to continue training from some point, should be state_dict that contains model_state_dict, optimizer_state_dict and epoch.

Results:

• final metrics and losses after 15 and 5 epochs of classification and semantic segmentation respectively:

ResNet18 VT-ResNet18 Training accuracy 0.664675 0.672889 Validation accuracy 0.691541 0.696929 Training loss 1.312150 1.249382 Validation loss 1.173559 1.114401

Panoptic FPN VT-FPN Training mIOU 8.0968 7.0343 Validation mIOU 4.3148 3.2351 Training loss 2.044084 2.068598 Validation loss 2.101253 2.120928

• loss and metric curves of classification and semantic segmentation:

cross entropy loss	accuracy

pixel-wise cross entropy loss	mIOU

• Efficiency and parameters

	Params (M)	FLOPs (M)	Forward-backward pass (s)
ResNet18	11.2	822	0.016
VT-ResNet18	12.7	543	0.02
Panoptic FPN	16.4	67412	0.08
VT-FPN	40.3	110019	0.062

Weights:

• classification: ResNet18, VT-ResNet18
• semantic segmentation: Panoptic FPN, VT-FPN