This repository is about using only autographer data to classify MART tasks. The backbone model applied in this repo is Efficientnet-B4.
The Data_Process notebook illustrated how to split dataset into train-val-test set. The result is stored in dataset folder. It also shown how to used the trained model to predict on a custom image or to evaluate on a train-val-test set.
The RUN folder stores the result after training a model including: model info, report at each epoch, tensorboard folder, and trained model.pth.tar
model.py: define model
mart_controller.py: define training and evaluating protocol
main.py: train model
run the command tensorboard --logdir= RUN_1/ to visualize the Loss, Accuracy, Learning rate of all experiments in RUN_1 folder
Currently, we merged both trainA and trainB into 1 dataset and split train-val-test subsets based on it. The proportion was 0.8, 0.1, 0.1 for each activity. The model achieved 90.5% accuracy on the validate set (about 170 images) using Adam optimizer based on RUN_0_Unfreeze. We first freeze the backbone (EfficientNet) and only train the transfer layers until converge then trained all layers. Download the pretrained model here