Please refer to requirements.txt for the requisite python modules.

You are free to choose the experiment hyperparameters by modifying TSP/config.json, where all notation is as per the paper. To run the experiment, execute large_tsp_solver.py

Usage: python train.py --epochs_stage1 400 --epochs_stage2 400

Please install the WIDER_train dataset from Wider Face, and place it in CNN/faceboxes/WIDER_train/images.

Train script usage:

python train_resnet_adcless.py -b [batch_size] --num_workers [num_workers] --no_wandb --w_bits [weight_bits] --a_bits [activation_bits] --xbar_size [xbar_size] --epochs_stage1 100 --epochs_stage2 100

Test script usage: Install the PASCAL datasets for the VOC 2007, 2008, 2009, 2010, 2011, and add them to faceboxes_data/PASCAL/images. The provided img_list serves as a reference of the images expected to be downloaded.

python test_resnet_adcless.py -m [path_to_trained_checkpoint.pth] --w_bits [weight_bits] --a_bits [activation_bits] --xbar_size [xbar_size] --test_mode [adc_less/wa_quant]

IMPORTANT: Specify the correct category of the checkpoint you're evaluating on (weights/activations quantized only (after first train stage) or adc_less (after the two train stages))

After running the test script, for PASCAL evaluation, please refer to: PASCAL eval