step 1: generate the rate-distortion curves of weights

• run -> python ./1_generate_rate_distortion_curves_weights/mobilenet_v1_original.py para_layer para_dz para_quant_level id_gpu
• para_layer: the quantization layer
• para_dz: the dead zone ratio
• para_quant_level: the number of quantization levels
• id_gpu: which GPU to run

step 2: generate the rate-distortion curves of activations

• run -> python ./1_generate_rate_distortion_curves_activations/mobilenet_v1_original.py para_layer para_quant_level id_gpu
• para_layer: the quantization layer
• para_quant_level: the number of quantization levels
• id_gpu: which GPU to run

step 3: solve optimal bit allocation under paredo contidion:

• compile -> pareto_condition.cpp
• run -> pareto_condition

step 4: perform inference of unequal bit allocation framework on imagenet: CUDA_VISIBLE_DEVICES=$id_gpu python ./slim/nets/mobilenet_v1_eval.py
--checkpoint_dir=/home/wangzhe/Documents/exp/exp_2019_4_p2/4_23_2_eval_mobilenet_v1_testing_2/mobilenet_v1_1.0_224/mobilenet_v1_1.0_224.ckpt
--dataset_dir=$dataset_dir
--bit_allocation=bit_allocations/bit_allocations_$i.txt
--dir_weight_codebooks=/home/wangzhe/Documents/exp/mobilenet_v1_data/codebooks_mobilenet_v1_wei/
--dir_activation_codebooks=/home/wangzhe/Documents/exp/mobilenet_v1_data/codebooks_mobilenet_v1_activations_uni
|& tee logs_results_$i.txt