NICE

This code implements NICE papper List of links to pre-trained models:

DeepISP

http://www.mediafire.com/file/15anb9x44nxkkke/best_checkpoint.pth.tar/file

ResNet18-CIFAR10

http://www.mediafire.com/file/legt0epbrw8qii3/model_best.pth.tar/file

ResNet18-ImageNet

http://www.mediafire.com/file/l5qbobd2mm5wry5/model_best.pth.tar/file

ResNet34-ImageNet

http://www.mediafire.com/file/et7mvajxamm8sup/model_best.pth.tar/file

ResNet50-ImageNet

http://www.mediafire.com/file/93f7s5h66d6n8z1/model_best.pth.tar/file

Running instructions

---

DeepISP

python3 deep_isp_main.py --batch_size=16 --resume= --quant=True --quant_bitwidth=4 --inject_noise=True --inject_act_noise=False --act_quant=True --act_bitwidth=8 --quant_epoch_step=6 --quant_start_stage=0 --epochs=500 --learning_rate=3e-5 --gpus 0 --set_gpu=True --stage_only_clamp=False

Params

• --seed N - Random seed

• --start-epoch N - manual epoch number (useful on restarts)

• --epochs EPOCHS Number of epochs to train.

• --batch_size BATCH_SIZE Number of epochs to train.

• --num_denoise_layers NUM_DENOISE_LAYERS num of layers.

• --learning_rate LEARNING_RATE,

• -lr LEARNING_RATE The learning rate.

• --decay DECAY

• -d DECAY Weight decay (L2 penalty).

• --gpus GPUS List of GPUs used for training - e.g 0,1,3. Remove to run on CPU

• --datapath DATAPATH Path to MSR-Demosaicing dataset

• --resume RESUME Path to checkpoint file

• --out_dir OUT_DIR Path to save model and results

• --quant_epoch_step QUANT_EPOCH_STEP quant_bitwidth.

• --num_workers NUM_WORKERS Num of workers for data.

• --quant_start_stage QUANT_START_STAGE Num of workers for data.

• --inject_noise INJECT_NOISE use preproccesing for the grad

• --show_test_result SHOW_TEST_RESULT show figures of test result

• --quant QUANT use preproccesing for the grad

• --quant_bitwidth QUANT_BITWIDTH quant_bitwidth.

• --inject_act_noise INJECT_ACT_NOISE use preproccesing for the grad

• --act_quant ACT_QUANT use preproccesing for the grad

• --act_bitwidth ACT_BITWIDTH quant_bitwidth.

• --step STEP amount of split the layer in quant.

• --set_gpu SET_GPU - show figures of test result (False to run on CPU)

• --adaptive_lr ADAPTIVE_LR show figures of test result

• --enable_decay ENABLE_DECAY decay_enable

• --weight_relu WEIGHT_RELU weight_relu

• --weight_grad_after_quant WEIGHT_GRAD_AFTER_QUANT weight_grad_after_quant

• --random_inject_noise RANDOM_INJECT_NOISE random_inject_noise

• --stage_only_clamp STAGE_ONLY_CLAMP stage_only_clamp

• --wrpn WRPN wrpn quantization

• --copy_statistics COPY_STATISTICS copy_statistics

• --quant_decay QUANT_DECAY quant decay.

• --val_part VAL_PART quant decay.

ResNet18 CIFAR10

python3 main.py --model resnet --depth 18 --bitwidth --act-bitwidth --step 21 --gpus 0 --epochs 120 -b 256 --dataset cifar10 --start-from-zero --resume --learning_rate=0.01 --quant_start_stage=0 --quant_epoch_step=3 --datapath --schedule 300

ResNet18 ImageNet

python main.py --model resnet --depth 18 --bitwidth --act-bitwidth --step 21 --schedule 42 110 -lr 1e-4 --decay 4e-5 --gamma 0.93451921456 --gpus 0,1 --epochs 120 -b 128 --dataset imagenet --datapath --resume --quant_start_stage=0 --quant_epoch_step=2 --no-quant-edges --noise_mask 0.05 --act_stats_batch_size 64

ResNet34 ImageNet

python main.py --model resnet --depth 34 --bitwidth --act-bitwidth --step 37 --schedule 73 110 -lr 1e-4 --decay 4e-5 --gamma 0.88296999554 --gpus 0,1 --epochs 120 -b 128 --dataset imagenet --datapath --resume --quant_start_stage=0 --quant_epoch_step=2 --no-quant-edges --noise_mask 0.05 --act_stats_batch_size 64

ResNet50 ImageNet

python main.py --model resnet --depth 50 --bitwidth --act-bitwidth --step 37 --schedule 83 110 -lr 1e-4 --decay 4e-5 --gamma 0.843190929--gpus 0,1 --epochs 120 -b 128 --dataset imagenet --datapath --resume --quant_start_stage=0 --quant_epoch_step=1.5 --no-quant-edges --noise_mask 0.05 --act_stats_batch_size 64