This repo contains the official Pytorch reimplementation of the paper "NetAdapt: Platform-Aware Neural Network Adaptation for Mobile Applications" [paper] [project]. The results in the paper were generated by the Tensorflow implementation from Google AI.


The code base is tested with the following setting:

  1. Python 3.7.0
  2. CUDA 10.0
  3. Pytorch 1.2.0
  4. torchvision 0.4.0
  5. numpy 1.17.0
  6. scipy 1.3.1

First clone the repo in the directory you want to work:

    git clone  
    cd netadapt

In the following context, we assume you are at the repo root.

If the versions of Python and CUDA are the same as yours, you can download the python packages using:

    pip install -r requirements.txt

To verify the downloaded code base is correct, please run either

    sh scripts/


    sh scripts/
    sh scripts/
    sh scripts/

If it is correct, you should not see any FAIL.


In order to apply NetAdapt, run:

    python [-h] [-gp GPUS [GPUS ...]] [-re] [-im INIT_MODEL_PATH]
             [-mi MAX_ITERS] [-lr FINETUNE_LR] [-bu BUDGET]
             [-bur BUDGET_RATIO] [-rt RESOURCE_TYPE]
             [-ir INIT_RESOURCE_REDUCTION]
             [-dp DATASET_PATH] [-a ARCH] [-si SAVE_INTERVAL]
             working_folder input_data_shape input_data_shape
  • working_folder: Root folder where models, related files and history information are saved. You can see how models are pruned progressively in working_folder/master/history.txt.

  • input_data_shape: Input data shape (C, H, W) (default: 3 224 224). If you want to apply NetAdapt to different tasks, you might need to change data shape.

  • -h, --help: Show this help message and exit.

  • -gp GPUS [GPUS ...], --gpus GPUS [GPUS ...]: Indices of available gpus (default: 0).

  • -re, --resume: Resume from previous iteration. In order to resume, specify --resume and specify working_folder as the one you want to resume.
    The resumed arguments will overwrite the arguments provided here.
    For example, if you want to simplify a model by pruning and finetuning for 30 iterations (under working_folder), however, your program terminated after 20 iterations.
    Then you can use --resume to restore and continue for the last 10 iterations.

  • -im INIT_MODEL_PATH, --init_model_path INIT_MODEL_PATH: Path to pretrained model.

  • -mi MAX_ITERS, --max_iters MAX_ITERS: Maximum iteration of removing filters and short-term fine-tune (default: 10).

  • -lr FINETUNE_LR, --finetune_lr FINETUNE_LR: Short-term fine-tune learning rate (default: 0.001).

  • -bu BUDGET, --budget BUDGET: Resource constraint. If resource < budget, the process is terminated.

  • -bur BUDGET_RATIO, --budget_ratio BUDGET_RATIO: If --budget is not specified, buget = budget_ratio*(pretrained model resource) (default: 0.25).

  • -rt RESOURCE_TYPE, --resource_type RESOURCE_TYPE: Resource constraint type (default: FLOPS). We currently support FLOPS, WEIGHTS, and LATENCY (device cuda:0). If you want to add other resource
    types, please modify def compute_resource(...) in network_util python files (e.g. network_utils/network_utils_alexnet).

  • -ir INIT_RESOURCE_REDUCTION, --init_resource_reduction INIT_RESOURCE_REDUCTION: For each iteration, target resource = current resource - init_resource_reduction*(resource_reduction_decay**(iteration-1)).

  • -irr INIT_RESOURCE_REDUCTION_RATIO, --init_resource_reduction_ratio INIT_RESOURCE_REDUCTION_RATIO: If --init_resource_reduction is not specified,
    init_resource_reduction = init_resource_reduction_ratio*(pretrained model resource) (default: 0.025).

  • -rd RESOURCE_REDUCTION_DECAY, --resource_reduction_decay RESOURCE_REDUCTION_DECAY: For each iteration, target resource = current resource - init_resource_reduction*(resource_reduction_decay**(iteration-1)) (default: 0.96).

  • -st SHORT_TERM_FINE_TUNE_ITERATION, --short_term_fine_tune_iteration SHORT_TERM_FINE_TUNE_ITERATION: Short-term fine-tune iteration (default: 10).

  • -lt LOOKUP_TABLE_PATH, --lookup_table_path LOOKUP_TABLE_PATH: Path to lookup table.

  • -dp DATASET_PATH, --dataset_path DATASET_PATH: Path to dataset.

  • -a ARCH, --arch ARCH network_utils: Defines how networks are pruned, fine-tuned, and evaluated. If you want to use
    your own method, please see Customization and specify here. (default: alexnet)

  • -si SAVE_INTERVAL, --save_interval SAVE_INTERVAL: Interval of iterations that all pruned models at the same iteration will be saved.
    Use -1 to save only the best model at each iteration.
    Use 1 to save all models at each iteration. (default: -1).


We provide a simple example of applying NetAdapt to a very small network:

    sh scripts/

Detailed examples of applying NetAdapt to AlexNet/MobileNet on CIFAR-10 are shown here (AlexNet) and here (MobileNet).

photo not available

If you want to apply the algorithm to different networks or even different tasks,
please see the following Customization section.


To apply NetAdapt to differenct networks or different tasks, please follow the instructions:

  1. Create your own network_utils python file (said and place it under network_utils.

  2. Implement functions described in

  3. As we provide an example of applying NetAdapt to AlexNet, you can also build your network_utils based on

        cd network_utils
        cp ./
  4. Add from .network_utils_yourNetworkOrTask import * to, which is under the same directory.

  5. Modify class networkUtils_alexnet(...) in line 44 in to class networkUtils_yourNetworkOrTask(...).

  6. Modify def alexnet(...) in line 325-326 to:

        def yourNetworkOrTask(model, input_data_shape, dataset_path, finetune_lr=1e-3):
            return networkUtils_yourNetworkOrTask(model, input_data_shape, dataset_path, finetune_lr)
  7. Specify training/validation data loader, loss functions, optimizers, network architecture, training method, and evaluation method in if there is any difference from the AlexNet example:

    • Modify data loader and loss functionsin function def __init__(...): in line 52.

    • Specify additive skip connections if there is any and modify def simplify_network_def_based_on_constraint(...) in
      You can see how our implementation uses additive skip connections here.

    • Modify training method (short-term finetune) in function def fine_tune(...): in line 245.

    • Modify evaluation method in function def evaluate(...): in line 291.

    You can see how these methods are utilized by the framework here.

  8. Our current code base supports pruning Conv2d, ConvTranspose2d, and Linear with additive skip connection.
    If your network architecture is not supported, please modify this.
    If you want to use other metrics (resource type) to prune networks, please modify this.

  9. We can apply NetAdapt to different networks or tasks by using --arch yourNetworkOrTask in scripts/
    As for the values of other arguments, please see Usage.
    Generally, if you want to apply NetAdapt to a different task, you might change input_data_shape.
    If your network architecture is very different from that of MobileNet, you would have to modify the values of --init_resource_reduction_ratio and --resource_reduction_decay to get a different resource reduction schedule.