fengxingxiang / modelcompression-2019

Thesis Test Bench

Prerequisites

• Python 3.6+
• Pip3
• Virtualenv (optional)

Setup

To get started, set up a virtualenv (if so desired):

~ virtualenv venv
~ . venv/bin/activate

Next, install all dependencies:

~ pip install -r requirements.txt

Everything should now be ready.

Usage

Train a model from scratch

In order to re-train one of the available models from scratch (without pruning or quantization), simply run the equivalent *_classifier.py file. Example:

~ python cifar_classifier.py --save-model

By default, this will train a model and save it to models/cifar_classifier.pt.

Evaluate Pruning

Step 1

In order to prune and evaluate a pruned model, ensure that a trained model exists. If you are not sure how to do this, follow the instructions here.

Step 2

Once confirming a trained model exists, ensure that there exists an appropriate test time configuration for that model. This can be found in main.py, under the configurations variable. This will change in the future to a more robust solution.

Step 3 (WILL CHANGE)

Run the command:

python main.py

You will be presented with a list of available configurations:

Select a model type to prune. Models available:
    0:{'model': <class 'cifar_classifier.MaskedCifar'>, 'dataset': <class 'torchvision.datasets.cifar.CIFAR10'>, 'transforms': [ToTensor(), Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))], 'loss_fn': <function cross_entropy at 0x000001C3C9EB1E18>}
    
    1:{'model': <class 'mnist_classifier.MaskedMNist'>, 'dataset': <class 'torchvision.datasets.mnist.MNIST'>, 'transforms': [ToTensor(), Normalize(mean=(0.1307,), std=(0.3081,))], 'loss_fn': <function nll_loss at 0x000001C3C9EB1AE8>}

Select the appropraite configuration by selecting the correct index:

Enter selected index and press enter: 0

Then, select the appropriate trained model parameters to load:

Select a model params file to load by index. Models available:
    0:cifar_classifier.pt
    1:mnist_classifier.pt
    2:mnist_cnn.pt

Enter selected index and press enter: 0

Finally, input the desired pruning percentage:

Select pruning percentage: (0-100)%: 50

The script will evaluate the pre-pruned model, prune the model, print relevant pruning statistics by layer, and finally evaluate the pruned model.

Loading file cifar_classifier.pt for model {'model': <class 'cifar_classifier.MaskedCifar'>, 'dataset': <class 'torchvision.datasets.cifar.CIFAR10'>, 'transforms': [ToTensor(), Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))], 'loss_fn': <function cross_entropy at 0x000001C3C9EB1E18>}

Testing pre-pruned model..

Test set: Average loss: 1.1923, Accuracy: 5865/10000 (59%)

Pruning model..
Layer 1 | Conv layer | 15.11% parameters pruned
Layer 2 | Conv layer | 26.54% parameters pruned
Layer 3 | Linear layer | 56.20% parameters pruned
Layer 4 | Linear layer | 30.25% parameters pruned
Layer 5 | Linear layer | 18.21% parameters pruned
Final pruning rate: 49.81%

Evaluating pruned model..

Test set: Average loss: 1.2210, Accuracy: 5714/10000 (57%)