create branch for v2.9

examples/model_compress/pruning/legacy/finetune_kd_torch.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+'''
+NNI exmaple for fine-tuning the pruned model with KD.
+Run basic_pruners_torch.py first to get the masks of the pruned model. Then pass the mask as argument for model speedup. The compressed model is further used for fine-tuning.
+'''
+
+import argparse
+import os
+import sys
+from copy import deepcopy
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from nni.compression.pytorch import ModelSpeedup
+from torch.optim.lr_scheduler import MultiStepLR
+from basic_pruners_torch import get_data
+
+from pathlib import Path
+sys.path.append(str(Path(__file__).absolute().parents[2] / 'models'))
+from mnist.lenet import LeNet
+from cifar10.vgg import VGG
+
+class DistillKL(nn.Module):
+    """Distilling the Knowledge in a Neural Network"""
+    def __init__(self, T):
+        super(DistillKL, self).__init__()
+        self.T = T
+
+    def forward(self, y_s, y_t):
+        p_s = F.log_softmax(y_s/self.T, dim=1)
+        p_t = F.softmax(y_t/self.T, dim=1)
+        loss = F.kl_div(p_s, p_t, size_average=False) * (self.T**2) / y_s.shape[0]
+        return loss
+
+def get_dummy_input(args, device):
+    if args.dataset == 'mnist':
+        dummy_input = torch.randn([args.test_batch_size, 1, 28, 28]).to(device)
+    elif args.dataset in ['cifar10', 'imagenet']:
+        dummy_input = torch.randn([args.test_batch_size, 3, 32, 32]).to(device)
+    return dummy_input
+    
+def get_model_optimizer_scheduler(args, device, test_loader, criterion):
+    if args.model == 'LeNet':
+        model = LeNet().to(device)
+    elif args.model == 'vgg16':
+        model = VGG(depth=16).to(device)
+    elif args.model == 'vgg19':
+        model = VGG(depth=19).to(device)
+    else:
+        raise ValueError("model not recognized")
+
+    # In this example, we set the architecture of teacher and student to be the same. It is feasible to set a different teacher architecture.
+    if args.teacher_model_dir is None:
+        raise NotImplementedError('please load pretrained teacher model first')
+    else:
+        model.load_state_dict(torch.load(args.teacher_model_dir))
+        best_acc = test(args, model, device, criterion, test_loader)
+
+    model_t = deepcopy(model)
+    model_s = deepcopy(model)
+
+    if args.student_model_dir is not None:
+        # load the pruned student model checkpoint
+        model_s.load_state_dict(torch.load(args.student_model_dir))
+
+    dummy_input = get_dummy_input(args, device)
+    m_speedup = ModelSpeedup(model_s, dummy_input, args.mask_path, device)
+    m_speedup.speedup_model()
+
+    module_list = nn.ModuleList([])
+    module_list.append(model_s)
+    module_list.append(model_t)
+
+    # setup opotimizer for fine-tuning studeng model
+    optimizer = torch.optim.SGD(model_s.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
+    scheduler = MultiStepLR(
+                optimizer, milestones=[int(args.fine_tune_epochs*0.5), int(args.fine_tune_epochs*0.75)], gamma=0.1)
+        
+    print('Pretrained teacher model acc:', best_acc)
+    return module_list, optimizer, scheduler
+
+
+def train(args, models, device, train_loader, criterion, optimizer, epoch):
+    # model.train()
+    model_s = models[0].train()
+    model_t = models[-1].eval()
+    cri_cls = criterion
+    cri_kd = DistillKL(args.kd_T)
+
+
+    for batch_idx, (data, target) in enumerate(train_loader):
+        data, target = data.to(device), target.to(device)
+        optimizer.zero_grad()
+        output_s = model_s(data)
+        output_t = model_t(data)
+
+        loss_cls = cri_cls(output_s, target)
+        loss_kd = cri_kd(output_s, output_t)
+        loss = loss_cls + loss_kd
+        loss.backward()
+
+        optimizer.step()
+        if batch_idx % args.log_interval == 0:
+            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
+                epoch, batch_idx * len(data), len(train_loader.dataset),
+                100. * batch_idx / len(train_loader), loss.item()))
+            if args.dry_run:
+                break
+
+def test(args, model, device, criterion, test_loader):
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for data, target in test_loader:
+            data, target = data.to(device), target.to(device)
+            output = model(data)
+            test_loss += criterion(output, target).item()
+            pred = output.argmax(dim=1, keepdim=True)
+            correct += pred.eq(target.view_as(pred)).sum().item()
+    test_loss /= len(test_loader.dataset)
+    acc = 100 * correct / len(test_loader.dataset)
+
+    print('Test Loss: {}  Accuracy: {}%\n'.format(
+        test_loss, acc))
+    return acc
+
+
+def main(args):
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    os.makedirs(args.experiment_data_dir, exist_ok=True)
+
+    # prepare model and data
+    train_loader, test_loader, criterion = get_data(args.dataset, args.data_dir, args.batch_size, args.test_batch_size)
+    models, optimizer, scheduler = get_model_optimizer_scheduler(args, device, test_loader, criterion)
+
+    best_top1 = 0
+    if args.test_only:
+        test(args, models[0], device, criterion, test_loader)
+
+    print('start fine-tuning...')
+    for epoch in range(args.fine_tune_epochs):
+        print('# Epoch {} #'.format(epoch))
+        train(args, models, device, train_loader, criterion, optimizer, epoch)
+        scheduler.step()
+
+        # test student only
+        top1 = test(args, models[0], device, criterion, test_loader)
+        if top1 > best_top1:
+            best_top1 = top1
+            torch.save(models[0].state_dict(), os.path.join(args.experiment_data_dir, 'model_trained.pth'))
+            print('Model trained saved to %s' % args.experiment_data_dir)
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='PyTorch Example for model comporession')
+
+    # dataset and model
+    parser.add_argument('--dataset', type=str, default='cifar10',
+                        help='dataset to use, mnist, cifar10 or imagenet')
+    parser.add_argument('--data-dir', type=str, default='./data/',
+                        help='dataset directory')
+    parser.add_argument('--model', type=str, default='vgg16',
+                        choices=['LeNet', 'vgg16' ,'vgg19', 'resnet18'],
+                        help='model to use')
+    parser.add_argument('--teacher-model-dir', type=str, default=None,
+                        help='path to the pretrained teacher model checkpoint')
+    parser.add_argument('--mask-path', type=str, default=None,
+                        help='path to the pruned student model mask file')
+    parser.add_argument('--student-model-dir', type=str, default=None,
+                        help='path to the pruned student model checkpoint')
+    parser.add_argument('--batch-size', type=int, default=128,
+                        help='input batch size for training')
+    parser.add_argument('--test-batch-size', type=int, default=200,
+                        help='input batch size for testing')
+    parser.add_argument('--fine-tune-epochs', type=int, default=160,
+                        help='epochs to fine tune')
+    parser.add_argument('--experiment-data-dir', type=str, default='./experiment_data',
+                        help='For saving output checkpoints')
+    parser.add_argument('--log-interval', type=int, default=100, metavar='N',
+                        help='how many batches to wait before logging training status')
+    parser.add_argument('--dry-run', action='store_true', default=False,
+                        help='quickly check a single pass')
+    parser.add_argument('--test-only', action='store_true', default=False,
+                        help='run test only')
+
+
+    # knowledge distillation
+    parser.add_argument('--kd_T', type=float, default=4,
+                        help='temperature for KD distillation')
+
+
+    args = parser.parse_args()
+    main(args)

examples/model_compress/pruning/legacy/lottery_torch_mnist_fc.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+'''
+NNI exmaple for reproducing Lottery Ticket Hypothesis.
+'''
+
+import argparse
+import copy
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.data
+import torchvision.datasets as datasets
+import torchvision.transforms as transforms
+from nni.algorithms.compression.pytorch.pruning import LotteryTicketPruner
+
+class fc1(nn.Module):
+
+    def __init__(self, num_classes=10):
+        super(fc1, self).__init__()
+        self.classifier = nn.Sequential(
+            nn.Linear(28 * 28, 300),
+            nn.ReLU(inplace=True),
+            nn.Linear(300, 100),
+            nn.ReLU(inplace=True),
+            nn.Linear(100, num_classes),
+        )
+
+    def forward(self, x):
+        x = torch.flatten(x, 1)
+        x = self.classifier(x)
+        return x
+
+def train(model, train_loader, optimizer, criterion):
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model.train()
+    for batch_idx, (imgs, targets) in enumerate(train_loader):
+        optimizer.zero_grad()
+        imgs, targets = imgs.to(device), targets.to(device)
+        output = model(imgs)
+        train_loss = criterion(output, targets)
+        train_loss.backward()
+        optimizer.step()
+        if batch_idx % 100 == 0:
+            print('{:2.0f}%  Loss {}'.format(
+                100 * batch_idx / len(train_loader), train_loss.item()))
+
+    return train_loss.item()
+
+def test(model, test_loader, criterion):
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for data, target in test_loader:
+            data, target = data.to(device), target.to(device)
+            output = model(data)
+            test_loss += F.nll_loss(output, target, reduction='sum').item()  # sum up batch loss
+            pred = output.data.max(1, keepdim=True)[1]  # get the index of the max log-probability
+            correct += pred.eq(target.data.view_as(pred)).sum().item()
+        test_loss /= len(test_loader.dataset)
+        accuracy = 100. * correct / len(test_loader.dataset)
+    return accuracy
+
+
+if __name__ == '__main__':
+    """
+    THE LOTTERY TICKET HYPOTHESIS: FINDING SPARSE, TRAINABLE NEURAL NETWORKS (https://arxiv.org/pdf/1803.03635.pdf)
+
+    The Lottery Ticket Hypothesis. A randomly-initialized, dense neural network contains a subnetwork that is
+    initialized such that—when trained in isolation—it can match the test accuracy of the original network after
+    training for at most the same number of iterations.
+
+    Identifying winning tickets. We identify a winning ticket by training a network and pruning its
+    smallest-magnitude weights. The remaining, unpruned connections constitute the architecture of the
+    winning ticket. Unique to our work, each unpruned connection’s value is then reset to its initialization
+    from original network before it was trained. This forms our central experiment:
+        1. Randomly initialize a neural network f(x; θ0) (where θ0 ∼ Dθ).
+        2. Train the network for j iterations, arriving at parameters θj .
+        3. Prune p% of the parameters in θj , creating a mask m.
+        4. Reset the remaining parameters to their values in θ0, creating the winning ticket f(x; mθ0).
+    As described, this pruning approach is one-shot: the network is trained once, p% of weights are
+    pruned, and the surviving weights are reset. However, in this paper, we focus on iterative pruning,
+    which repeatedly trains, prunes, and resets the network over n rounds; each round prunes p**(1/n) % of
+    the weights that survive the previous round. Our results show that iterative pruning finds winning tickets
+    that match the accuracy of the original network at smaller sizes than does one-shot pruning.
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--train_epochs", type=int, default=10, help="training epochs")
+    args = parser.parse_args()
+
+    transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,))])
+    traindataset = datasets.MNIST('./data', train=True, download=True, transform=transform)
+    testdataset = datasets.MNIST('./data', train=False, transform=transform)
+    train_loader = torch.utils.data.DataLoader(traindataset, batch_size=60, shuffle=True, num_workers=0, drop_last=False)
+    test_loader = torch.utils.data.DataLoader(testdataset, batch_size=60, shuffle=False, num_workers=0, drop_last=True)
+
+    model = fc1().to("cuda" if torch.cuda.is_available() else "cpu")
+    optimizer = torch.optim.Adam(model.parameters(), lr=1.2e-3)
+    criterion = nn.CrossEntropyLoss()
+
+    # Record the random intialized model weights
+    orig_state = copy.deepcopy(model.state_dict())
+
+    # train the model to get unpruned metrics
+    for epoch in range(args.train_epochs):
+        train(model, train_loader, optimizer, criterion)
+    orig_accuracy = test(model, test_loader, criterion)
+    print('unpruned model accuracy: {}'.format(orig_accuracy))
+
+    # reset model weights and optimizer for pruning
+    model.load_state_dict(orig_state)
+    optimizer = torch.optim.Adam(model.parameters(), lr=1.2e-3)
+
+    # Prune the model to find a winning ticket
+    configure_list = [{
+        'prune_iterations': 5,
+        'sparsity': 0.96,
+        'op_types': ['default']
+    }]
+    pruner = LotteryTicketPruner(model, configure_list, optimizer)
+    pruner.compress()
+
+    best_accuracy = 0.
+    best_state_dict = None
+
+    for i in pruner.get_prune_iterations():
+        pruner.prune_iteration_start()
+        loss = 0
+        accuracy = 0
+        for epoch in range(args.train_epochs):
+            loss = train(model, train_loader, optimizer, criterion)
+            accuracy = test(model, test_loader, criterion)
+            print('current epoch: {0}, loss: {1}, accuracy: {2}'.format(epoch, loss, accuracy))
+            if accuracy > best_accuracy:
+                best_accuracy = accuracy
+                # state dict of weights and masks
+                best_state_dict = copy.deepcopy(model.state_dict())
+        print('prune iteration: {0}, loss: {1}, accuracy: {2}'.format(i, loss, accuracy))
+
+    if best_accuracy > orig_accuracy:
+        # load weights and masks
+        pruner.bound_model.load_state_dict(best_state_dict)
+        # reset weights to original untrained model and keep masks unchanged to export winning ticket
+        pruner.load_model_state_dict(orig_state)
+        pruner.export_model('model_winning_ticket.pth', 'mask_winning_ticket.pth')
+        print('winning ticket has been saved: model_winning_ticket.pth, mask_winning_ticket.pth')
+    else:
+        print('winning ticket is not found in this run, you can run it again.')

examples/model_compress/pruning/legacy/mobilenetv2_end2end/prepare_data.sh

+#!/bin/bash
+
+# download and preprocess the Stanford Dogs dataset
+
+mkdir -p data/stanford-dogs
+
+# download raw data (images, annotations, and train-test split)
+cd data/stanford-dogs
+
+if [ ! -d './Images' ] ; then
+  if [ ! -f 'images.tar' ] ; then
+    wget http://vision.stanford.edu/aditya86/ImageNetDogs/images.tar
+  fi
+  tar -xvf images.tar
+fi
+
+if [ ! -d './Annotation' ] ; then
+  if [ ! -f 'annotation.tar' ] ; then
+    wget http://vision.stanford.edu/aditya86/ImageNetDogs/annotation.tar
+  fi
+  tar -xvf annotation.tar
+fi
+
+if [ ! -f 'lists.tar' ] ; then
+  wget http://vision.stanford.edu/aditya86/ImageNetDogs/lists.tar
+fi
+tar -xvf lists.tar
+
+cd ../..
+
+# preprocess: train-valid-test splitting and image cropping
+python preprocess.py

examples/model_compress/pruning/legacy/mobilenetv2_end2end/preprocess.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import xml.etree.ElementTree
+from PIL import Image
+import numpy as np
+from sklearn.model_selection import train_test_split
+from scipy import io
+
+
+ROOT_DIR = './data/stanford-dogs/'
+NUM_CATEGORIES = 120
+OUT_IMAGE_SIZE = (224, 224)
+RANDOM_SEED = 42                        # for splitting train and validation
+TRAIN_RATIO = 0.9                       # train / (train + validation)
+
+
+def get_bounding_box(annotation_file):
+    """
+    Parse the annotation file and returns the bounding box information
+    Parameters
+    ----------
+    annotation_file: path to the annotation XML file
+
+    Returns
+    -------
+    A dict containing bounding box information
+    """
+    ret = {}
+    xml_root = xml.etree.ElementTree.parse(annotation_file).getroot()
+    bounding_box = xml_root.findall('object')[0].findall('bndbox')[0]
+    ret['X_min'] = int(bounding_box.findall('xmin')[0].text)
+    ret['X_max'] = int(bounding_box.findall('xmax')[0].text)
+    ret['Y_min'] = int(bounding_box.findall('ymin')[0].text)
+    ret['Y_max'] = int(bounding_box.findall('ymax')[0].text)
+
+    return ret
+
+
+def main(root_dir):
+    try:
+        os.mkdir(root_dir + 'Processed')
+        os.mkdir(root_dir + 'Processed/train')
+        os.mkdir(root_dir + 'Processed/valid')
+        os.mkdir(root_dir + 'Processed/test')
+    except:
+        print('Directory already exists. Nothing done.')
+        exit()
+        
+    # load train test splits
+    train_metadata = io.loadmat(root_dir + 'train_list.mat')
+    train_valid_file_list = [x[0][0] for x in train_metadata['file_list']]
+    train_valid_annotation_list = [x[0][0] for x in train_metadata['annotation_list']]
+    train_valid_labels = [x[0] - 1 for x in train_metadata['labels']]
+    train_valid_lists = [x for x in zip(train_valid_file_list, train_valid_annotation_list, train_valid_labels)]
+    train_lists, valid_lists = train_test_split(train_valid_lists, train_size=TRAIN_RATIO, random_state=RANDOM_SEED)
+    train_file_list, train_annotation_list, train_labels = zip(*train_lists)
+    valid_file_list, valid_annotation_list, valid_labels = zip(*valid_lists)
+
+    test_metadata = io.loadmat(root_dir + 'test_list.mat')
+    test_file_list = [x[0][0] for x in test_metadata['file_list']]
+    test_annotation_list = [x[0][0] for x in test_metadata['annotation_list']]
+    test_labels = [x[0] - 1 for x in test_metadata['labels']]
+
+    label2idx = {}
+    for split, file_list, annotation_list, labels in zip(['train', 'valid', 'test'],
+                                                         [train_file_list, valid_file_list, test_file_list],
+                                                         [train_annotation_list, valid_annotation_list, test_annotation_list],
+                                                         [train_labels, valid_labels, test_labels]):
+        print('Preprocessing {} set: {} cases'.format(split, len(file_list)))
+        for cur_file, cur_annotation, cur_label in zip(file_list, annotation_list, labels):
+            label_name = cur_file.split('/')[0].split('-')[-1].lower()
+            if label_name not in label2idx:
+                label2idx[label_name] = cur_label
+            image = Image.open(root_dir + '/Images/' + cur_file)
+
+            # cropping and reshape
+            annotation_file = root_dir + '/Annotation/' + cur_annotation
+            bounding_box = get_bounding_box(annotation_file)
+            image = image.crop([bounding_box['X_min'], bounding_box['Y_min'],
+                                bounding_box['X_max'], bounding_box['Y_max']])
+            image = image.convert('RGB')
+            image = image.resize(OUT_IMAGE_SIZE)
+
+            # Normalize and save the instance
+            X = np.array(image)
+            X = (X - np.mean(X, axis=(0, 1))) / np.std(X, axis=(0, 1))          # normalize each channel separately
+
+            # image.save(root_dir + 'Processed/' + split + '/' + image_name)
+            np.save(root_dir + 'Processed/' + split + '/' + cur_file.split('/')[-1].replace('.jpg', '.npy'),
+                    {'input': X, 'label': cur_label})
+
+    # save mapping from label name to index to a dict
+    with open(ROOT_DIR + '/category_dict.tsv', 'w') as dict_f:
+        final_dict_list = sorted(list(label2idx.items()), key=(lambda x: x[-1]))
+        for label, index in final_dict_list:
+            dict_f.write('{}\t{}\n'.format(index, label))
+        print(final_dict_list)
+
+
+if __name__ == '__main__':
+    main(ROOT_DIR)

examples/model_compress/pruning/legacy/mobilenetv2_end2end/pretrain.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import argparse
+from time import gmtime, strftime
+import torch
+import torch.nn as nn
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+import numpy as np
+
+from utils import *
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+def run_validation(model, valid_dataloader):
+    model.eval()
+    
+    loss_func = nn.CrossEntropyLoss()
+    acc_list, loss_list = [], []
+    with torch.no_grad():
+        for i, (inputs, labels) in enumerate(tqdm(valid_dataloader)):
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds= model(inputs)
+            pred_idx = preds.max(1).indices
+            acc = (pred_idx == labels).sum().item() / labels.size(0)
+            acc_list.append(acc)
+            loss = loss_func(preds, labels).item()
+            loss_list.append(loss)
+
+    valid_loss = np.array(loss_list).mean()
+    valid_acc = np.array(acc_list).mean()
+    
+    return valid_loss, valid_acc
+
+
+def run_pretrain(args):
+    print(args)
+    torch.set_num_threads(args.n_workers)
+    
+    model_type = 'mobilenet_v2_torchhub'   
+    pretrained = True                      # load imagenet weight
+    experiment_dir = 'pretrained_{}'.format(model_type) if args.experiment_dir is None else args.experiment_dir
+    os.mkdir(experiment_dir)
+    checkpoint = None
+    input_size = 224
+    n_classes = 120
+    
+    log = open(experiment_dir + '/pretrain.log', 'w')
+    
+    model = create_model(model_type=model_type, pretrained=pretrained, n_classes=n_classes,
+                         input_size=input_size, checkpoint=checkpoint)
+    model = model.to(device)
+    print(model)
+    # count_flops(model, device=device)
+
+    train_dataset = TrainDataset('./data/stanford-dogs/Processed/train')
+    train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)
+    valid_dataset = EvalDataset('./data/stanford-dogs/Processed/valid')
+    valid_dataloader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False)
+
+    criterion = nn.CrossEntropyLoss()
+    optimizer = torch.optim.SGD(model.parameters(), lr=args.learning_rate, momentum=0.9, weight_decay=args.weight_decay)
+
+    best_valid_acc = 0.0
+    for epoch in range(args.n_epochs):
+        print('Start training epoch {}'.format(epoch))
+        loss_list = []
+
+        # train
+        model.train()
+        for i, (inputs, labels) in enumerate(tqdm(train_dataloader)):
+            optimizer.zero_grad()
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds = model(inputs)
+            loss = criterion(preds, labels)
+            loss_list.append(loss.item())
+            loss.backward()
+            optimizer.step()
+            
+        # validation
+        valid_loss, valid_acc = run_validation(model, valid_dataloader)
+        train_loss = np.array(loss_list).mean()
+        print('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}'.format
+              (epoch, train_loss, valid_loss, valid_acc))
+        log.write('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}\n'.format
+                  (epoch, train_loss, valid_loss, valid_acc))
+        
+        # save
+        if valid_acc > best_valid_acc:
+            best_valid_acc = valid_acc
+            torch.save(model.state_dict(), experiment_dir + '/checkpoint_best.pt')
+
+    log.close()
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Example code for pruning MobileNetV2')
+
+    parser.add_argument('--experiment_dir', type=str, default=None,
+                        help='directory containing the pretrained model')
+    parser.add_argument('--checkpoint_name', type=str, default='checkpoint_best.pt',
+                         help='checkpoint of the pretrained model')
+    
+    # finetuning parameters
+    parser.add_argument('--n_workers', type=int, default=16,
+                        help='number of threads')
+    parser.add_argument('--n_epochs', type=int, default=180,
+                        help='number of epochs to train the model')
+    parser.add_argument('--learning_rate', type=float, default=1e-4)
+    parser.add_argument('--weight_decay', type=float, default=0.0)
+    parser.add_argument('--batch_size', type=int, default=32,
+                        help='input batch size for training and inference')
+
+    args = parser.parse_args()
+    return args
+
+    
+if __name__ == '__main__':
+    args = parse_args()
+    run_pretrain(args)

examples/model_compress/pruning/legacy/mobilenetv2_end2end/pruning_experiments.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import argparse
+import copy
+from time import gmtime, strftime
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+import numpy as np
+
+import nni
+from nni.compression.pytorch import ModelSpeedup
+from nni.algorithms.compression.pytorch.pruning import (
+    LevelPruner,
+    SlimPruner,
+    FPGMPruner,
+    TaylorFOWeightFilterPruner,
+    L1FilterPruner,
+    L2FilterPruner,
+    AGPPruner,
+    ActivationMeanRankFilterPruner,
+    ActivationAPoZRankFilterPruner
+)
+
+from utils import *
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+model_type = 'mobilenet_v2_torchhub'
+input_size = 224
+n_classes = 120
+
+pruner_type_to_class = {'level': LevelPruner,
+                        'l1': L1FilterPruner,
+                        'l2': L2FilterPruner,
+                        'slim': SlimPruner,
+                        'fpgm': FPGMPruner,
+                        'taylorfo': TaylorFOWeightFilterPruner,
+                        'agp': AGPPruner,
+                        'mean_activation': ActivationMeanRankFilterPruner,
+                        'apoz': ActivationAPoZRankFilterPruner}
+
+
+def run_eval(model, dataloader, device):
+    model.eval()
+    loss_func = nn.CrossEntropyLoss()
+    acc_list, loss_list = [], []
+    with torch.no_grad():
+        for i, (inputs, labels) in enumerate(tqdm(dataloader)):
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds= model(inputs)
+            pred_idx = preds.max(1).indices
+            acc = (pred_idx == labels).sum().item() / labels.size(0)
+            acc_list.append(acc)
+            loss = loss_func(preds, labels).item()
+            loss_list.append(loss)
+
+    final_loss = np.array(loss_list).mean()
+    final_acc = np.array(acc_list).mean()
+
+    return final_loss, final_acc
+
+
+def run_finetune(model, train_dataloader, valid_dataloader, device,
+                 n_epochs=2, learning_rate=1e-4, weight_decay=0.0, log=None):    
+    criterion = nn.CrossEntropyLoss()
+    optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)
+
+    best_valid_acc = 0.0
+    best_model = None
+    for epoch in range(n_epochs):
+        print('Start finetuning epoch {}'.format(epoch))
+        loss_list = []
+
+        # train
+        model.train()
+        for i, (inputs, labels) in enumerate(tqdm(train_dataloader)):
+            optimizer.zero_grad()
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds = model(inputs)
+            loss = criterion(preds, labels)
+            loss_list.append(loss.item())
+            loss.backward()
+            optimizer.step()
+            
+        # validation
+        valid_loss, valid_acc = run_eval(model, valid_dataloader, device)
+        train_loss = np.array(loss_list).mean()
+        print('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}'.format
+              (epoch, train_loss, valid_loss, valid_acc))
+        if log is not None:
+            log.write('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}'.format
+                      (epoch, train_loss, valid_loss, valid_acc))
+
+        if valid_acc > best_valid_acc:
+            best_valid_acc = valid_acc
+            best_model = copy.deepcopy(model).to(device)
+
+    print("Best validation accuracy: {}".format(best_valid_acc))
+    if log is not None:
+        log.write("Best validation accuracy: {}".format(best_valid_acc))
+    
+    model = best_model
+    return model
+
+
+def run_finetune_distillation(student_model, teacher_model, train_dataloader, valid_dataloader, device,
+                              alpha, temperature,
+                              n_epochs=2, learning_rate=1e-4, weight_decay=0.0, log=None):
+    optimizer = torch.optim.Adam(student_model.parameters(), lr=learning_rate, weight_decay=weight_decay)
+    # optimizer = torch.optim.SGD(student_model.parameters(), lr=learning_rate, momentum=0.9)            
+
+    best_valid_acc = 0.0
+    best_model = None
+    for epoch in range(n_epochs):
+        print('Start finetuning with distillation epoch {}'.format(epoch))
+        loss_list = []
+
+        # train
+        student_model.train()
+        for i, (inputs, labels) in enumerate(tqdm(train_dataloader)):
+            optimizer.zero_grad()
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            with torch.no_grad():
+                teacher_preds = teacher_model(inputs)
+            
+            preds = student_model(inputs)
+            soft_loss = nn.KLDivLoss()(F.log_softmax(preds/temperature, dim=1),
+                                       F.softmax(teacher_preds/temperature, dim=1))
+            hard_loss = F.cross_entropy(preds, labels)
+            loss = soft_loss * (alpha * temperature * temperature) + hard_loss * (1. - alpha)
+            loss_list.append(loss.item())
+            loss.backward()
+            optimizer.step()
+
+        # validation
+        valid_loss, valid_acc = run_eval(student_model, valid_dataloader, device)
+        train_loss = np.array(loss_list).mean()
+        print('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}'.format
+              (epoch, train_loss, valid_loss, valid_acc))
+        if log is not None:
+            log.write('Epoch {}: train loss {:.4f}, valid loss {:.4f}, valid acc {:.4f}'.format
+                      (epoch, train_loss, valid_loss, valid_acc))
+        
+        if valid_acc > best_valid_acc:
+            best_valid_acc = valid_acc
+            best_model = copy.deepcopy(student_model).to(device)
+
+    print("Best validation accuracy: {}".format(best_valid_acc))
+    if log is not None:
+        log.write("Best validation accuracy: {}".format(best_valid_acc))
+
+    student_model = best_model
+    return student_model
+
+
+def trainer_helper(model, criterion, optimizer, dataloader, device):
+    print("Running trainer in tuner")
+    for epoch in range(1):
+        model.train()
+        for i, (inputs, labels) in enumerate(tqdm(dataloader)):
+            optimizer.zero_grad()
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds = model(inputs)
+            loss = criterion(preds, labels)
+            loss.backward()
+            optimizer.step()
+
+
+def trainer_helper_with_distillation(model, teacher_model, alpha, temperature, optimizer, dataloader, device):
+    print("Running trainer in tuner")
+    for epoch in range(1):
+        model.train()
+        for i, (inputs, labels) in enumerate(tqdm(dataloader)):
+            optimizer.zero_grad()
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            
+            with torch.no_grad():
+                teacher_preds = teacher_model(inputs)
+            preds = model(inputs)
+            soft_loss = nn.KLDivLoss()(F.log_softmax(preds/temperature, dim=1),
+                                       F.softmax(teacher_preds/temperature, dim=1))
+            hard_loss = F.cross_entropy(preds, labels)
+            loss = soft_loss * (alpha * temperature * temperature) + hard_loss * (1. - alpha)
+            loss.backward()
+            optimizer.step()
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description='Example code for pruning MobileNetV2')
+
+    parser.add_argument('--experiment_dir', type=str, required=True,
+                        help='directory containing the pretrained model')
+    parser.add_argument('--checkpoint_name', type=str, default='checkpoint_best.pt',
+                         help='checkpoint of the pretrained model')
+    
+    # pruner
+    parser.add_argument('--pruning_mode', type=str, default='conv1andconv2',
+                        choices=['conv0', 'conv1', 'conv2', 'conv1andconv2', 'all'])
+    parser.add_argument('--sparsity', type=float, default=0.5,
+                        help='target sparsity')
+    parser.add_argument('--pruner_name', type=str, default='l1',
+                        choices=['l1', 'l2', 'slim', 'agp',
+                                 'fpgm', 'mean_activation', 'apoz', 'taylorfo'],
+                        help='pruner to use')
+    # for agp only
+    parser.add_argument('--agp_pruning_alg', default='l1',
+                        choices=['l1', 'l2', 'slim', 'fpgm',
+                                 'mean_activation', 'apoz', 'taylorfo'],
+                        help='pruner to use for agp')
+    parser.add_argument('--agp_n_iters', type=int, default=64,
+                        help='number of iterations for agp')
+    parser.add_argument('--agp_n_epochs_per_iter', type=int, default=1,
+                        help='number of epochs per iteration for agp')
+
+    # speedup
+    parser.add_argument('--speedup', action='store_true', default=False,
+                        help='Whether to speedup the pruned model')
+
+    # finetuning parameters
+    parser.add_argument('--n_workers', type=int, default=16,
+                        help='number of threads')
+    parser.add_argument('--finetune_epochs', type=int, default=180,
+                        help='number of epochs to finetune the model')
+    parser.add_argument('--learning_rate', type=float, default=1e-4)
+    parser.add_argument('--weight_decay', type=float, default=0.0)
+    parser.add_argument('--batch_size', type=int, default=32,
+                        help='input batch size for training and inference')
+    parser.add_argument('--kd', action='store_true', default=False,
+                        help='Whether to use knowledge distillation')
+    parser.add_argument('--alpha', type=float, default=0.99,
+                        help='Alpha for knowledge distillation loss')
+    parser.add_argument('--temp', type=float, default=8,
+                        help='Temperature for knowledge distillation loss')
+
+    args = parser.parse_args()
+    return args
+
+
+def run_pruning(args):
+    print(args)
+    torch.set_num_threads(args.n_workers)
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+    log = open(args.experiment_dir + '/pruning_{}_{}_sparsity{}_{}.log'.format(
+        args.pruner_name, args.pruning_mode, args.sparsity,
+        strftime("%Y%m%d%H%M", gmtime())), 'w')
+    
+    train_dataset = TrainDataset('./data/stanford-dogs/Processed/train')
+    train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)
+    train_dataset_for_pruner = EvalDataset('./data/stanford-dogs/Processed/train')
+    train_dataloader_for_pruner = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False)
+    valid_dataset = EvalDataset('./data/stanford-dogs/Processed/valid')
+    valid_dataloader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False)
+    test_dataset = EvalDataset('./data/stanford-dogs/Processed/test')
+    test_dataloader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False)
+
+    model = create_model(model_type=model_type, pretrained=False, n_classes=n_classes,
+                         input_size=input_size, checkpoint=args.experiment_dir + '/' + args.checkpoint_name)
+    model = model.to(device)
+
+    teacher_model = None
+    if args.kd:
+        teacher_model = copy.deepcopy(model)
+
+    # evaluation before pruning
+    # count_flops(model, log, device)
+    initial_loss, initial_acc = run_eval(model, test_dataloader, device)
+    print('Before Pruning:\nLoss: {}\nAccuracy: {}'.format(initial_loss, initial_acc))
+    log.write('Before Pruning:\nLoss: {}\nAccuracy: {}\n'.format(initial_loss, initial_acc))
+
+    # set up config list and pruner
+    config_list = []
+    if 'conv0' in args.pruning_mode or args.pruning_mode == 'all':
+        if args.pruner_name == 'slim' or (args.pruner_name == 'agp' and args.agp_pruning_alg == 'slim'):
+            config_list.append({
+                'op_names': ['features.{}.conv.0.1'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+        else:
+            config_list.append({
+                'op_names': ['features.{}.conv.0.0'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+    if 'conv1' in args.pruning_mode or args.pruning_mode == 'all':
+        if args.pruner_name == 'slim' or (args.pruner_name == 'agp' and args.agp_pruning_alg == 'slim'):
+            config_list.append({
+                'op_names': ['features.{}.conv.1.1'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+        else:
+            config_list.append({
+                'op_names': ['features.{}.conv.1.0'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+    if 'conv2' in args.pruning_mode or args.pruning_mode == 'all':
+        if args.pruner_name == 'slim' or (args.pruner_name == 'agp' and args.agp_pruning_alg == 'slim'):
+            config_list.append({
+                'op_names': ['features.{}.conv.3'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+        else:
+            config_list.append({
+                'op_names': ['features.{}.conv.2'.format(x) for x in range(2, 18)],
+                'sparsity': args.sparsity
+            })
+    print(config_list)
+
+    kwargs = {}
+    if args.pruner_name in ['slim', 'taylorfo', 'mean_activation', 'apoz', 'agp']:
+        def trainer(model, optimizer, criterion, epoch):
+            if not args.kd:
+                return trainer_helper(model, criterion, optimizer, train_dataloader, device)
+            else:
+                return trainer_helper_with_distillation(model, teacher_model, args.alpha, args.temp, optimizer, train_dataloader, device)
+        kwargs = {
+            'trainer': trainer,
+            'optimizer': torch.optim.Adam(model.parameters()),
+            'criterion': nn.CrossEntropyLoss()
+        }
+        if args.pruner_name  == 'agp':
+            kwargs['pruning_algorithm'] = args.agp_pruning_alg
+            kwargs['num_iterations'] = args.agp_n_iters
+            kwargs['epochs_per_iteration'] = args.agp_n_epochs_per_iter
+        if args.pruner_name == 'slim':
+            kwargs['sparsifying_training_epochs'] = 10
+
+    # pruning
+    pruner = pruner_type_to_class[args.pruner_name](model, config_list, **kwargs)
+    pruner.compress()
+    pruner.export_model(args.experiment_dir + '/model_temp.pth', args.experiment_dir + './mask_temp.pth')
+    
+    # model speedup
+    pruner._unwrap_model()
+    if args.speedup:
+        dummy_input = torch.rand(1,3,224,224).to(device)
+        ms = ModelSpeedup(model, dummy_input, args.experiment_dir + './mask_temp.pth')
+        ms.speedup_model()
+        print(model)
+        count_flops(model, log)
+
+    intermediate_loss, intermediate_acc = run_eval(model, test_dataloader, device)
+    print('Before Finetuning:\nLoss: {}\nAccuracy: {}'.format(intermediate_loss, intermediate_acc))
+    log.write('Before Finetuning:\nLoss: {}\nAccuracy: {}\n'.format(intermediate_loss, intermediate_acc))
+
+    # finetuning
+    if args.kd:
+        model = run_finetune_distillation(model, teacher_model, train_dataloader, valid_dataloader, device,
+                                          args.alpha, args.temp, n_epochs=args.finetune_epochs,
+                                          learning_rate=args.learning_rate, weight_decay=args.weight_decay)
+    else:
+        model = run_finetune(model, train_dataloader, valid_dataloader, device, n_epochs=args.finetune_epochs,
+                             learning_rate=args.learning_rate, weight_decay=args.weight_decay)
+        
+    # final evaluation
+    final_loss, final_acc = run_eval(model, test_dataloader, device)
+    print('After Pruning:\nLoss: {}\nAccuracy: {}'.format(final_loss, final_acc))
+    log.write('After Pruning:\nLoss: {}\nAccuracy: {}'.format(final_loss, final_acc))
+
+    # clean up
+    filePaths = [args.experiment_dir + '/model_tmp.pth', args.experiment_dir + '/mask_tmp.pth']
+    for f in filePaths:
+        if os.path.exists(f):
+            os.remove(f)
+
+    log.close()
+    
+    
+if __name__ == '__main__':
+    args = parse_args()
+    run_pruning(args)

examples/model_compress/pruning/legacy/mobilenetv2_end2end/test.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import torch
+import torch.nn as nn
+from torch.utils.data import DataLoader
+from tqdm import tqdm
+import numpy as np
+
+from utils import create_model, EvalDataset, count_flops
+
+
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+
+model_type = 'mobilenet_v2_torchhub'    # 'mobilenet_v1' 'mobilenet_v2' 'mobilenet_v2_torchhub'
+pretrained = False                      # load imagenet weight (only for 'mobilenet_v2_torchhub')
+checkpoint_dir = './pretrained_{}/'.format(model_type)
+checkpoint = checkpoint_dir + '/checkpoint_best.pt'    # model checkpoint produced by pretrain.py
+input_size = 224
+n_classes = 120
+batch_size = 32
+
+
+def run_test():
+    model = create_model(model_type=model_type, pretrained=pretrained, n_classes=n_classes,
+                         input_size=input_size, checkpoint=checkpoint)
+    model = model.to(device)
+    print(model)
+    # count_flops(model, device=device)
+    
+    test_dataset = EvalDataset('./data/stanford-dogs/Processed/test')
+    test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
+
+    model.eval()
+    loss_func = nn.CrossEntropyLoss()
+    acc_list, loss_list = [], []
+    with torch.no_grad():
+        for i, (inputs, labels) in enumerate(tqdm(test_dataloader)):
+            inputs, labels = inputs.float().to(device), labels.to(device)
+            preds= model(inputs)
+            pred_idx = preds.max(1).indices
+            acc = (pred_idx == labels).sum().item() / labels.size(0)
+            acc_list.append(acc)
+            loss = loss_func(preds, labels).item()
+            loss_list.append(loss)
+
+    final_loss = np.array(loss_list).mean()
+    final_acc = np.array(acc_list).mean()
+    print('Test loss: {}\nTest accuracy: {}'.format(final_loss, final_acc))
+
+
+if __name__ == '__main__':
+    run_test()
+    

examples/model_compress/pruning/legacy/mobilenetv2_end2end/utils.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+import os
+import sys
+import torch
+from torch.utils.data import Dataset, DataLoader
+import torchvision.transforms as transforms
+import numpy as np
+from nni.compression.pytorch.utils import count_flops_params
+
+from pathlib import Path
+sys.path.append(str(Path(__file__).absolute().parents[3] / 'models'))
+from mobilenet import MobileNet
+from mobilenet_v2 import MobileNetV2
+
+
+def create_model(model_type=None, n_classes=120, input_size=224, checkpoint=None, pretrained=False, width_mult=1.):
+    if model_type == 'mobilenet_v1':
+        model = MobileNet(n_class=n_classes, profile='normal')
+    elif model_type == 'mobilenet_v2':
+        model = MobileNetV2(n_class=n_classes, input_size=input_size, width_mult=width_mult)
+    elif model_type == 'mobilenet_v2_torchhub':
+        model = torch.hub.load('pytorch/vision:v0.8.1', 'mobilenet_v2', pretrained=pretrained)
+        # model = torch.hub.load('pytorch/vision:v0.10.0', 'mobilenet_v2', pretrained=pretrained)
+        feature_size = model.classifier[1].weight.data.size()[1]
+        replace_classifier = torch.nn.Linear(feature_size, n_classes)
+        model.classifier[1] = replace_classifier
+    elif model_type is None:
+        model = None
+    else:
+        raise RuntimeError('Unknown model_type.')
+
+    if checkpoint is not None:
+        model.load_state_dict(torch.load(checkpoint))
+
+    return model
+
+
+def get_dataloader(dataset_type, data_path, batch_size=32, shuffle=True):
+    assert dataset_type in ['train', 'eval']
+    if dataset_type == 'train':
+        ds = TrainDataset(data_path)
+    else:
+        ds = EvalDataset(data_path)
+    return DataLoader(ds, batch_size, shuffle=shuffle)
+
+
+class TrainDataset(Dataset):
+    def __init__(self, npy_dir):
+        self.root_dir = npy_dir
+        self.case_names = [self.root_dir + '/' + x for x in os.listdir(self.root_dir)]
+        
+        transform_set = [transforms.Lambda(lambda x: x),
+                         transforms.RandomRotation(30),
+                         transforms.ColorJitter(),
+                         transforms.RandomHorizontalFlip(p=1)]
+        self.transform = transforms.RandomChoice(transform_set)
+        
+    def __len__(self):
+        return len(self.case_names)
+
+    def __getitem__(self, index):
+        instance = np.load(self.case_names[index], allow_pickle=True).item()
+        x = instance['input'].transpose(2, 0, 1)     # (C, H, W)
+        x = torch.from_numpy(x).type(torch.float)    # convert to Tensor to use torchvision.transforms
+        x = self.transform(x)
+        return x, instance['label']
+
+
+class EvalDataset(Dataset):
+    def __init__(self, npy_dir):
+        self.root_dir = npy_dir
+        self.case_names = [self.root_dir + '/' + x for x in os.listdir(self.root_dir)]
+
+    def __len__(self):
+        return len(self.case_names)
+
+    def __getitem__(self, index):
+        instance = np.load(self.case_names[index], allow_pickle=True).item()
+        x = instance['input'].transpose(2, 0, 1)
+        x = torch.from_numpy(x).type(torch.float)
+        return x, instance['label']
+
+
+def count_flops(model, log=None, device=None):
+    dummy_input = torch.rand([1, 3, 256, 256])
+    if device is not None:
+        dummy_input = dummy_input.to(device)
+    flops, params, results = count_flops_params(model, dummy_input)
+    print(f"FLOPs: {flops}, params: {params}")
+    if log is not None:
+        log.write(f"FLOPs: {flops}, params: {params}\n")
+    return flops, params

examples/model_compress/pruning/legacy/naive_prune_tf.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+'''
+NNI example for quick start of pruning.
+In this example, we use level pruner to prune the LeNet on MNIST.
+'''
+
+import argparse
+
+import tensorflow as tf
+from tensorflow.keras import Model
+from tensorflow.keras.layers import (Conv2D, Dense, Dropout, Flatten, MaxPool2D, BatchNormalization)
+
+from nni.algorithms.compression.tensorflow.pruning import LevelPruner, SlimPruner
+
+class LeNet(Model):
+    """
+    LeNet-5 Model with customizable hyper-parameters
+    """
+    def __init__(self, conv_size=3, hidden_size=32, dropout_rate=0.5):
+        """
+        Initialize hyper-parameters.
+
+        Parameters
+        ----------
+        conv_size : int
+            Kernel size of convolutional layers.
+        hidden_size : int
+            Dimensionality of last hidden layer.
+        dropout_rate : float
+            Dropout rate between two fully connected (dense) layers, to prevent co-adaptation.
+        """
+        super().__init__()
+        self.conv1 = Conv2D(filters=32, kernel_size=conv_size, activation='relu')
+        self.pool1 = MaxPool2D(pool_size=2)
+        self.bn1 = BatchNormalization()
+        self.conv2 = Conv2D(filters=64, kernel_size=conv_size, activation='relu')
+        self.pool2 = MaxPool2D(pool_size=2)
+        self.bn2 = BatchNormalization()
+        self.flatten = Flatten()
+        self.fc1 = Dense(units=hidden_size, activation='relu')
+        self.dropout = Dropout(rate=dropout_rate)
+        self.fc2 = Dense(units=10, activation='softmax')
+
+    def call(self, x):
+        """Override ``Model.call`` to build LeNet-5 model."""
+        x = self.conv1(x)
+        x = self.pool1(x)
+        x = self.bn1(x)
+        x = self.conv2(x)
+        x = self.pool2(x)
+        x = self.bn2(x)
+        x = self.flatten(x)
+        x = self.fc1(x)
+        x = self.dropout(x)
+        return self.fc2(x)
+
+
+def get_dataset(dataset_name='mnist'):
+    assert dataset_name == 'mnist'
+
+    (x_train, y_train), (x_test, y_test) = tf.keras.datasets.mnist.load_data()
+    x_train = x_train[..., tf.newaxis] / 255.0
+    x_test = x_test[..., tf.newaxis] / 255.0
+    return (x_train, y_train), (x_test, y_test)
+
+
+# def create_model(model_name='naive'):
+#     assert model_name == 'naive'
+#     return tf.keras.Sequential([
+#         tf.keras.layers.Conv2D(filters=20, kernel_size=5),
+#         tf.keras.layers.BatchNormalization(),
+#         tf.keras.layers.ReLU(),
+#         tf.keras.layers.MaxPool2D(pool_size=2),
+#         tf.keras.layers.Conv2D(filters=20, kernel_size=5),
+#         tf.keras.layers.BatchNormalization(),
+#         tf.keras.layers.ReLU(),
+#         tf.keras.layers.MaxPool2D(pool_size=2),
+#         tf.keras.layers.Flatten(),
+#         tf.keras.layers.Dense(units=500),
+#         tf.keras.layers.ReLU(),
+#         tf.keras.layers.Dense(units=10),
+#         tf.keras.layers.Softmax()
+#     ])
+
+def main(args):
+    train_set, test_set = get_dataset('mnist')
+    model = LeNet()
+
+    print('start training')
+
+    optimizer = tf.keras.optimizers.SGD(learning_rate=0.1, momentum=0.9, decay=1e-4)
+
+    if args.pruner_name == 'slim':
+        def slim_loss(y_true, y_pred):
+            loss_1 = tf.keras.losses.sparse_categorical_crossentropy(y_true=y_true, y_pred=y_pred)
+            weight_list = []
+            for layer in [model.bn1, model.bn2]:
+                weight_list.append([w for w in layer.weights if '/gamma:' in w.name][0].read_value())
+            loss_2 = 0.0001 * tf.reduce_sum([tf.reduce_sum(tf.abs(w)) for w in weight_list])
+            return loss_1 + loss_2
+        model.compile(
+            optimizer=optimizer,
+            loss=slim_loss,
+            metrics=['accuracy']
+        )
+    else:
+        model.compile(
+            optimizer=optimizer,
+            loss='sparse_categorical_crossentropy',
+            metrics=['accuracy']
+        )
+
+    model.fit(
+        train_set[0],
+        train_set[1],
+        batch_size=args.batch_size,
+        epochs=args.pretrain_epochs,
+        validation_data=test_set
+    )
+
+    print('start pruning')
+    optimizer_finetune = tf.keras.optimizers.SGD(learning_rate=0.001, momentum=0.9, decay=1e-4)
+
+    # create_pruner
+    if args.pruner_name == 'level':
+        prune_config = [{
+            'sparsity': args.sparsity,
+            'op_types': ['default'],
+        }]
+        pruner = LevelPruner(model, prune_config)
+    elif args.pruner_name == 'slim':
+        prune_config = [{
+            'sparsity': args.sparsity,
+            'op_types': ['BatchNormalization'],
+        }]
+        pruner = SlimPruner(model, prune_config)
+
+    model = pruner.compress()
+
+    model.compile(
+        optimizer=optimizer_finetune,
+        loss='sparse_categorical_crossentropy',
+        metrics=['accuracy'],
+        run_eagerly=True  # NOTE: Important, model compression does not work in graph mode!
+    )
+
+    # fine-tuning
+    model.fit(
+        train_set[0],
+        train_set[1],
+        batch_size=args.batch_size,
+        epochs=args.prune_epochs,
+        validation_data=test_set
+    )
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--pruner_name', type=str, default='level', choices=['level', 'slim'])
+    parser.add_argument('--batch-size', type=int, default=256)
+    parser.add_argument('--pretrain_epochs', type=int, default=10)
+    parser.add_argument('--prune_epochs', type=int, default=10)
+    parser.add_argument('--sparsity', type=float, default=0.5)
+
+    args = parser.parse_args()
+    main(args)

examples/model_compress/pruning/legacy/naive_prune_torch.py

+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+'''
+NNI example for quick start of pruning.
+In this example, we use level pruner to prune the LeNet on MNIST.
+'''
+
+import logging
+
+import argparse
+import torch
+import torch.nn.functional as F
+import torch.optim as optim
+from torchvision import datasets, transforms
+from torch.optim.lr_scheduler import StepLR
+
+from nni.algorithms.compression.pytorch.pruning import LevelPruner
+
+import sys
+sys.path.append('../models')
+from mnist.lenet import LeNet
+
+_logger = logging.getLogger('mnist_example')
+_logger.setLevel(logging.INFO)
+
+def train(args, model, device, train_loader, optimizer, epoch):
+    model.train()
+    for batch_idx, (data, target) in enumerate(train_loader):
+        data, target = data.to(device), target.to(device)
+        optimizer.zero_grad()
+        output = model(data)
+        loss = F.nll_loss(output, target)
+        loss.backward()
+        optimizer.step()
+        if batch_idx % args.log_interval == 0:
+            print('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}'.format(
+                epoch, batch_idx * len(data), len(train_loader.dataset),
+                100. * batch_idx / len(train_loader), loss.item()))
+            if args.dry_run:
+                break
+                
+
+def test(model, device, test_loader):
+    model.eval()
+    test_loss = 0
+    correct = 0
+    with torch.no_grad():
+        for data, target in test_loader:
+            data, target = data.to(device), target.to(device)
+            output = model(data)
+            test_loss += F.nll_loss(output, target, reduction='sum').item()
+            pred = output.argmax(dim=1, keepdim=True)
+            correct += pred.eq(target.view_as(pred)).sum().item()
+
+    test_loss /= len(test_loader.dataset)
+    acc = 100 * correct / len(test_loader.dataset)
+
+    print('\nTest set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%)\n'.format(
+        test_loss, correct, len(test_loader.dataset), acc))
+
+    return acc
+
+def main(args):
+    torch.manual_seed(args.seed)
+    use_cuda = not args.no_cuda and torch.cuda.is_available()
+
+    device = torch.device("cuda" if use_cuda else "cpu")
+
+    train_kwargs = {'batch_size': args.batch_size}
+    test_kwargs = {'batch_size': args.test_batch_size}
+    if use_cuda:
+        cuda_kwargs = {'num_workers': 1,
+                       'pin_memory': True,
+                       'shuffle': True}
+        train_kwargs.update(cuda_kwargs)
+        test_kwargs.update(cuda_kwargs)
+
+    transform=transforms.Compose([
+        transforms.ToTensor(),
+        transforms.Normalize((0.1307,), (0.3081,))
+        ])
+
+    dataset1 = datasets.MNIST('./data', train=True, download=True,
+                       transform=transform)
+    dataset2 = datasets.MNIST('./data', train=False,
+                       transform=transform)
+    train_loader = torch.utils.data.DataLoader(dataset1,**train_kwargs)
+    test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)
+
+    model = LeNet().to(device)
+    optimizer = optim.Adadelta(model.parameters(), lr=args.lr)
+
+    print('start pre-training')
+    scheduler = StepLR(optimizer, step_size=1, gamma=args.gamma)
+    for epoch in range(1, args.epochs + 1):
+        train(args, model, device, train_loader, optimizer, epoch)
+        test(model, device, test_loader)
+        scheduler.step()
+        
+    torch.save(model.state_dict(), "pretrain_mnist_lenet.pt")
+
+    print('start pruning')
+    optimizer_finetune = torch.optim.SGD(model.parameters(), lr=0.01)
+
+    # create pruner
+    prune_config = [{
+        'sparsity': args.sparsity,
+        'op_types': ['default'],
+    }]
+
+    pruner = LevelPruner(model, prune_config)
+    model = pruner.compress()
+
+    # fine-tuning
+    best_top1 = 0
+    for epoch in range(1, args.epochs + 1):
+        pruner.update_epoch(epoch)
+        train(args, model, device, train_loader, optimizer_finetune, epoch)
+        top1 = test(model, device, test_loader)
+
+        if top1 > best_top1:
+            best_top1 = top1
+            # Export the best model, 'model_path' stores state_dict of the pruned model,
+            # mask_path stores mask_dict of the pruned model
+            pruner.export_model(model_path='pruend_mnist_lenet.pt', mask_path='mask_mnist_lenet.pt')
+
+if __name__ == '__main__':
+     # Training settings
+    parser = argparse.ArgumentParser(description='PyTorch MNIST Example for model comporession')
+    parser.add_argument('--batch-size', type=int, default=64, metavar='N',
+                        help='input batch size for training (default: 64)')
+    parser.add_argument('--test-batch-size', type=int, default=1000, metavar='N',
+                        help='input batch size for testing (default: 1000)')
+    parser.add_argument('--epochs', type=int, default=10, metavar='N',
+                        help='number of epochs to train (default: 10)')
+    parser.add_argument('--lr', type=float, default=1.0, metavar='LR',
+                        help='learning rate (default: 1.0)')
+    parser.add_argument('--gamma', type=float, default=0.7, metavar='M',
+                        help='Learning rate step gamma (default: 0.7)')
+    parser.add_argument('--no-cuda', action='store_true', default=False,
+                        help='disables CUDA training')
+    parser.add_argument('--dry-run', action='store_true', default=False,
+                        help='quickly check a single pass')
+    parser.add_argument('--seed', type=int, default=1, metavar='S',
+                        help='random seed (default: 1)')
+    parser.add_argument('--log-interval', type=int, default=10, metavar='N',
+                        help='how many batches to wait before logging training status')
+    parser.add_argument('--sparsity', type=float, default=0.5,
+                        help='target overall target sparsity')
+    args = parser.parse_args()
+
+    main(args)

examples/model_compress/pruning/legacy/speedup/model_speedup.py

+import os
+import sys
+import argparse
+import time
+import torch
+
+from pathlib import Path
+sys.path.append(str(Path(__file__).absolute().parents[3] / 'models'))
+from cifar10.vgg import VGG
+from mnist.lenet import LeNet
+
+from nni.compression.pytorch import apply_compression_results, ModelSpeedup
+
+torch.manual_seed(0)
+use_mask = True
+use_speedup = True
+compare_results = True
+
+config = {
+    'apoz': {
+        'model_name': 'vgg16',
+        'input_shape': [64, 3, 32, 32],
+        'masks_file': './experiment_data/mask_vgg16_cifar10_apoz.pth'
+    },
+    'l1filter': {
+        'model_name': 'vgg16',
+        'input_shape': [64, 3, 32, 32],
+        'masks_file': './experiment_data/mask_vgg16_cifar10_l1filter.pth'
+    },
+    'fpgm': {
+        'model_name': 'vgg16',
+        'input_shape': [64, 3, 32, 32],
+        'masks_file': './experiment_data/mask_vgg16_cifar10_fpgm.pth'
+    },
+    'slim': {
+        'model_name': 'vgg19',
+        'input_shape': [64, 3, 32, 32],
+        'masks_file': './experiment_data/mask_vgg19_cifar10_slim.pth'
+    }
+}
+
+def model_inference(config):
+    masks_file = config['masks_file']
+    device = torch.device(
+        'cuda') if torch.cuda.is_available() else torch.device('cpu')
+
+    # device = torch.device(config['device'])
+    if config['model_name'] == 'vgg16':
+        model = VGG(depth=16)
+    elif config['model_name'] == 'vgg19':
+        model = VGG(depth=19)
+    elif config['model_name'] == 'lenet':
+        model = LeNet()
+
+    model.to(device)
+    model.eval()
+
+    dummy_input = torch.randn(config['input_shape']).to(device)
+    use_mask_out = use_speedup_out = None
+    # must run use_mask before use_speedup because use_speedup modify the model
+    if use_mask:
+        apply_compression_results(model, masks_file, device)
+        start = time.time()
+        for _ in range(32):
+            use_mask_out = model(dummy_input)
+        print('elapsed time when use mask: ', time.time() - start)
+    if use_speedup:
+        m_speedup = ModelSpeedup(model, dummy_input, masks_file, device)
+        m_speedup.speedup_model()
+        start = time.time()
+        for _ in range(32):
+            use_speedup_out = model(dummy_input)
+        print('elapsed time when use speedup: ', time.time() - start)
+    if compare_results:
+        if torch.allclose(use_mask_out, use_speedup_out, atol=1e-07):
+            print('the outputs from use_mask and use_speedup are the same')
+        else:
+            raise RuntimeError('the outputs from use_mask and use_speedup are different')
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser("speedup")
+    parser.add_argument("--example_name", type=str, default="slim", help="the name of pruning example")
+    parser.add_argument("--masks_file", type=str, default=None, help="the path of the masks file")
+    args = parser.parse_args()
+
+    if args.example_name != 'all':
+        if args.masks_file is not None:
+            config[args.example_name]['masks_file'] = args.masks_file
+        if not os.path.exists(config[args.example_name]['masks_file']):
+            msg = '{} does not exist! You should specify masks_file correctly, ' \
+                  'or use default one which is generated by model_prune_torch.py'
+            raise RuntimeError(msg.format(config[args.example_name]['masks_file']))
+        model_inference(config[args.example_name])
+    else:
+        model_inference(config['fpgm'])
+        model_inference(config['slim'])
+        model_inference(config['l1filter'])
+        model_inference(config['apoz'])

examples/model_compress/pruning/legacy/speedup/speedup_mobilnetv2.py

+import torch
+from torchvision.models import mobilenet_v2
+from nni.compression.pytorch import ModelSpeedup
+from nni.algorithms.compression.pytorch.pruning import L1FilterPruner
+
+
+model = mobilenet_v2(pretrained=True)
+dummy_input  = torch.rand(8, 3, 416, 416)
+
+cfg_list = [{'op_types':['Conv2d'], 'sparsity':0.5}]
+pruner = L1FilterPruner(model, cfg_list)
+pruner.compress()
+pruner.export_model('./model', './mask')
+# need call _unwrap_model if you want run the speedup on the same model
+pruner._unwrap_model()
+
+# Speedup the nanodet
+ms = ModelSpeedup(model, dummy_input, './mask')
+ms.speedup_model()
+
+model(dummy_input)
\ No newline at end of file