Transfer Learning for Computer Vision Tutorial

Author: Sasank Chilamkurthy

In this tutorial, you will learn how to train a convolutional neural network for image classification using transfer learning. You can read more about the transfer learning at cs231n notes

Quoting these notes,

In practice, very few people train an entire Convolutional Network from scratch (with random initialization), because it is relatively rare to have a dataset of sufficient size. Instead, it is common to pretrain a ConvNet on a very large dataset (e.g. ImageNet, which contains 1.2 million images with 1000 categories), and then use the ConvNet either as an initialization or a fixed feature extractor for the task of interest.

These two major transfer learning scenarios look as follows:

• Finetuning the convnet: Instead of random initialization, we initialize the network with a pretrained network, like the one that is trained on imagenet 1000 dataset. Rest of the training looks as usual.
• ConvNet as fixed feature extractor: Here, we will freeze the weights for all of the network except that of the final fully connected layer. This last fully connected layer is replaced with a new one with random weights and only this layer is trained.

# License: BSD
# Author: Sasank Chilamkurthy

from __future__ import print_function, division

import torch
import torch.nn as nn
import torch.optim as optim
from torch.optim import lr_scheduler
import torch.backends.cudnn as cudnn
import numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy

cudnn.benchmark = True
plt.ion()   # interactive mode

Load Data

We will use torchvision and torch.utils.data packages for loading the data.

The problem we’re going to solve today is to train a model to classify ants and bees. We have about 120 training images each for ants and bees. There are 75 validation images for each class. Usually, this is a very small dataset to generalize upon, if trained from scratch. Since we are using transfer learning, we should be able to generalize reasonably well.

This dataset is a very small subset of imagenet.

Note

Download the data from here and extract it to the current directory.

# Data augmentation and normalization for training
# Just normalization for validation
data_transforms = {
    'train': transforms.Compose([
        transforms.RandomResizedCrop(224),
        transforms.RandomHorizontalFlip(),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
    'val': transforms.Compose([
        transforms.Resize(256),
        transforms.CenterCrop(224),
        transforms.ToTensor(),
        transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ]),
}

data_dir = 'data/hymenoptera_data'
image_datasets = {x: datasets.ImageFolder(os.path.join(data_dir, x),
                                          data_transforms[x])
                  for x in ['train', 'val']}
dataloaders = {x: torch.utils.data.DataLoader(image_datasets[x], batch_size=4,
                                             shuffle=True, num_workers=4)
              for x in ['train', 'val']}
dataset_sizes = {x: len(image_datasets[x]) for x in ['train', 'val']}
class_names = image_datasets['train'].classes

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

Visualize a few images

Let’s visualize a few training images so as to understand the data augmentations.

def imshow(inp, title=None):
    """Imshow for Tensor."""
    inp = inp.numpy().transpose((1, 2, 0))
    mean = np.array([0.485, 0.456, 0.406])
    std = np.array([0.229, 0.224, 0.225])
    inp = std * inp + mean
    inp = np.clip(inp, 0, 1)
    plt.imshow(inp)
    if title is not None:
        plt.title(title)
    plt.pause(0.001)  # pause a bit so that plots are updated


# Get a batch of training data
inputs, classes = next(iter(dataloaders['train']))

# Make a grid from batch
out = torchvision.utils.make_grid(inputs)

imshow(out, title=[class_names[x] for x in classes])

Training the model

Now, let’s write a general function to train a model. Here, we will illustrate:

• Scheduling the learning rate
• Saving the best model

In the following, parameter scheduler is an LR scheduler object from torch.optim.lr_scheduler.

def train_model(model, criterion, optimizer, scheduler, num_epochs=25):
    since = time.time()

    best_model_wts = copy.deepcopy(model.state_dict())
    best_acc = 0.0

    for epoch in range(num_epochs):
        print(f'Epoch {epoch}/{num_epochs - 1}')
        print('-' * 10)

        # Each epoch has a training and validation phase
        for phase in ['train', 'val']:
            if phase == 'train':
                model.train()  # Set model to training mode
            else:
                model.eval()   # Set model to evaluate mode

            running_loss = 0.0
            running_corrects = 0

            # Iterate over data.
            for inputs, labels in dataloaders[phase]:
                inputs = inputs.to(device)
                labels = labels.to(device)

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                # track history if only in train
                with torch.set_grad_enabled(phase == 'train'):
                    outputs = model(inputs)
                    _, preds = torch.max(outputs, 1)
                    loss = criterion(outputs, labels)

                    # backward + optimize only if in training phase
                    if phase == 'train':
                        loss.backward()
                        optimizer.step()

                # statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += torch.sum(preds == labels.data)
            if phase == 'train':
                scheduler.step()

            epoch_loss = running_loss / dataset_sizes[phase]
            epoch_acc = running_corrects.double() / dataset_sizes[phase]

            print(f'{phase} Loss: {epoch_loss:.4f} Acc: {epoch_acc:.4f}')

            # deep copy the model
            if phase == 'val' and epoch_acc > best_acc:
                best_acc = epoch_acc
                best_model_wts = copy.deepcopy(model.state_dict())

        print()

    time_elapsed = time.time() - since
    print(f'Training complete in {time_elapsed // 60:.0f}m {time_elapsed % 60:.0f}s')
    print(f'Best val Acc: {best_acc:4f}')

    # load best model weights
    model.load_state_dict(best_model_wts)
    return model

Visualizing the model predictions

Generic function to display predictions for a few images

def visualize_model(model, num_images=6):
    was_training = model.training
    model.eval()
    images_so_far = 0
    fig = plt.figure()

    with torch.no_grad():
        for i, (inputs, labels) in enumerate(dataloaders['val']):
            inputs = inputs.to(device)
            labels = labels.to(device)

            outputs = model(inputs)
            _, preds = torch.max(outputs, 1)

            for j in range(inputs.size()[0]):
                images_so_far += 1
                ax = plt.subplot(num_images//2, 2, images_so_far)
                ax.axis('off')
                ax.set_title(f'predicted: {class_names[preds[j]]}')
                imshow(inputs.cpu().data[j])

                if images_so_far == num_images:
                    model.train(mode=was_training)
                    return
        model.train(mode=was_training)

Finetuning the convnet

Load a pretrained model and reset final fully connected layer.

model_ft = models.resnet18(pretrained=True)
num_ftrs = model_ft.fc.in_features
# Here the size of each output sample is set to 2.
# Alternatively, it can be generalized to nn.Linear(num_ftrs, len(class_names)).
model_ft.fc = nn.Linear(num_ftrs, 2)

model_ft = model_ft.to(device)

criterion = nn.CrossEntropyLoss()

# Observe that all parameters are being optimized
optimizer_ft = optim.SGD(model_ft.parameters(), lr=0.001, momentum=0.9)

# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_ft, step_size=7, gamma=0.1)

Train and evaluate

It should take around 15-25 min on CPU. On GPU though, it takes less than a minute.

model_ft = train_model(model_ft, criterion, optimizer_ft, exp_lr_scheduler,
                       num_epochs=25)

Out:

Epoch 0/24
----------
train Loss: 0.6114 Acc: 0.6967
val Loss: 0.5982 Acc: 0.7190

Epoch 1/24
----------
train Loss: 0.4216 Acc: 0.8361
val Loss: 0.3403 Acc: 0.8562

Epoch 2/24
----------
train Loss: 0.6654 Acc: 0.6803
val Loss: 0.7903 Acc: 0.6601

Epoch 3/24
----------
train Loss: 0.4905 Acc: 0.8033
val Loss: 0.4686 Acc: 0.8562

Epoch 4/24
----------
train Loss: 0.4547 Acc: 0.8197
val Loss: 0.3586 Acc: 0.8497

Epoch 5/24
----------
train Loss: 0.4229 Acc: 0.8320
val Loss: 0.1841 Acc: 0.9346

Epoch 6/24
----------
train Loss: 0.5599 Acc: 0.7951
val Loss: 0.1951 Acc: 0.9281

Epoch 7/24
----------
train Loss: 0.3074 Acc: 0.8689
val Loss: 0.1726 Acc: 0.9477

Epoch 8/24
----------
train Loss: 0.2723 Acc: 0.8811
val Loss: 0.2053 Acc: 0.9281

Epoch 9/24
----------
train Loss: 0.3379 Acc: 0.8730
val Loss: 0.2019 Acc: 0.9346

Epoch 10/24
----------
train Loss: 0.3513 Acc: 0.8689
val Loss: 0.2019 Acc: 0.9412

Epoch 11/24
----------
train Loss: 0.2054 Acc: 0.9344
val Loss: 0.2025 Acc: 0.9477

Epoch 12/24
----------
train Loss: 0.3692 Acc: 0.8443
val Loss: 0.2086 Acc: 0.9346

Epoch 13/24
----------
train Loss: 0.2468 Acc: 0.8975
val Loss: 0.2337 Acc: 0.8954

Epoch 14/24
----------
train Loss: 0.3197 Acc: 0.8648
val Loss: 0.2432 Acc: 0.9346

Epoch 15/24
----------
train Loss: 0.2706 Acc: 0.8852
val Loss: 0.2013 Acc: 0.9281

Epoch 16/24
----------
train Loss: 0.2445 Acc: 0.9139
val Loss: 0.1969 Acc: 0.9346

Epoch 17/24
----------
train Loss: 0.3132 Acc: 0.8279
val Loss: 0.2041 Acc: 0.9346

Epoch 18/24
----------
train Loss: 0.2742 Acc: 0.8770
val Loss: 0.1941 Acc: 0.9477

Epoch 19/24
----------
train Loss: 0.2991 Acc: 0.8770
val Loss: 0.1983 Acc: 0.9412

Epoch 20/24
----------
train Loss: 0.3321 Acc: 0.8689
val Loss: 0.2100 Acc: 0.9216

Epoch 21/24
----------
train Loss: 0.2930 Acc: 0.8770
val Loss: 0.2098 Acc: 0.9346

Epoch 22/24
----------
train Loss: 0.2786 Acc: 0.8770
val Loss: 0.2036 Acc: 0.9412

Epoch 23/24
----------
train Loss: 0.2645 Acc: 0.8975
val Loss: 0.1924 Acc: 0.9412

Epoch 24/24
----------
train Loss: 0.2697 Acc: 0.8607
val Loss: 0.2012 Acc: 0.9542

Training complete in 1m 9s
Best val Acc: 0.954248

visualize_model(model_ft)

ConvNet as fixed feature extractor

Here, we need to freeze all the network except the final layer. We need to set requires_grad = False to freeze the parameters so that the gradients are not computed in backward().

You can read more about this in the documentation here.

model_conv = torchvision.models.resnet18(pretrained=True)
for param in model_conv.parameters():
    param.requires_grad = False

# Parameters of newly constructed modules have requires_grad=True by default
num_ftrs = model_conv.fc.in_features
model_conv.fc = nn.Linear(num_ftrs, 2)

model_conv = model_conv.to(device)

criterion = nn.CrossEntropyLoss()

# Observe that only parameters of final layer are being optimized as
# opposed to before.
optimizer_conv = optim.SGD(model_conv.fc.parameters(), lr=0.001, momentum=0.9)

# Decay LR by a factor of 0.1 every 7 epochs
exp_lr_scheduler = lr_scheduler.StepLR(optimizer_conv, step_size=7, gamma=0.1)

Train and evaluate

On CPU this will take about half the time compared to previous scenario. This is expected as gradients don’t need to be computed for most of the network. However, forward does need to be computed.

model_conv = train_model(model_conv, criterion, optimizer_conv,
                         exp_lr_scheduler, num_epochs=25)

Out:

Epoch 0/24
----------
train Loss: 0.8283 Acc: 0.6680
val Loss: 0.2313 Acc: 0.9412

Epoch 1/24
----------
train Loss: 0.3411 Acc: 0.8566
val Loss: 0.2130 Acc: 0.9085

Epoch 2/24
----------
train Loss: 0.4557 Acc: 0.7951
val Loss: 0.2120 Acc: 0.9346

Epoch 3/24
----------
train Loss: 0.3622 Acc: 0.8402
val Loss: 0.2383 Acc: 0.9085

Epoch 4/24
----------
train Loss: 0.5417 Acc: 0.7623
val Loss: 0.1973 Acc: 0.9412

Epoch 5/24
----------
train Loss: 0.4069 Acc: 0.8402
val Loss: 0.2052 Acc: 0.9477

Epoch 6/24
----------
train Loss: 0.4072 Acc: 0.7869
val Loss: 0.2844 Acc: 0.8824

Epoch 7/24
----------
train Loss: 0.3897 Acc: 0.8361
val Loss: 0.1749 Acc: 0.9477

Epoch 8/24
----------
train Loss: 0.3548 Acc: 0.8197
val Loss: 0.1757 Acc: 0.9477

Epoch 9/24
----------
train Loss: 0.3292 Acc: 0.8484
val Loss: 0.1728 Acc: 0.9477

Epoch 10/24
----------
train Loss: 0.4108 Acc: 0.8115
val Loss: 0.1748 Acc: 0.9412

Epoch 11/24
----------
train Loss: 0.4019 Acc: 0.8033
val Loss: 0.1815 Acc: 0.9477

Epoch 12/24
----------
train Loss: 0.4609 Acc: 0.7869
val Loss: 0.1933 Acc: 0.9281

Epoch 13/24
----------
train Loss: 0.4383 Acc: 0.7828
val Loss: 0.1774 Acc: 0.9477

Epoch 14/24
----------
train Loss: 0.2799 Acc: 0.8730
val Loss: 0.1831 Acc: 0.9412

Epoch 15/24
----------
train Loss: 0.3141 Acc: 0.8443
val Loss: 0.1811 Acc: 0.9477

Epoch 16/24
----------
train Loss: 0.2609 Acc: 0.9139
val Loss: 0.1956 Acc: 0.9346

Epoch 17/24
----------
train Loss: 0.3234 Acc: 0.8279
val Loss: 0.1788 Acc: 0.9477

Epoch 18/24
----------
train Loss: 0.3325 Acc: 0.8607
val Loss: 0.1541 Acc: 0.9477

Epoch 19/24
----------
train Loss: 0.3555 Acc: 0.8361
val Loss: 0.1735 Acc: 0.9477

Epoch 20/24
----------
train Loss: 0.3300 Acc: 0.8443
val Loss: 0.1767 Acc: 0.9608

Epoch 21/24
----------
train Loss: 0.3155 Acc: 0.8607
val Loss: 0.1737 Acc: 0.9477

Epoch 22/24
----------
train Loss: 0.3697 Acc: 0.8443
val Loss: 0.1803 Acc: 0.9281

Epoch 23/24
----------
train Loss: 0.2814 Acc: 0.8648
val Loss: 0.1667 Acc: 0.9477

Epoch 24/24
----------
train Loss: 0.3470 Acc: 0.8525
val Loss: 0.2084 Acc: 0.9281

Training complete in 0m 41s
Best val Acc: 0.960784

visualize_model(model_conv)

plt.ioff()
plt.show()

Further Learning

If you would like to learn more about the applications of transfer learning, checkout our Quantized Transfer Learning for Computer Vision Tutorial.