Author: Sasank Chilamkurthy <https://chsasank.github.io>
_
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 <https://cs231n.github.io/transfer-learning/>
__
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:
%matplotlib inline
# 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 numpy as np
import torchvision
from torchvision import datasets, models, transforms
import matplotlib.pyplot as plt
import time
import os
import copy
plt.ion() # interactive mode
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 <https://download.pytorch.org/tutorial/hymenoptera_data.zip>
_
and extract it to the current directory.
# dowload the data
!rm -rf hymenoptera_data*
!wget https://download.pytorch.org/tutorial/hymenoptera_data.zip
!unzip hymenoptera_data.zip
# Data augmentation and normalization for training
# Just normalization for validation
normalize_mean = [0.485, 0.456, 0.406]
normalize_std = [0.229, 0.224, 0.225]
data_transforms = {
'train': transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(normalize_mean, normalize_std)
]),
'val': transforms.Compose([
transforms.Resize(256),
transforms.CenterCrop(224),
transforms.ToTensor(),
transforms.Normalize(normalize_mean, normalize_std)
]),
}
data_dir = '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 test and training images so as to understand the data augmentations.
def imshow(inp, title=None, rescale=True):
"""Imshow for Tensor."""
if len(inp.shape)==2:
inp = np.tile(inp[None,:,:],(3,1,1))
if rescale:
inp = (inp - np.mean(inp)) / (4*np.std(inp)) + .5
inp = inp.transpose((1, 2, 0))
inp = np.clip(inp, 0, 1)
plt.imshow(inp)
if title is not None:
plt.title(title)
# Display a row of teest data and a row of train data
for data_type in 'val', 'train':
inputs, classes = next(iter(dataloaders[data_type]))
inputs = torchvision.utils.make_grid(inputs)
inputs = inputs.numpy() * np.array(normalize_std)[:,None,None] + np.array(normalize_mean)[:,None,None]
print(data_type,'data:')
plt.figure(figsize=(16,8))
imshow(inputs, title=[class_names[x] for x in classes])
plt.show()
val data:
train data:
Now, let's write a general function to train a model. Here, we will illustrate:
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('Epoch {}/{}'.format(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('{} Loss: {:.4f} Acc: {:.4f}'.format(
phase, epoch_loss, epoch_acc))
# 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('Training complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print('Best val Acc: {:4f}'.format(best_acc))
# 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=8):
was_training = model.training
model.eval()
images_so_far = 0
plt.figure(figsize=(16,8))
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
plt.subplot(num_images//4, 4, images_so_far)
plt.axis('off')
plt.title('predicted: {}'.format(class_names[preds[j]]))
inp = inputs.cpu().data[j].numpy() * np.array(normalize_std)[:,None,None] + np.array(normalize_mean)[:,None,None]
imshow(inp)
if images_so_far == num_images:
model.train(mode=was_training)
return
model.train(mode=was_training)
plt.show()
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)
Epoch 0/24 ---------- train Loss: 0.5480 Acc: 0.6967 val Loss: 0.1589 Acc: 0.9477 Epoch 1/24 ---------- train Loss: 0.5262 Acc: 0.7869 val Loss: 0.4204 Acc: 0.8431 Epoch 2/24 ---------- train Loss: 0.5985 Acc: 0.7541 val Loss: 0.2325 Acc: 0.9346 Epoch 3/24 ---------- train Loss: 0.4967 Acc: 0.8197 val Loss: 0.6398 Acc: 0.8170 Epoch 4/24 ---------- train Loss: 0.6347 Acc: 0.7992 val Loss: 0.2850 Acc: 0.8954 Epoch 5/24 ---------- train Loss: 0.4595 Acc: 0.7992 val Loss: 0.2748 Acc: 0.9020 Epoch 6/24 ---------- train Loss: 0.4351 Acc: 0.8238 val Loss: 0.2355 Acc: 0.9281 Epoch 7/24 ---------- train Loss: 0.3323 Acc: 0.8525 val Loss: 0.2540 Acc: 0.8889 Epoch 8/24 ---------- train Loss: 0.2437 Acc: 0.9098 val Loss: 0.2338 Acc: 0.9216 Epoch 9/24 ---------- train Loss: 0.2640 Acc: 0.8975 val Loss: 0.2067 Acc: 0.9346 Epoch 10/24 ---------- train Loss: 0.3221 Acc: 0.8730 val Loss: 0.2520 Acc: 0.8889 Epoch 11/24 ---------- train Loss: 0.2610 Acc: 0.9098 val Loss: 0.2462 Acc: 0.9020 Epoch 12/24 ---------- train Loss: 0.2633 Acc: 0.8852 val Loss: 0.1984 Acc: 0.9281 Epoch 13/24 ---------- train Loss: 0.3475 Acc: 0.8566 val Loss: 0.2112 Acc: 0.9020 Epoch 14/24 ---------- train Loss: 0.3400 Acc: 0.8607 val Loss: 0.2152 Acc: 0.9216 Epoch 15/24 ---------- train Loss: 0.2774 Acc: 0.8852 val Loss: 0.2070 Acc: 0.9281 Epoch 16/24 ---------- train Loss: 0.3748 Acc: 0.8443 val Loss: 0.2140 Acc: 0.9281 Epoch 17/24 ---------- train Loss: 0.2584 Acc: 0.8893 val Loss: 0.2109 Acc: 0.9281 Epoch 18/24 ---------- train Loss: 0.2574 Acc: 0.8893 val Loss: 0.2160 Acc: 0.9150 Epoch 19/24 ---------- train Loss: 0.2850 Acc: 0.8893 val Loss: 0.2133 Acc: 0.9085 Epoch 20/24 ---------- train Loss: 0.2969 Acc: 0.8811 val Loss: 0.2182 Acc: 0.9281 Epoch 21/24 ---------- train Loss: 0.3045 Acc: 0.8689 val Loss: 0.2203 Acc: 0.9216 Epoch 22/24 ---------- train Loss: 0.2784 Acc: 0.8730 val Loss: 0.2160 Acc: 0.9281 Epoch 23/24 ---------- train Loss: 0.3538 Acc: 0.8525 val Loss: 0.2224 Acc: 0.9150 Epoch 24/24 ---------- train Loss: 0.2920 Acc: 0.8811 val Loss: 0.2173 Acc: 0.9346 Training complete in 2m 19s Best val Acc: 0.947712
visualize_model(model_ft)
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 <https://pytorch.org/docs/notes/autograd.html#excluding-subgraphs-from-backward>
__.
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)
Epoch 0/24 ---------- train Loss: 0.5779 Acc: 0.6885 val Loss: 0.2990 Acc: 0.8824 Epoch 1/24 ---------- train Loss: 0.4138 Acc: 0.8033 val Loss: 0.3838 Acc: 0.8235 Epoch 2/24 ---------- train Loss: 0.3642 Acc: 0.8484 val Loss: 0.1846 Acc: 0.9412 Epoch 3/24 ---------- train Loss: 0.4944 Acc: 0.7705 val Loss: 0.1736 Acc: 0.9542 Epoch 4/24 ---------- train Loss: 0.4354 Acc: 0.7787 val Loss: 0.1898 Acc: 0.9477 Epoch 5/24 ---------- train Loss: 0.3721 Acc: 0.8156 val Loss: 0.2034 Acc: 0.9412 Epoch 6/24 ---------- train Loss: 0.5675 Acc: 0.7541 val Loss: 0.1895 Acc: 0.9477 Epoch 7/24 ---------- train Loss: 0.3493 Acc: 0.8197 val Loss: 0.1915 Acc: 0.9412 Epoch 8/24 ---------- train Loss: 0.3252 Acc: 0.8566 val Loss: 0.1998 Acc: 0.9412 Epoch 9/24 ---------- train Loss: 0.3123 Acc: 0.8525 val Loss: 0.1888 Acc: 0.9477 Epoch 10/24 ---------- train Loss: 0.3736 Acc: 0.8279 val Loss: 0.1985 Acc: 0.9412 Epoch 11/24 ---------- train Loss: 0.3022 Acc: 0.8484 val Loss: 0.2128 Acc: 0.9412 Epoch 12/24 ---------- train Loss: 0.3492 Acc: 0.8607 val Loss: 0.1887 Acc: 0.9412 Epoch 13/24 ---------- train Loss: 0.3045 Acc: 0.8566 val Loss: 0.2300 Acc: 0.9150 Epoch 14/24 ---------- train Loss: 0.3104 Acc: 0.8566 val Loss: 0.2081 Acc: 0.9346 Epoch 15/24 ---------- train Loss: 0.4159 Acc: 0.8074 val Loss: 0.2458 Acc: 0.9150 Epoch 16/24 ---------- train Loss: 0.3827 Acc: 0.8320 val Loss: 0.2186 Acc: 0.9412 Epoch 17/24 ---------- train Loss: 0.2844 Acc: 0.8852 val Loss: 0.2139 Acc: 0.9281 Epoch 18/24 ---------- train Loss: 0.3633 Acc: 0.8320 val Loss: 0.2081 Acc: 0.9346 Epoch 19/24 ---------- train Loss: 0.3385 Acc: 0.8811 val Loss: 0.2111 Acc: 0.9346 Epoch 20/24 ---------- train Loss: 0.2994 Acc: 0.8811 val Loss: 0.2622 Acc: 0.9216 Epoch 21/24 ---------- train Loss: 0.3227 Acc: 0.8484 val Loss: 0.2109 Acc: 0.9477 Epoch 22/24 ---------- train Loss: 0.3107 Acc: 0.8607 val Loss: 0.2298 Acc: 0.9346 Epoch 23/24 ---------- train Loss: 0.3471 Acc: 0.8648 val Loss: 0.2281 Acc: 0.9281 Epoch 24/24 ---------- train Loss: 0.2726 Acc: 0.8730 val Loss: 0.2385 Acc: 0.9216 Training complete in 1m 31s Best val Acc: 0.954248
visualize_model(model_conv)
plt.ioff()
plt.show()
If you would like to learn more about the applications of transfer learning,
checkout our Quantized Transfer Learning for Computer Vision Tutorial <https://pytorch.org/tutorials/intermediate/quantized_transfer_learning_tutorial.html>
_.
This part is adapted from: https://debuggercafe.com/visualizing-filters-and-feature-maps-in-convolutional-neural-networks-using-pytorch/
First we extract all filters from the neural networks:
model_weights = [] # we will save the conv layer weights in this list
conv_layers = [] # we will save the 49 conv layers in this list
# get all the model children as list
model_children = list(model_ft.children())
# counter to keep count of the conv layers
counter = 0
# append all the conv layers and their respective weights to the list
for i in range(len(model_children)):
if type(model_children[i]) == nn.Conv2d:
counter += 1
model_weights.append(model_children[i].weight)
conv_layers.append(model_children[i])
elif type(model_children[i]) == nn.Sequential:
for j in range(len(model_children[i])):
for child in model_children[i][j].children():
if type(child) == nn.Conv2d:
counter += 1
model_weights.append(child.weight)
conv_layers.append(child)
print(f"Total convolutional layers: {counter}")
# take a look at the conv layers and the respective weights
for weight, conv in zip(model_weights, conv_layers):
# print(f"WEIGHT: {weight} \nSHAPE: {weight.shape}")
print(f"CONV: {conv} ====> SHAPE: {weight.shape}")
Total convolutional layers: 17 CONV: Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False) ====> SHAPE: torch.Size([64, 3, 7, 7]) CONV: Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([64, 64, 3, 3]) CONV: Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([64, 64, 3, 3]) CONV: Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([64, 64, 3, 3]) CONV: Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([64, 64, 3, 3]) CONV: Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([128, 64, 3, 3]) CONV: Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([128, 128, 3, 3]) CONV: Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([128, 128, 3, 3]) CONV: Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([128, 128, 3, 3]) CONV: Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([256, 128, 3, 3]) CONV: Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([256, 256, 3, 3]) CONV: Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([256, 256, 3, 3]) CONV: Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([256, 256, 3, 3]) CONV: Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([512, 256, 3, 3]) CONV: Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([512, 512, 3, 3]) CONV: Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([512, 512, 3, 3]) CONV: Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False) ====> SHAPE: torch.Size([512, 512, 3, 3])
# visualize the first conv layer filters
# There are 3*64 filters (number of input channels times number of output channels)
# so we can visualize them as 64 RGB images
plt.figure(figsize=(20, 17))
for i, filter in enumerate(model_weights[0]):
plt.subplot(8, 8, i+1) # (8, 8) because in conv0 we have 64 filters
imshow(filter.detach().cpu().numpy())
plt.axis('off')
plt.show()
# Visualize the second conv layer filters
# Here there are 64*64 filters, so we visualize only one slice of 64 filters
input_channel = 0 # whatever between 0 and 63
plt.figure(figsize=(20, 17))
for i, filter in enumerate(model_weights[1]):
plt.subplot(8, 8, i+1)
imshow(filter[input_channel, :, :].detach().cpu().numpy())
plt.axis('off')
plt.show()
inputs, _ = next(iter(dataloaders['val']))
# select on image
inp = inputs[0]
# show the image
imshow(inp.cpu().numpy())
plt.show()
inp = inp.to(device)[None,:,:,:]
# pass the image through all the layers
results = [conv_layers[0](inp)]
for i in range(1, len(conv_layers)):
# pass the result from the last layer to the next layer
results.append(conv_layers[i](results[-1]))
# make a copy of the `results`
outputs = results
def VisuLayer(fmaps):
plt.figure(figsize=(30, 30))
layer_viz = torch.squeeze(fmaps.data, dim=0)
sz = int(np.sqrt(layer_viz.shape[0]))+1
for i, fmap in enumerate(layer_viz):
plt.subplot(sz, sz, i + 1)
imshow(fmap.cpu().numpy())
plt.axis("off")
plt.show()
# visualize all feature maps for all layers !!
for num_layer in range(len(outputs)):
print('Layer',num_layer,':')
VisuLayer(outputs[num_layer])
Layer 0 :
Layer 1 :
Layer 2 :
Layer 3 :
Layer 4 :
Layer 5 :
Layer 6 :
Layer 7 :