Introduction || Tensors || Autograd || Building Models || TensorBoard Support || Training Models || Model Understanding

Training with PyTorch

Follow along with the video below or on youtube.

Introduction

In past videos, we’ve discussed and demonstrated:

• Building models with the neural network layers and functions of the torch.nn module

• The mechanics of automated gradient computation, which is central to gradient-based model training

• Using TensorBoard to visualize training progress and other activities

In this video, we’ll be adding some new tools to your inventory:

• We’ll get familiar with the dataset and dataloader abstractions, and how they ease the process of feeding data to your model during a training loop

• We’ll discuss specific loss functions and when to use them

• We’ll look at PyTorch optimizers, which implement algorithms to adjust model weights based on the outcome of a loss function

Finally, we’ll pull all of these together and see a full PyTorch training loop in action.

Dataset and DataLoader

The Dataset and DataLoader classes encapsulate the process of pulling your data from storage and exposing it to your training loop in batches.

The Dataset is responsible for accessing and processing single instances of data.

The DataLoader pulls instances of data from the Dataset (either automatically or with a sampler that you define), collects them in batches, and returns them for consumption by your training loop. The DataLoader works with all kinds of datasets, regardless of the type of data they contain.

For this tutorial, we’ll be using the Fashion-MNIST dataset provided by TorchVision. We use torchvision.transforms.Normalize() to zero-center and normalize the distribution of the image tile content, and download both training and validation data splits.

import torch
import torchvision
import torchvision.transforms as transforms

# PyTorch TensorBoard support
from torch.utils.tensorboard import SummaryWriter
from datetime import datetime


transform = transforms.Compose(
    [transforms.ToTensor(),
    transforms.Normalize((0.5,), (0.5,))])

# Create datasets for training & validation, download if necessary
training_set = torchvision.datasets.FashionMNIST('./data', train=True, transform=transform, download=True)
validation_set = torchvision.datasets.FashionMNIST('./data', train=False, transform=transform, download=True)

# Create data loaders for our datasets; shuffle for training, not for validation
training_loader = torch.utils.data.DataLoader(training_set, batch_size=4, shuffle=True)
validation_loader = torch.utils.data.DataLoader(validation_set, batch_size=4, shuffle=False)

# Class labels
classes = ('T-shirt/top', 'Trouser', 'Pullover', 'Dress', 'Coat',
        'Sandal', 'Shirt', 'Sneaker', 'Bag', 'Ankle Boot')

# Report split sizes
print('Training set has {} instances'.format(len(training_set)))
print('Validation set has {} instances'.format(len(validation_set)))

Training set has 60000 instances
Validation set has 10000 instances

As always, let’s visualize the data as a sanity check:

import matplotlib.pyplot as plt
import numpy as np

# Helper function for inline image display
def matplotlib_imshow(img, one_channel=False):
    if one_channel:
        img = img.mean(dim=0)
    img = img / 2 + 0.5     # unnormalize
    npimg = img.numpy()
    if one_channel:
        plt.imshow(npimg, cmap="Greys")
    else:
        plt.imshow(np.transpose(npimg, (1, 2, 0)))

dataiter = iter(training_loader)
images, labels = next(dataiter)

# Create a grid from the images and show them
img_grid = torchvision.utils.make_grid(images)
matplotlib_imshow(img_grid, one_channel=True)
print('  '.join(classes[labels[j]] for j in range(4)))

Sandal  Sneaker  Coat  Sneaker

The Model

The model we’ll use in this example is a variant of LeNet-5 - it should be familiar if you’ve watched the previous videos in this series.

import torch.nn as nn
import torch.nn.functional as F

# PyTorch models inherit from torch.nn.Module
class GarmentClassifier(nn.Module):
    def __init__(self):
        super(GarmentClassifier, self).__init__()
        self.conv1 = nn.Conv2d(1, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 4 * 4, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 4 * 4)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x


model = GarmentClassifier()

Loss Function

For this example, we’ll be using a cross-entropy loss. For demonstration purposes, we’ll create batches of dummy output and label values, run them through the loss function, and examine the result.

loss_fn = torch.nn.CrossEntropyLoss()

# NB: Loss functions expect data in batches, so we're creating batches of 4
# Represents the model's confidence in each of the 10 classes for a given input
dummy_outputs = torch.rand(4, 10)
# Represents the correct class among the 10 being tested
dummy_labels = torch.tensor([1, 5, 3, 7])

print(dummy_outputs)
print(dummy_labels)

loss = loss_fn(dummy_outputs, dummy_labels)
print('Total loss for this batch: {}'.format(loss.item()))

tensor([[0.7026, 0.1489, 0.0065, 0.6841, 0.4166, 0.3980, 0.9849, 0.6701, 0.4601,
         0.8599],
        [0.7461, 0.3920, 0.9978, 0.0354, 0.9843, 0.0312, 0.5989, 0.2888, 0.8170,
         0.4150],
        [0.8408, 0.5368, 0.0059, 0.8931, 0.3942, 0.7349, 0.5500, 0.0074, 0.0554,
         0.1537],
        [0.7282, 0.8755, 0.3649, 0.4566, 0.8796, 0.2390, 0.9865, 0.7549, 0.9105,
         0.5427]])
tensor([1, 5, 3, 7])
Total loss for this batch: 2.428950309753418

Optimizer

For this example, we’ll be using simple stochastic gradient descent with momentum.

It can be instructive to try some variations on this optimization scheme:

• Learning rate determines the size of the steps the optimizer takes. What does a different learning rate do to the your training results, in terms of accuracy and convergence time?

• Momentum nudges the optimizer in the direction of strongest gradient over multiple steps. What does changing this value do to your results?

• Try some different optimization algorithms, such as averaged SGD, Adagrad, or Adam. How do your results differ?

# Optimizers specified in the torch.optim package
optimizer = torch.optim.SGD(model.parameters(), lr=0.001, momentum=0.9)

The Training Loop

Below, we have a function that performs one training epoch. It enumerates data from the DataLoader, and on each pass of the loop does the following:

• Gets a batch of training data from the DataLoader

• Zeros the optimizer’s gradients

• Performs an inference - that is, gets predictions from the model for an input batch

• Calculates the loss for that set of predictions vs. the labels on the dataset

• Calculates the backward gradients over the learning weights

• Tells the optimizer to perform one learning step - that is, adjust the model’s learning weights based on the observed gradients for this batch, according to the optimization algorithm we chose

• It reports on the loss for every 1000 batches.

• Finally, it reports the average per-batch loss for the last 1000 batches, for comparison with a validation run

def train_one_epoch(epoch_index, tb_writer):
    running_loss = 0.
    last_loss = 0.

    # Here, we use enumerate(training_loader) instead of
    # iter(training_loader) so that we can track the batch
    # index and do some intra-epoch reporting
    for i, data in enumerate(training_loader):
        # Every data instance is an input + label pair
        inputs, labels = data

        # Zero your gradients for every batch!
        optimizer.zero_grad()

        # Make predictions for this batch
        outputs = model(inputs)

        # Compute the loss and its gradients
        loss = loss_fn(outputs, labels)
        loss.backward()

        # Adjust learning weights
        optimizer.step()

        # Gather data and report
        running_loss += loss.item()
        if i % 1000 == 999:
            last_loss = running_loss / 1000 # loss per batch
            print('  batch {} loss: {}'.format(i + 1, last_loss))
            tb_x = epoch_index * len(training_loader) + i + 1
            tb_writer.add_scalar('Loss/train', last_loss, tb_x)
            running_loss = 0.

    return last_loss

Per-Epoch Activity

There are a couple of things we’ll want to do once per epoch:

• Perform validation by checking our relative loss on a set of data that was not used for training, and report this

• Save a copy of the model

Here, we’ll do our reporting in TensorBoard. This will require going to the command line to start TensorBoard, and opening it in another browser tab.

# Initializing in a separate cell so we can easily add more epochs to the same run
timestamp = datetime.now().strftime('%Y%m%d_%H%M%S')
writer = SummaryWriter('runs/fashion_trainer_{}'.format(timestamp))
epoch_number = 0

EPOCHS = 5

best_vloss = 1_000_000.

for epoch in range(EPOCHS):
    print('EPOCH {}:'.format(epoch_number + 1))

    # Make sure gradient tracking is on, and do a pass over the data
    model.train(True)
    avg_loss = train_one_epoch(epoch_number, writer)


    running_vloss = 0.0
    # Set the model to evaluation mode, disabling dropout and using population
    # statistics for batch normalization.
    model.eval()

    # Disable gradient computation and reduce memory consumption.
    with torch.no_grad():
        for i, vdata in enumerate(validation_loader):
            vinputs, vlabels = vdata
            voutputs = model(vinputs)
            vloss = loss_fn(voutputs, vlabels)
            running_vloss += vloss

    avg_vloss = running_vloss / (i + 1)
    print('LOSS train {} valid {}'.format(avg_loss, avg_vloss))

    # Log the running loss averaged per batch
    # for both training and validation
    writer.add_scalars('Training vs. Validation Loss',
                    { 'Training' : avg_loss, 'Validation' : avg_vloss },
                    epoch_number + 1)
    writer.flush()

    # Track best performance, and save the model's state
    if avg_vloss < best_vloss:
        best_vloss = avg_vloss
        model_path = 'model_{}_{}'.format(timestamp, epoch_number)
        torch.save(model.state_dict(), model_path)

    epoch_number += 1

EPOCH 1:
  batch 1000 loss: 1.6334228742718697
  batch 2000 loss: 0.8325267498586327
  batch 3000 loss: 0.7357839432582259
  batch 4000 loss: 0.6203198612027336
  batch 5000 loss: 0.5985794507635291
  batch 6000 loss: 0.5540174103826284
  batch 7000 loss: 0.5262376921508694
  batch 8000 loss: 0.49943213945179016
  batch 9000 loss: 0.4791228069158969
  batch 10000 loss: 0.48048832945676984
  batch 11000 loss: 0.4555416843077692
  batch 12000 loss: 0.4327957744560554
  batch 13000 loss: 0.4215374990614946
  batch 14000 loss: 0.4116538490946405
  batch 15000 loss: 0.42197128416545454
LOSS train 0.42197128416545454 valid 0.42573070526123047
EPOCH 2:
  batch 1000 loss: 0.391129400132515
  batch 2000 loss: 0.3594876553445356
  batch 3000 loss: 0.38024698745517527
  batch 4000 loss: 0.36007948506608956
  batch 5000 loss: 0.36804933209921
  batch 6000 loss: 0.36939992672245714
  batch 7000 loss: 0.380406323489733
  batch 8000 loss: 0.37918244087623315
  batch 9000 loss: 0.3311092873358284
  batch 10000 loss: 0.3492796658105071
  batch 11000 loss: 0.3560263846115849
  batch 12000 loss: 0.347208712031148
  batch 13000 loss: 0.3392629021731736
  batch 14000 loss: 0.35167778954130335
  batch 15000 loss: 0.3399065617335145
LOSS train 0.3399065617335145 valid 0.34940749406814575
EPOCH 3:
  batch 1000 loss: 0.32995265404548263
  batch 2000 loss: 0.2898239671024494
  batch 3000 loss: 0.3086822715619201
  batch 4000 loss: 0.3226545439936017
  batch 5000 loss: 0.3076599842306387
  batch 6000 loss: 0.3404498044280117
  batch 7000 loss: 0.31397077315032945
  batch 8000 loss: 0.2985334051818354
  batch 9000 loss: 0.31255276522028724
  batch 10000 loss: 0.31668289129548066
  batch 11000 loss: 0.3182725422651274
  batch 12000 loss: 0.3132900773736983
  batch 13000 loss: 0.3010104660175784
  batch 14000 loss: 0.2975951905332986
  batch 15000 loss: 0.3183587776140194
LOSS train 0.3183587776140194 valid 0.331748366355896
EPOCH 4:
  batch 1000 loss: 0.27423156507417296
  batch 2000 loss: 0.27768196498346515
  batch 3000 loss: 0.28617818430181796
  batch 4000 loss: 0.2953875640570623
  batch 5000 loss: 0.3034211081086105
  batch 6000 loss: 0.29898134227376433
  batch 7000 loss: 0.2829316312874653
  batch 8000 loss: 0.26064652617712275
  batch 9000 loss: 0.30588441241285547
  batch 10000 loss: 0.2790648038830368
  batch 11000 loss: 0.2757911273932623
  batch 12000 loss: 0.28895767328885447
  batch 13000 loss: 0.28029485422050765
  batch 14000 loss: 0.3011846994490734
  batch 15000 loss: 0.2972985631366173
LOSS train 0.2972985631366173 valid 0.3738856613636017
EPOCH 5:
  batch 1000 loss: 0.26176671784678185
  batch 2000 loss: 0.25637523230407167
  batch 3000 loss: 0.26169903839824654
  batch 4000 loss: 0.2785256232180291
  batch 5000 loss: 0.27159029314570216
  batch 6000 loss: 0.2748634670724714
  batch 7000 loss: 0.25473591768076587
  batch 8000 loss: 0.2906644816369808
  batch 9000 loss: 0.2861707231853361
  batch 10000 loss: 0.249603123359022
  batch 11000 loss: 0.27283034429119746
  batch 12000 loss: 0.2827443071613561
  batch 13000 loss: 0.27192901143227755
  batch 14000 loss: 0.24838552182626153
  batch 15000 loss: 0.2789099564121534
LOSS train 0.2789099564121534 valid 0.30068570375442505

To load a saved version of the model:

saved_model = GarmentClassifier()
saved_model.load_state_dict(torch.load(PATH))

Once you’ve loaded the model, it’s ready for whatever you need it for - more training, inference, or analysis.

Note that if your model has constructor parameters that affect model structure, you’ll need to provide them and configure the model identically to the state in which it was saved.

Other Resources

• Docs on the data utilities, including Dataset and DataLoader, at pytorch.org

• A note on the use of pinned memory for GPU training

• Documentation on the datasets available in TorchVision, TorchText, and TorchAudio

• Documentation on the loss functions available in PyTorch

• Documentation on the torch.optim package, which includes optimizers and related tools, such as learning rate scheduling

• A detailed tutorial on saving and loading models

• The Tutorials section of pytorch.org contains tutorials on a broad variety of training tasks, including classification in different domains, generative adversarial networks, reinforcement learning, and more