Finetuning a Pytorch Image Classifier with Ray Train#

This example fine tunes a pre-trained ResNet model with Ray Train.

For this example, the network architecture consists of the intermediate layer output of a pre-trained ResNet model, which feeds into a randomly initialized linear layer that outputs classification logits for our new task.

Load and preprocess finetuning dataset#

This example is adapted from Pytorch’s Finetuning Torchvision Models tutorial. We will use hymenoptera_data as the finetuning dataset, which contains two classes (bees and ants) and 397 total images (across training and validation). This is a quite small dataset and used only for demonstration purposes.

The dataset is publicly available here. Note that it is structured with directory names as the labels. Use torchvision.datasets.ImageFolder() to load the images and their corresponding labels.

import os
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets, models, transforms
import numpy as np

# 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(224),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
        ]
    ),
}

def download_datasets():
    os.system(
        "wget https://download.pytorch.org/tutorial/hymenoptera_data.zip >/dev/null 2>&1"
    )
    os.system("unzip hymenoptera_data.zip >/dev/null 2>&1")

# Download and build torch datasets
def build_datasets():
    torch_datasets = {}
    for split in ["train", "val"]:
        torch_datasets[split] = datasets.ImageFolder(
            os.path.join("./hymenoptera_data", split), data_transforms[split]
        )
    return torch_datasets

Initialize Model and Fine-tuning configs#

Next, let’s define the training configuration that will be passed into the training loop function later.

train_loop_config = {
    "input_size": 224,  # Input image size (224 x 224)
    "batch_size": 32,  # Batch size for training
    "num_epochs": 10,  # Number of epochs to train for
    "lr": 0.001,  # Learning Rate
    "momentum": 0.9,  # SGD optimizer momentum
}

Next, let’s define our model. You can either create a model from pre-trained weights or reload the model checkpoint from a previous run.

import os
import torch
from ray.train import Checkpoint

# Option 1: Initialize model with pretrained weights
def initialize_model():
    # Load pretrained model params
    model = models.resnet50(pretrained=True)

    # Replace the original classifier with a new Linear layer
    num_features = model.fc.in_features
    model.fc = nn.Linear(num_features, 2)

    # Ensure all params get updated during finetuning
    for param in model.parameters():
        param.requires_grad = True
    return model


# Option 2: Initialize model with an Train checkpoint
# Replace this with your own uri
CHECKPOINT_FROM_S3 = Checkpoint(
    path="s3://air-example-data/finetune-resnet-checkpoint/TorchTrainer_4f69f_00000_0_2023-02-14_14-04-09/checkpoint_000001/"
)


def initialize_model_from_checkpoint(checkpoint: Checkpoint):
    with checkpoint.as_directory() as tmpdir:
        state_dict = torch.load(os.path.join(tmpdir, "checkpoint.pt"))
    resnet50 = initialize_model()
    resnet50.load_state_dict(state_dict["model"])
    return resnet50

Define the Training Loop#

The train_loop_per_worker function defines the fine-tuning procedure for each worker.

1. Prepare dataloaders for each worker:

  • This tutorial assumes you are using PyTorch’s native torch.utils.data.Dataset for data input. train.torch.prepare_data_loader() prepares your dataLoader for distributed execution. You can also use Ray Data for more efficient preprocessing. For more details on using Ray Data for for images, see the Working with Images Ray Data user guide.

2. Prepare your model:

  • train.torch.prepare_model() prepares the model for distributed training. Under the hood, it converts your torch model to DistributedDataParallel model, which synchronize its weights across all workers.

3. Report metrics and checkpoint:

import os
from tempfile import TemporaryDirectory

import ray.train as train
from ray.train import Checkpoint



def evaluate(logits, labels):
    _, preds = torch.max(logits, 1)
    corrects = torch.sum(preds == labels).item()
    return corrects


def train_loop_per_worker(configs):
    import warnings

    warnings.filterwarnings("ignore")

    # Calculate the batch size for a single worker
    worker_batch_size = configs["batch_size"] // train.get_context().get_world_size()

    # Download dataset once on local rank 0 worker
    if train.get_context().get_local_rank() == 0:
        download_datasets()
    torch.distributed.barrier()

    # Build datasets on each worker
    torch_datasets = build_datasets()

    # Prepare dataloader for each worker
    dataloaders = dict()
    dataloaders["train"] = DataLoader(
        torch_datasets["train"], batch_size=worker_batch_size, shuffle=True
    )
    dataloaders["val"] = DataLoader(
        torch_datasets["val"], batch_size=worker_batch_size, shuffle=False
    )

    # Distribute
    dataloaders["train"] = train.torch.prepare_data_loader(dataloaders["train"])
    dataloaders["val"] = train.torch.prepare_data_loader(dataloaders["val"])

    device = train.torch.get_device()

    # Prepare DDP Model, optimizer, and loss function
    model = initialize_model()
    model = train.torch.prepare_model(model)

    optimizer = optim.SGD(
        model.parameters(), lr=configs["lr"], momentum=configs["momentum"]
    )
    criterion = nn.CrossEntropyLoss()

    # Start training loops
    for epoch in range(configs["num_epochs"]):
        # 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

            if train.get_context().get_world_size() > 1:
                dataloaders[phase].sampler.set_epoch(epoch)

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

                # zero the parameter gradients
                optimizer.zero_grad()

                # forward
                with torch.set_grad_enabled(phase == "train"):
                    # Get model outputs and calculate loss
                    outputs = model(inputs)
                    loss = criterion(outputs, labels)

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

                # calculate statistics
                running_loss += loss.item() * inputs.size(0)
                running_corrects += evaluate(outputs, labels)

            size = len(torch_datasets[phase]) // train.get_context().get_world_size()
            epoch_loss = running_loss / size
            epoch_acc = running_corrects / size

            if train.get_context().get_world_rank() == 0:
                print(
                    "Epoch {}-{} Loss: {:.4f} Acc: {:.4f}".format(
                        epoch, phase, epoch_loss, epoch_acc
                    )
                )

            # Report metrics and checkpoint every epoch
            if phase == "val":
                with TemporaryDirectory() as tmpdir:
                    state_dict = {
                        "epoch": epoch,
                        "model": model.module.state_dict(),
                        "optimizer_state_dict": optimizer.state_dict(),
                    }
                    torch.save(state_dict, os.path.join(tmpdir, "checkpoint.pt"))
                    train.report(
                        metrics={"loss": epoch_loss, "acc": epoch_acc},
                        checkpoint=Checkpoint.from_directory(tmpdir),
                    )

Next, setup the TorchTrainer:

from ray.train.torch import TorchTrainer
from ray.train import ScalingConfig, RunConfig, CheckpointConfig

# Scale out model training across 4 GPUs.
scaling_config = ScalingConfig(
    num_workers=4, use_gpu=True, resources_per_worker={"CPU": 1, "GPU": 1}
)

# Save the latest checkpoint
checkpoint_config = CheckpointConfig(num_to_keep=1)

# Set experiment name and checkpoint configs
run_config = RunConfig(
    name="finetune-resnet",
    storage_path="/tmp/ray_results",
    checkpoint_config=checkpoint_config,
)

trainer = TorchTrainer(
    train_loop_per_worker=train_loop_per_worker,
    train_loop_config=train_loop_config,
    scaling_config=scaling_config,
    run_config=run_config,
)

result = trainer.fit()
print(result)

Load the checkpoint for prediction:#

The metadata and checkpoints have already been saved in the storage_path specified in TorchTrainer:

We now need to load the trained model and evaluate it on test data. The best model parameters have been saved in log_dir. We can load the resulting checkpoint from our fine-tuning run using the previously defined initialize_model_from_checkpoint() function.

model = initialize_model_from_checkpoint(result.checkpoint)
device = torch.device("cuda")

Finally, define a simple evaluation loop and check the performance of the checkpoint model.

model = model.to(device)
model.eval()

download_datasets()
torch_datasets = build_datasets()
dataloader = DataLoader(torch_datasets["val"], batch_size=32, num_workers=4)
corrects = 0
for inputs, labels in dataloader:
    inputs = inputs.to(device)
    labels = labels.to(device)
    preds = model(inputs)
    corrects += evaluate(preds, labels)

print("Accuracy: ", corrects / len(dataloader.dataset))

Accuracy:  0.934640522875817