metabatch 0.9.1

MetaBatch: A micro-framework for efficient batching of tasks in PyTorch.

Introduction

MetaBatch provides convenient Taskset and TaskLoader classes for batch-aware online task creation for meta-learning.

Efficient batching

Training meta-learning models efficiently can be a challenge, especially when it comes to creating random tasks of a consistent shape in one batch. The task creation process can be time-consuming and typically requires all tasks in the batch to have the same amount of context and target points. This can be a bottleneck during training:

# Sample code for creating a batch of tasks with traditional approach
class MyTaskDataset(Dataset):
    ...
    def __getitem__(self, idx):
        task = self.task_data[idx]
        return task

class Model(Module):
    ...
    def forward(self, tasks):
        ctx_batch = tasks['context']
        tgt_batch = tasks['target']
        ...

# create dataset
task_data = [{'images': [...], 'label': 'dog'},
             {'images': [...], 'label': 'cat'}, ...]
dataset = MyTaskDataset(task_data)
dataloader = DataLoader(dataset, batch_size=16, workers=8)

for batch in dataloader:
    ...
    # Construct batch of random tasks in the training loop (bottleneck!)
    n_context = random.randint(low=1, high=5)
    n_target = random.randint(low=1, high=10)
    tasks = {'context': [], 'target': []}
    for task in batch:
        context_images = sample_n_images(task['images'], n_context)
        target_images = sample_n_images(task['images'], n_target)
        tasks['context'].append(context_images)
        tasks['target'].append(target_images)
    model(tasks)
    ...

Multiprocessing

Wouldn't it be better to offload the task creation to the dataloader, so that it can be done in parallel on multiple cores? With MetaBatch, we simplify the process by allowing you to do just that. We provide a TaskSet wrapper, where you can implement the __gettask__(self, index, n_context, n_target)__ method instead of PyTorch's __getitem(self, index)__. Our TaskLoader and custom sampler take care of synchronizing n_context and n_target for each batch element dispatched to all workers. With MetaBatch, the training bottleneck can be removed from the above example:

# Sample code for creating a batch of tasks with MetaBatch
from metabatch import TaskSet, TaskLoader

class MyTaskSet(TaskSet):
    ...
    def __gettask__(self, idx, n_context, n_target):
        data = self.task_data[idx]
        context_images = sample_n_images(data['images'], n_context)
        target_images = sample_n_images(data['images'], n_target)
        return {
            'context': context_images
            'target': target_images
        }

class Model(Module):
    ...
    def forward(self, tasks):
        ctx_batch = tasks['context']
        tgt_batch = tasks['target']
        ...

# create dataset
task_data = [{'images': [...], 'label': 'dog'},
             {'images': [...], 'label': 'cat'}, ...]
dataset = MyTaskSet(task_data, min_pts=1, max_ctx_pts=5, max_tgt_pts=10)
dataloader = TaskLoader(dataset, batch_size=16, workers=8)

for batch in dataloader:
    ...
    # Simply access the batch of constructed tasks (no bottleneck!)
    model(batch)
    ...

Installation & usage

Install it: pip install metabatch

Requirements:

• pytorch

Look at the example above for an idea or how to use TaskLoader with TaskSet, or go through the examples in examples/ (TODO).

Advantages

• MetaBatch allows for efficient task creation and batching during training, resulting in more task variations since you are no longer limited to precomputed tasks.
• Reduces boilerplate needed to precompute and load tasks.

MetaBatch is a micro-framework for meta-learning in PyTorch that provides convenient tools for (potentially faster) meta-training. It simplifies the task creation process and allows for efficient batching, making it a useful tool for researchers and engineers working on meta-learning projects.

How much faster?

TODO: benchmark MAML and CNP examples with typical implementation and other repos.

License

MetaBatch is released under the MIT License. See the LICENSE file for more information.