torchbench 0.0.18

Easily benchmark PyTorch models on selected tasks and datasets.

Installation

Requires Python 3.6+.

pip install torchbench

Usage

This library can be used together with the sotabench website, or standalone. Below we'll describe the usage with the sotabench website.

Steps to benchmark your model on the sotabench website:

1. Create a sotabench.py in the root of your repository. Below you can see an example sotabench.py file added to the torchvision repository to test one of its constituent models:

from torchbench.image_classification import ImageNet
from torchvision.models.resnet import resnext101_32x8d
import torchvision.transforms as transforms
import PIL

# Define the transforms need to convert ImageNet data to expected model input
normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
input_transform = transforms.Compose([
    transforms.Resize(256, PIL.Image.BICUBIC),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    normalize,
])

# Run the benchmark
ImageNet.benchmark(
    model=resnext101_32x8d(pretrained=True),
    paper_model_name='ResNeXt-101-32x8d',
    paper_arxiv_id='1611.05431',
    input_transform=input_transform,
    batch_size=256,
    num_gpu=1
)

2. Run it locally on your machine to verify it works:

python sotabench.py

Alternatively you can run the same logic within a Notebook if that is your preferred workflow.

3. Login and connect your repository to sotabench. After you connect your repository the website will re-evaluate your model on every commit, to ensure the model is working and results are up-to-date - including if you add additional models to the benchmark file.

You can also use the library without the sotabench website, by simply omitting step 3. In that case you also don't need to put in the paper details into the benchmark() method.

Benchmarks

Image Classification on ImageNet

Image Classification on ImageNet benchmark is implemented in the image_classification.ImageNet class.

Benchmarking Pipeline

1. The model is put into evaluation mode and sent to the device
2. The ImageNet validation dataset is loaded and transformed using input_transform
3. The dataset is put into a DataLoader with options batch_size and num_workers
4. The model and dataset are passed into an evaluation function for the task, along with an optional model_output_transform function that can transform the outputs after inference
5. The transformed output is compared to expected output and Top 1 and Top 5 accuracy are calculated

Expected Inputs/Outputs

• Model output (following model.forward() and optionally model_output_transform) should be a 2D torch.Tensor containing the model output; first dimension should be output for each example (length batch_size) and second dimension should be output for each class in ImageNet (length 1000).

Language Modelling on WikiText-103

Language Modelling on WikiText-103 benchmark is implemented in the language_modelling.WikiText103 class.

Benchmarking Pipeline

1. The model is put into evaluation mode and sent to the device
2. The WikiText-103 text dataset is loaded and:

• encoeded using encoder: this should be an object with an encode method that takes in raw text and produces a list of tokens, i.e. token_list = encoder.encode(rawtext). This is the same interface as in Hugging Face.

• takes a context length context_length (default 1024 - same as GPT-2).

3. The dataset is put into a DataLoader with options batch_size and num_workers
4. The model and dataset are passed into an evaluation function for the task, along with an optional model_output_transform function that can transform the outputs after inference. The expected output is logits.
5. The logits and labels are shifted to perform predictive language modelling, and the Perplexity metric is calculated.

Expected Inputs/Outputs

• Model output (following model.forward() and optionally model_output_transform) should be a 3D torch.Tensor containing the model output; first dimension should be output for each example (length batch_size), second dimension should be output for each token (length=context_length), third dimension should be output for each vocab (length = vocab size). This is the same interface as in Hugging Face.

More benchmarks coming soon...

Contributing

All contributions welcome!