aibench 0.0.5

No project description provided

 

AIBenchmark

Benchmark your model against other models

About   |   Features   |   Technologies   |   Starting   |   License   |   Author

Installation

Run this script in your terminal:

$ pip install aibench

About

AIBenchmark is a package which lets you quickly get the benchmark of your model based on the popular datasets and compare with existing leaderboard. It also has a nice collection of metrics which you could easily import.

We currently support 14 text-based and 2 image-based datasets for AutoBenchmarking aiming for regression/classification tasks. Available datasets could be found in aibenchmark/dataset.py file.

Or run the following code:

from aibenchmark.dataset import DatasetsList

print(list(DatasetsList.get_available_datasets()))

Code example for benchmarking:

from aibenchmark.benchmark import Benchmark
from aibenchmark.dataset import DatasetInfo, DatasetsList


benchmark = Benchmark(DatasetsList.Texts.SST)
dataset_info: DatasetInfo = benchmark.dataset_info
print(dataset_info)

test_features = dataset_info.data['Texts']
model = torch.load(...)
# Implement your code based on the type of model you use, your pre- and post-processing etc.
outputs = model.predict(test_features)

# Results of your model based on predictions
benchmark_results = benchmark.run(predictions=outputs, metrics=['accuracy', 'precision', 'recall', 'f1_score']) 

# Metrics
print(benchmark_results)
# Existing leaderboard for this dataset
print(benchmark.get_existing_benchmarks())

Features

1. Fast comparison of metrics of your model and other SOTA models for particular dataset
2. Supporting 16+ most populat datasets, the list is always updating. Soon we willl support more than 1000 datasets
3. All metrics in one place and we are adding new ones in a standardised way

Technologies

The following tools were used in this project:

• Pytorch
• Transformers
• Scikit-learn

:memo: License

This project is under license from MIT. For more details, see the LICENSE file.

Made by Igor and Tim

 

Back to top