Perceiver IO
Project description
Perceiver IO
A PyTorch implementation of
- Perceiver: General Perception with Iterative Attention
- Perceiver IO: A General Architecture for Structured Inputs & Outputs
This project supports training of Perceiver IO models with Pytorch Lightning. Training examples are given in section Tasks, inference examples in section Notebooks. Perceiver IO models are constructed with generic encoder and decoder classes and task-specific input and output adapters (see Model API). The command line interface is implemented with Lighting CLI.
Setup
From sources
conda env create -f environment.yml
conda activate perceiver-io
poetry install
Via pip
pip install perceiver-io
When installing via pip
make sure you have a CUDA toolkit installed as well (see environment.yml
for details).
Tasks
In the following subsections, Perceiver IO models are trained on a rather small scale. In particular, hyperparameters are set such that parallel training on two NVIDIA GTX 1080 GPUs (8 GB memory each) works quite well. I didn't really tune model architectures and other hyperparameters, so you'll probably get better results with a bit of experimentation. Support for more datasets and tasks will be added later.
Masked language modeling
Pretrain a Perceiver IO model on masked language modeling (MLM) with text from the IMDB training set. The pretrained encoder is then used for training a sentiment classification model. Predictions of masked tokens are logged to Tensorboard.
python -m perceiver.scripts.mlm fit \
--model.num_latent_channels=64 \
--model.encoder.num_layers=3 \
--model.encoder.dropout=0.0 \
--model.decoder.dropout=0.0 \
--data=IMDBDataModule \
--data.max_seq_len=512 \
--data.batch_size=64 \
--optimizer.lr=3e-3 \
--optimizer.weight_decay=0.0 \
--lr_scheduler.pct_start=0.1 \
--trainer.default_root_dir=logs \
--trainer.accelerator=gpu \
--trainer.devices=-1 \
--trainer.max_steps=50000 \
--trainer.check_val_every_n_epoch=5
For saving GPU memory and scaling model training, activation checkpointing can be enabled with
--model.activation_checkpoint=true
(disabled by default).
Sentiment classification
Train a classification decoder using a frozen encoder from masked language modeling.
If you ran MLM yourself you'll need to modify the --model.mlm_ckpt
argument accordingly, otherwise download
checkpoints from here and extract them in the root directory of
this project.
python -m perceiver.scripts.seq_clf fit \
--model.mlm_ckpt='logs/mlm/version_0/checkpoints/epoch=254-val_loss=4.556.ckpt' \
--model.num_latent_channels=64 \
--model.encoder.num_layers=3 \
--model.encoder.freeze=true \
--model.encoder.dropout=0.0 \
--model.decoder.dropout=0.0 \
--data=IMDBDataModule \
--data.max_seq_len=512 \
--data.batch_size=128 \
--optimizer.lr=1e-3 \
--optimizer.weight_decay=0.01 \
--trainer.accelerator=gpu \
--trainer.devices=-1 \
--trainer.max_epochs=30
Unfreeze the encoder and jointly fine-tune it together with the decoder that has been trained in the previous step.
If you ran the previous step yourself you'll need to modify the --model.clf_ckpt
argument accordingly, otherwise
download checkpoints from here.
python -m perceiver.scripts.seq_clf fit \
--model.clf_ckpt='logs/seq_clf/version_0/checkpoints/epoch=024-val_loss=0.352.ckpt' \
--model.num_latent_channels=64 \
--model.encoder.num_layers=3 \
--model.encoder.dropout=0.1 \
--model.decoder.dropout=0.1 \
--data=IMDBDataModule \
--data.max_seq_len=512 \
--data.batch_size=128 \
--optimizer.lr=1e-4 \
--optimizer.weight_decay=0.01 \
--trainer.accelerator=gpu \
--trainer.devices=-1 \
--trainer.max_epochs=30
Image classification
Classify MNIST images. See also Model API for details about the underlying Perceiver IO model.
python -m perceiver.scripts.img_clf fit \
--model.num_latent_channels=128 \
--model.encoder.num_layers=3 \
--model.encoder.dropout=0.0 \
--model.decoder.dropout=0.0 \
--data=MNISTDataModule \
--data.batch_size=128 \
--optimizer.lr=1e-3 \
--optimizer.weight_decay=0.01 \
--trainer.accelerator=gpu \
--trainer.devices=-1 \
--trainer.max_epochs=20
Notebooks
Start the notebook server with:
PYTHONPATH=.. jupyter notebook
Model API
The model API is based on generic encoder and decoder classes (PerceiverEncoder
and
PerceiverDecoder
) and task-specific input and output adapters. The following snippet
shows how they can be used to create an MNIST image classifier, for example:
from perceiver.model import (
PerceiverIO,
PerceiverEncoder,
PerceiverDecoder,
ImageInputAdapter,
ClassificationOutputAdapter,
)
# Fourier-encode pixel positions and flatten along spatial dimensions
input_adapter = ImageInputAdapter(image_shape=(28, 28, 1), num_frequency_bands=32)
# Project generic Perceiver decoder output to specified number of classes
output_adapter = ClassificationOutputAdapter(num_classes=10, num_output_channels=128)
# Generic Perceiver encoder
encoder = PerceiverEncoder(
input_adapter=input_adapter,
num_latents=32,
num_latent_channels=128,
num_layers=3,
num_cross_attention_heads=4,
num_self_attention_heads=4,
num_self_attention_layers_per_block=3,
dropout=0.0,
)
# Generic Perceiver decoder
decoder = PerceiverDecoder(
output_adapter=output_adapter,
num_latent_channels=128,
num_cross_attention_heads=1,
dropout=0.0,
)
# MNIST classifier implemented as Perceiver IO model
mnist_classifier = PerceiverIO(encoder, decoder)
Development environment
Update the project dependencies in the conda environment:
invoke install
Install the pre-commit hooks:
invoke precommit-install
Run code quality checks:
invoke cc
Run tests:
invoke test
The project and task structure presented here is based on the Python Project Template.
Citations
@misc{jaegle2021perceiver,
title = {Perceiver: General Perception with Iterative Attention},
author = {Andrew Jaegle and Felix Gimeno and Andrew Brock and Andrew Zisserman and Oriol Vinyals and Joao Carreira},
year = {2021},
eprint = {2103.03206},
archivePrefix = {arXiv},
primaryClass = {cs.CV}
}
@misc{jaegle2021perceiver,
title = {Perceiver IO: A General Architecture for Structured Inputs & Outputs},
author = {Andrew Jaegle and Sebastian Borgeaud and Jean-Baptiste Alayrac and Carl Doersch and Catalin Ionescu and David Ding and Skanda Koppula and Andrew Brock and Evan Shelhamer and Olivier Hénaff and Matthew M. Botvinick and Andrew Zisserman and Oriol Vinyals and João Carreira},
year = {2021},
eprint = {2107.14795},
archivePrefix = {arXiv},
primaryClass = {cs.LG}
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for perceiver_io-0.1.2-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 9b21446c3d24d06d616b4d4fe40227b595ab1bd43356bb51eb5ac4c85fe2faff |
|
MD5 | 1eca93fe27f4b8e80118e2a99b5e0ccc |
|
BLAKE2b-256 | 87669ad193cb164239fcad541f1170b4489abc41b13816cb6e3464a9196b47cf |