Code for "Unsupervised State Representation Learning in Atari"

Last update: Jan 03, 2023

Overview

Unsupervised State Representation Learning in Atari

Ankesh Anand*, Evan Racah*, Sherjil Ozair*, Yoshua Bengio, Marc-Alexandre Côté, R Devon Hjelm

This repo provides code for the benchmark and techniques introduced in the paper Unsupervised State Representation Learning in Atari

📦 Install -- Install relevant dependencies and the project
🔧 Usage -- Learn how to use AtariARI and train representations with Spatio-Temporal DeepInfomax (ST-DIM)
🕹️ RAM Annotations -- Mapping of RAM indexes to semantic state variables
📜 Poster -- Summary of the paper in a poster format
🎞️ Slides -- Slides describing the paper

Install

AtariARI Wrapper

You can do a minimal install to get just the AtariARI (Atari Annotated RAM Interface) wrapper by doing:

pip install 'gym[atari]'
pip install git+git://github.com/mila-iqia/atari-representation-learning.git

This just requires gym[atari] and it gives you the ability to play around with the AtariARI wrapper. If you want to use the code for training representation learning methods and probing them, you will need a full installation:

Full installation (AtariARI Wrapper + Training & Probing Code)

# PyTorch and scikit learn
conda install pytorch torchvision -c pytorch
conda install scikit-learn

# Baselines for Atari preprocessing
# Tensorflow is a dependency, but you don't need to install the GPU version
conda install tensorflow
pip install git+git://github.com/openai/baselines

# pytorch-a2c-ppo-acktr for RL utils
pip install git+git://github.com/ankeshanand/pytorch-a2c-ppo-acktr-gail

# Clone and install our package
pip install -r requirements.txt
pip install git+git://github.com/mila-iqia/atari-representation-learning.git

Usage

Atari Annotated RAM Interface (AtariARI):

AtariARI exposes the ground truth labels for different state variables for each observation. We have made AtariARI available as a Gym wrapper, to use it simply wrap an Atari gym env with AtariARIWrapper.

import gym
from atariari.benchmark.wrapper import AtariARIWrapper
env = AtariARIWrapper(gym.make('MsPacmanNoFrameskip-v4'))
obs = env.reset()
obs, reward, done, info = env.step(1)

Now, info is a dictionary of the form:

{'ale.lives': 3,
 'labels': {'enemy_sue_x': 88,
  'enemy_inky_x': 88,
  'enemy_pinky_x': 88,
  'enemy_blinky_x': 88,
  'enemy_sue_y': 80,
  'enemy_inky_y': 80,
  'enemy_pinky_y': 80,
  'enemy_blinky_y': 50,
  'player_x': 88,
  'player_y': 98,
  'fruit_x': 0,
  'fruit_y': 0,
  'ghosts_count': 3,
  'player_direction': 3,
  'dots_eaten_count': 0,
  'player_score': 0,
  'num_lives': 2}}

Note: In our experiments, we use additional preprocessing for Atari environments mainly following Minh et. al, 2014. See atariari/benchmark/envs.py for more info!

If you want the raw RAM annotations (which parts of ram correspond to each state variable), check out atariari/benchmark/ram_annotations.py

Probing

⚠️ Important ⚠️ : The RAM labels are meant for full-sized Atari observations (210 * 160). Probing results won't be accurate if you downsample the observations.

We provide an interface for the included probing tasks.

First, get episodes for train, val and, test:

from atariari.benchmark.episodes import get_episodes

tr_episodes, val_episodes,\
tr_labels, val_labels,\
test_episodes, test_labels = get_episodes(env_name="PitfallNoFrameskip-v4", 
                                     steps=50000, 
                                     collect_mode="random_agent")

Then probe them using ProbeTrainer and your encoder (my_encoder):

from atariari.benchmark.probe import ProbeTrainer

probe_trainer = ProbeTrainer(my_encoder, representation_len=my_encoder.feature_size)
probe_trainer.train(tr_episodes, val_episodes,
                     tr_labels, val_labels,)
final_accuracies, final_f1_scores = probe_trainer.test(test_episodes, test_labels)

To see how we use ProbeTrainer, check out scripts/run_probe.py

Here is an example of my_encoder:

# get your encoder
import torch.nn as nn
import torch
class MyEncoder(nn.Module):
    def __init__(self, input_channels, feature_size):
        super().__init__()
        self.feature_size = feature_size
        self.input_channels = input_channels
        self.final_conv_size = 64 * 9 * 6
        self.cnn = nn.Sequential(
            nn.Conv2d(input_channels, 32, 8, stride=4),
            nn.ReLU(),
            nn.Conv2d(32, 64, 4, stride=2),
            nn.ReLU(),
            nn.Conv2d(64, 128, 4, stride=2),
            nn.ReLU(),
            nn.Conv2d(128, 64, 3, stride=1),
            nn.ReLU()
        )
        self.fc = nn.Linear(self.final_conv_size, self.feature_size)

    def forward(self, inputs):
        x = self.cnn(inputs)
        x = x.view(x.size(0), -1)
        return self.fc(x)
        

my_encoder = MyEncoder(input_channels=1,feature_size=256)
# load in weights
my_encoder.load_state_dict(torch.load(open("path/to/my/weights.pt", "rb")))

Spatio-Temporal DeepInfoMax:

src/ contains implementations of several representation learning methods, along with ST-DIM. Here's a sample usage:

python -m scripts.run_probe --method infonce-stdim --env-name {env_name}

where env_name is of the form {game}NoFrameskip-v4, such as PongNoFrameskip-v4

Citation

@article{anand2019unsupervised,
  title={Unsupervised State Representation Learning in Atari},
  author={Anand, Ankesh and Racah, Evan and Ozair, Sherjil and Bengio, Yoshua and C{\^o}t{\'e}, Marc-Alexandre and Hjelm, R Devon},
  journal={arXiv preprint arXiv:1906.08226},
  year={2019}
}

Code for "Unsupervised State Representation Learning in Atari"

Related tags

Overview

Unsupervised State Representation Learning in Atari

Install

AtariARI Wrapper

Full installation (AtariARI Wrapper + Training & Probing Code)

Usage

Atari Annotated RAM Interface (AtariARI):

Probing

Spatio-Temporal DeepInfoMax:

Citation

Owner

Mila

Unofficial PyTorch implementation of Fastformer based on paper "Fastformer: Additive Attention Can Be All You Need"."

This tool uses Deep Learning to help you draw and write with your hand and webcam.

The modify PyTorch version of Siam-trackers which are speed-up by TensorRT.

Safe Control for Black-box Dynamical Systems via Neural Barrier Certificates

Code repository for "Reducing Underflow in Mixed Precision Training by Gradient Scaling" presented at IJCAI '20

Official PyTorch implementation of paper: Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation (ICCV 2021 Oral Presentation)

A series of Python scripts to access measurements from Fluke 28X meters. Fluke IR Remote Interface required.

Hybrid CenterNet - Hybrid-supervised object detection / Weakly semi-supervised object detection

N-RPG - Novel role playing game da turfu

Towards Understanding Quality Challenges of the Federated Learning: A First Look from the Lens of Robustness

SHIFT15M: multiobjective large-scale fashion dataset with distributional shifts

Real-time ground filtering algorithm of cloud points acquired using Terrestrial Laser Scanner (TLS)

Harmonic Memory Networks for Graph Completion

A Momentumized, Adaptive, Dual Averaged Gradient Method for Stochastic Optimization

Code for "LoRA: Low-Rank Adaptation of Large Language Models"

Bunch of different tools which helps visualizing and annotating images for semantic/instance segmentation tasks

Meta-meta-learning with evolution and plasticity

Head2Toe: Utilizing Intermediate Representations for Better OOD Generalization

Awesome Transformers in Medical Imaging

Content shared at DS-OX Meetup