Initial commit

Co-authored-by: conglu1997 <[email protected]>
Co-authored-by: philipjball <[email protected]>

.gitignore

@@ -0,0 +1,148 @@
+bin/
+logs/
+wandb/
+online_logs/
+
+fuse.cfg
+
+*.ai
+
+# Generation results
+results/
+
+ray/auth.json
+
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# Custom ignores
+.idea/
+.vscode/
+*.DS_Store

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "synther/REDQ"]
+	path = synther/REDQ
+	url = https://github.com/watchernyu/REDQ

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2023 Cong Lu, Philip J. Ball.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,94 @@
+# Synthetic Experience Replay
+
+<p align="center">
+  <img src="figs/diffusion.png" />
+</p>
+
+Synthetic Experience Replay (SynthER) is a diffusion-based approach to arbitrarily upsample an RL agent's collected
+experience, leading to large gains in sample efficiency and scaling benefits. We integrate SynthER into a variety of
+offline and online algorithms in this codebase, including SAC, TD3+BC, IQL, EDAC, and CQL. For further details, please
+see the paper:
+
+**_Synthetic Experience Replay_**; Cong Lu*, Philip J. Ball*, Yee Whye Teh, Jack Parker-Holder.
+
+<p align="center">
+  <a href=https://arxiv.org/abs/2303.06614>View on arXiv</a>
+</p>
+
+## Setup
+
+To install, clone the repository and run the following:
+
+```bash 
+git submodule update --init --recursive
+pip install -r requirements.txt
+```
+
+## Running Instructions
+
+### Offline RL
+
+Diffusion model training (this automatically generates samples and saves them):
+
+```bash
+python3 synther/diffusion/train_diffuser.py --dataset halfcheetah-medium_replay-v2
+```
+
+Baseline without SynthER (e.g. on TD3+BC):
+
+```bash
+python3 synther/corl/algorithms/td3_bc.py --config synther/corl/yaml/td3_bc/halfcheetah/medium_replay_v2.yaml --checkpoints_path corl_logs/
+```
+
+Offline RL training with SynthER:
+
+```bash
+# Generating diffusion samples on the fly.
+python3 synther/corl/algorithms/td3_bc.py --config synther/corl/yaml/td3_bc/halfcheetah/medium_replay_v2.yaml --checkpoints_path corl_logs/ --name SynthER --diffusion_path path/to/model-100000.pt
+
+# Using saved diffusion samples.
+python3 synther/corl/algorithms/td3_bc.py --config synther/corl/yaml/td3_bc/halfcheetah/medium_replay_v2.yaml --checkpoints_path corl_logs/ --name SynthER --diffusion_path path/to/samples.npz
+```
+
+### Online RL
+
+Baselines (SAC, REDQ):
+
+```bash
+# SAC.
+python3 synther/online/online_exp.py --env quadruped-walk-v0 --results_folder online_logs/ --exp_name SAC --gin_config_files 'config/online/sac.gin'
+
+# REDQ.
+python3 synther/online/online_exp.py --env quadruped-walk-v0 --results_folder online_logs/ --exp_name REDQ --gin_config_files 'config/online/redq.gin'
+```
+
+SynthER (SAC):
+
+```bash
+# DMC environments.
+python3 synther/online/online_exp.py --env quadruped-walk-v0 --results_folder online_logs/ --exp_name SynthER --gin_config_files 'config/online/sac_synther_dmc.gin' --gin_params 'redq_sac.utd_ratio = 20' 'redq_sac.num_samples = 1000000'
+
+# OpenAI environments (different gin config).
+python3 synther/online/online_exp.py --env HalfCheetah-v2 --results_folder online_logs/ --exp_name SynthER --gin_config_files 'config/online/sac_synther_openai.gin' --gin_params 'redq_sac.utd_ratio = 20' 'redq_sac.num_samples = 1000000'
+```
+
+## Thinking of adding SynthER to your own algorithm?
+
+Our codebase has everything you need for diffusion with low-dimensional data along with example integrations with RL algorithms.
+For a custom use-case, we recommend starting from the training script and `SimpleDiffusionGenerator` class
+in `synther/diffusion/train_diffuser.py`. You can modify the hyperparameters specified in `config/resmlp_denoiser.gin`
+to suit your own needs.
+
+## Acknowledgements
+
+SynthER builds upon many works and open-source codebases in both diffusion modelling and reinforcement learning. We
+would like to particularly thank the authors of:
+
+- [denoising-diffusion-pytorch](https://github.com/lucidrains/denoising-diffusion-pytorch/tree/main/denoising_diffusion_pytorch)
+- [REDQ](https://github.com/watchernyu/REDQ)
+- [CORL](https://github.com/tinkoff-ai/CORL)
+
+## Contact
+
+Please contact [Cong Lu](mailto:[email protected]) or [Philip Ball](mailto:[email protected]) for any queries.
+We welcome any suggestions or contributions! 

config/online/redq.gin

@@ -0,0 +1,6 @@
+# REDQ baseline.
+
+include 'config/online/sac.gin'
+
+redq_sac.utd_ratio = 20
+redq_sac.num_Q = 10

config/online/sac.gin

@@ -0,0 +1,7 @@
+# SAC baseline.
+
+redq_sac.start_steps = 5000
+redq_sac.utd_ratio = 1
+redq_sac.num_Q = 2
+redq_sac.n_evals_per_epoch = 10
+redq_sac.evaluate_bias = False

config/online/sac_synther_dmc.gin

@@ -0,0 +1,45 @@
+# SAC (SynthER) for DMC environments.
+
+include 'config/online/sac.gin'
+
+# Higher UTD with synthetic data.
+redq_sac.utd_ratio = 20
+
+# SAC/REDQ diffusion args.
+redq_sac.disable_diffusion = False
+redq_sac.retrain_diffusion_every = 10000
+redq_sac.num_samples = 200000
+redq_sac.diffusion_sample_ratio = 0.5
+redq_sac.diffusion_start = 0
+redq_sac.skip_reward_norm = True
+
+# Online setup.
+construct_diffusion_model.denoising_network = @ResidualMLPDenoiser
+construct_diffusion_model.normalizer_type = 'standard'
+
+# No terminals in DMC.
+modelled_terminals = False
+redq_sac.model_terminals = %modelled_terminals
+split_diffusion_samples.modelled_terminals = %modelled_terminals
+
+# Diffusion denoising network.
+ResidualMLPDenoiser.dim_t = 256
+ResidualMLPDenoiser.mlp_width = 1024
+ResidualMLPDenoiser.num_layers = 6
+ResidualMLPDenoiser.learned_sinusoidal_cond = False
+ResidualMLPDenoiser.random_fourier_features = True
+ResidualMLPDenoiser.learned_sinusoidal_dim = 16
+ResidualMLPDenoiser.activation = 'relu'
+ResidualMLPDenoiser.layer_norm = False
+
+# Diffusion training.
+REDQTrainer.train_batch_size = 256
+REDQTrainer.train_lr = 3e-4
+REDQTrainer.lr_scheduler = "cosine"
+REDQTrainer.weight_decay = 0
+REDQTrainer.train_num_steps = 100000
+# Don't save.
+REDQTrainer.save_and_sample_every = 100000000
+
+# Diffusion sampling.
+SimpleDiffusionGenerator.num_sample_steps = 128

config/online/sac_synther_openai.gin

@@ -0,0 +1,11 @@
+# SAC (SynthER) for OpenAI environments.
+
+include 'config/online/sac_synther_dmc.gin'
+
+# Normalize rewards.
+redq_sac.skip_reward_norm = False
+
+# Model terminals.
+modelled_terminals = True
+redq_sac.model_terminals = %modelled_terminals
+split_diffusion_samples.modelled_terminals = %modelled_terminals

config/resmlp_denoiser.gin

@@ -0,0 +1,43 @@
+# Default Residual MLP configuration.
+
+# Model the terminals and discretize them with threshold 0.5.
+modelled_terminals = True
+make_inputs.modelled_terminals = %modelled_terminals
+split_diffusion_samples.modelled_terminals = %modelled_terminals
+split_diffusion_samples.terminal_threshold = 0.5
+
+construct_diffusion_model.normalizer_type = 'standard'
+construct_diffusion_model.denoising_network = @ResidualMLPDenoiser
+# No normalization on the terminals.
+construct_diffusion_model.disable_terminal_norm = True
+
+# Network.
+ResidualMLPDenoiser.dim_t = 128
+ResidualMLPDenoiser.mlp_width = 2048
+ResidualMLPDenoiser.num_layers = 6
+ResidualMLPDenoiser.learned_sinusoidal_cond = False
+ResidualMLPDenoiser.random_fourier_features = True
+ResidualMLPDenoiser.learned_sinusoidal_dim = 16
+ResidualMLPDenoiser.activation = 'relu'
+ResidualMLPDenoiser.layer_norm = False
+
+# Diffusion Model.
+ElucidatedDiffusion.num_sample_steps = 128
+ElucidatedDiffusion.sigma_data = 1.0
+ElucidatedDiffusion.S_churn = 80
+ElucidatedDiffusion.S_tmin = 0.05
+ElucidatedDiffusion.S_tmax = 50
+ElucidatedDiffusion.S_noise = 1.003
+
+# Training.
+Trainer.train_batch_size = 1024
+Trainer.small_batch_size = 256
+Trainer.train_lr = 3e-4
+Trainer.lr_scheduler = 'cosine'
+Trainer.weight_decay = 0
+Trainer.train_num_steps = 100000
+Trainer.save_and_sample_every = 10000
+
+# Sampling.
+SimpleDiffusionGenerator.num_sample_steps = 128
+SimpleDiffusionGenerator.sample_batch_size = 100000

config/resmlp_w1024_denoiser.gin

@@ -0,0 +1,6 @@
+# Default Residual MLP configuration (lower width for some datasets).
+
+include 'config/resmlp_denoiser.gin'
+
+# Network.
+ResidualMLPDenoiser.mlp_width = 1024