hongshaorou/Tianshou
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
Tianshou/examples/atari/README.md

98 lines
10 KiB
Markdown
Raw Normal View History

Fix critic network for Discrete CRR (#485) - Fixes an inconsistency in the implementation of Discrete CRR. Now it uses `Critic` class for its critic, following conventions in other actor-critic policies; - Updates several offline policies to use `ActorCritic` class for its optimizer to eliminate randomness caused by parameter sharing between actor and critic; - Add `writer.flush()` in TensorboardLogger to ensure real-time result; - Enable `test_collector=None` in 3 trainers to turn off testing during training; - Updates the Atari offline results in README.md; - Moves Atari offline RL examples to `examples/offline`; tests to `test/offline` per review comments.
2021-11-28 07:10:28 -08:00
# Atari
DQN Atari examples (#187) This PR aims to provide the script of Atari DQN setting: - A speedrun of PongNoFrameskip-v4 (finished, about half an hour in i7-8750 + GTX1060 with 1M environment steps) - A general script for all atari game Since we use multiple env for simulation, the result is slightly different from the original paper, but consider to be acceptable. It also adds another parameter save_only_last_obs for replay buffer in order to save the memory. Co-authored-by: Trinkle23897 <463003665@qq.com>
2020-08-30 05:48:09 +08:00
The sample speed is \~3000 env step per second (\~12000 Atari frame per second in fact since we use frame_stack=4) under the normal mode (use a CNN policy and a collector, also storing data into the buffer). The main bottleneck is training the convolutional neural network.
The Atari env seed cannot be fixed due to the discussion [here](https://github.com/openai/gym/issues/1478), but it is not a big issue since on Atari it will always have the similar results.
The env wrapper is a crucial thing. Without wrappers, the agent cannot perform well enough on Atari games. Many existing RL codebases use [OpenAI wrapper](https://github.com/openai/baselines/blob/master/baselines/common/atari_wrappers.py), but it is not the original DeepMind version ([related issue](https://github.com/openai/baselines/issues/240)). Dopamine has a different [wrapper](https://github.com/google/dopamine/blob/master/dopamine/discrete_domains/atari_lib.py) but unfortunately it cannot work very well in our codebase.
# DQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters | time cost |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ | ------------------- |
change API of train_fn and test_fn (#229) train_fn(epoch) -> train_fn(epoch, num_env_step) test_fn(epoch) -> test_fn(epoch, num_env_step)
2020-09-26 16:35:37 +08:00
| PongNoFrameskip-v4 | 20 | ![](results/dqn/Pong_rew.png) | `python3 atari_dqn.py --task "PongNoFrameskip-v4" --batch-size 64` | ~30 min (~15 epoch) |
| BreakoutNoFrameskip-v4 | 316 | ![](results/dqn/Breakout_rew.png) | `python3 atari_dqn.py --task "BreakoutNoFrameskip-v4" --test-num 100` | 3~4h (100 epoch) |
| EnduroNoFrameskip-v4 | 670 | ![](results/dqn/Enduro_rew.png) | `python3 atari_dqn.py --task "EnduroNoFrameskip-v4 " --test-num 100` | 3~4h (100 epoch) |
| QbertNoFrameskip-v4 | 7307 | ![](results/dqn/Qbert_rew.png) | `python3 atari_dqn.py --task "QbertNoFrameskip-v4" --test-num 100` | 3~4h (100 epoch) |
| MsPacmanNoFrameskip-v4 | 2107 | ![](results/dqn/MsPacman_rew.png) | `python3 atari_dqn.py --task "MsPacmanNoFrameskip-v4" --test-num 100` | 3~4h (100 epoch) |
| SeaquestNoFrameskip-v4 | 2088 | ![](results/dqn/Seaquest_rew.png) | `python3 atari_dqn.py --task "SeaquestNoFrameskip-v4" --test-num 100` | 3~4h (100 epoch) |
| SpaceInvadersNoFrameskip-v4 | 812.2 | ![](results/dqn/SpaceInvader_rew.png) | `python3 atari_dqn.py --task "SpaceInvadersNoFrameskip-v4" --test-num 100` | 3~4h (100 epoch) |
Note: The `eps_train_final` and `eps_test` in the original DQN paper is 0.1 and 0.01, but [some works](https://github.com/google/dopamine/tree/master/baselines) found that smaller eps helps improve the performance. Also, a large batchsize (say 64 instead of 32) will help faster convergence but will slow down the training speed.
DQN Atari examples (#187) This PR aims to provide the script of Atari DQN setting: - A speedrun of PongNoFrameskip-v4 (finished, about half an hour in i7-8750 + GTX1060 with 1M environment steps) - A general script for all atari game Since we use multiple env for simulation, the result is slightly different from the original paper, but consider to be acceptable. It also adds another parameter save_only_last_obs for replay buffer in order to save the memory. Co-authored-by: Trinkle23897 <463003665@qq.com>
2020-08-30 05:48:09 +08:00
We haven't tuned this result to the best, so have fun with playing these hyperparameters!
Add C51 algorithm (#266) This is the PR for C51algorithm: https://arxiv.org/abs/1707.06887 1. add C51 policy in tianshou/policy/modelfree/c51.py. 2. add C51 net in tianshou/utils/net/discrete.py. 3. add C51 atari example in examples/atari/atari_c51.py. 4. add C51 statement in tianshou/policy/__init__.py. 5. add C51 test in test/discrete/test_c51.py. 6. add C51 atari results in examples/atari/results/c51/. By running "python3 atari_c51.py --task "PongNoFrameskip-v4" --batch-size 64", get best_result': '20.50 ± 0.50', in epoch 9. By running "python3 atari_c51.py --task "BreakoutNoFrameskip-v4" --n-step 1 --epoch 40", get best_reward: 407.400000 ± 31.155096 in epoch 39.
2021-01-06 10:17:45 +08:00
# C51 (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ |
| PongNoFrameskip-v4 | 20 | ![](results/c51/Pong_rew.png) | `python3 atari_c51.py --task "PongNoFrameskip-v4" --batch-size 64` |
| BreakoutNoFrameskip-v4 | 536.6 | ![](results/c51/Breakout_rew.png) | `python3 atari_c51.py --task "BreakoutNoFrameskip-v4" --n-step 1` |
| EnduroNoFrameskip-v4 | 1032 | ![](results/c51/Enduro_rew.png) | `python3 atari_c51.py --task "EnduroNoFrameskip-v4 " ` |
| QbertNoFrameskip-v4 | 16245 | ![](results/c51/Qbert_rew.png) | `python3 atari_c51.py --task "QbertNoFrameskip-v4"` |
| MsPacmanNoFrameskip-v4 | 3133 | ![](results/c51/MsPacman_rew.png) | `python3 atari_c51.py --task "MsPacmanNoFrameskip-v4"` |
| SeaquestNoFrameskip-v4 | 6226 | ![](results/c51/Seaquest_rew.png) | `python3 atari_c51.py --task "SeaquestNoFrameskip-v4"` |
| SpaceInvadersNoFrameskip-v4 | 988.5 | ![](results/c51/SpaceInvader_rew.png) | `python3 atari_c51.py --task "SpaceInvadersNoFrameskip-v4"` |
Add offline trainer and discrete BCQ algorithm (#263) The result needs to be tuned after `done` issue fixed. Co-authored-by: n+e <trinkle23897@gmail.com>
2021-01-20 02:13:04 -08:00
Note: The selection of `n_step` is based on Figure 6 in the [Rainbow](https://arxiv.org/abs/1710.02298) paper.
Add QR-DQN algorithm (#276) This is the PR for QR-DQN algorithm: https://arxiv.org/abs/1710.10044 1. add QR-DQN policy in tianshou/policy/modelfree/qrdqn.py. 2. add QR-DQN net in examples/atari/atari_network.py. 3. add QR-DQN atari example in examples/atari/atari_qrdqn.py. 4. add QR-DQN statement in tianshou/policy/init.py. 5. add QR-DQN unit test in test/discrete/test_qrdqn.py. 6. add QR-DQN atari results in examples/atari/results/qrdqn/. 7. add compute_q_value in DQNPolicy and C51Policy for simplify forward function. 8. move `with torch.no_grad():` from `_target_q` to BasePolicy By running "python3 atari_qrdqn.py --task "PongNoFrameskip-v4" --batch-size 64", get best_result': '19.8 ± 0.40', in epoch 8.
2021-01-28 09:27:05 +08:00
# QRDQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ |
| PongNoFrameskip-v4 | 20 | ![](results/qrdqn/Pong_rew.png) | `python3 atari_qrdqn.py --task "PongNoFrameskip-v4" --batch-size 64` |
| BreakoutNoFrameskip-v4 | 409.2 | ![](results/qrdqn/Breakout_rew.png) | `python3 atari_qrdqn.py --task "BreakoutNoFrameskip-v4" --n-step 1` |
| EnduroNoFrameskip-v4 | 1055.9 | ![](results/qrdqn/Enduro_rew.png) | `python3 atari_qrdqn.py --task "EnduroNoFrameskip-v4"` |
| QbertNoFrameskip-v4 | 14990 | ![](results/qrdqn/Qbert_rew.png) | `python3 atari_qrdqn.py --task "QbertNoFrameskip-v4"` |
| MsPacmanNoFrameskip-v4 | 2886 | ![](results/qrdqn/MsPacman_rew.png) | `python3 atari_qrdqn.py --task "MsPacmanNoFrameskip-v4"` |
| SeaquestNoFrameskip-v4 | 5676 | ![](results/qrdqn/Seaquest_rew.png) | `python3 atari_qrdqn.py --task "SeaquestNoFrameskip-v4"` |
| SpaceInvadersNoFrameskip-v4 | 938 | ![](results/qrdqn/SpaceInvader_rew.png) | `python3 atari_qrdqn.py --task "SpaceInvadersNoFrameskip-v4"` |
Add Implicit Quantile Network (#371)
2021-05-28 18:44:23 -07:00
# IQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ |
update iqn results and reward plots (#377)
2021-06-09 18:05:25 -07:00
| PongNoFrameskip-v4 | 20.3 | ![](results/iqn/Pong_rew.png) | `python3 atari_iqn.py --task "PongNoFrameskip-v4" --batch-size 64` |
| BreakoutNoFrameskip-v4 | 496.7 | ![](results/iqn/Breakout_rew.png) | `python3 atari_iqn.py --task "BreakoutNoFrameskip-v4" --n-step 1` |
| EnduroNoFrameskip-v4 | 1545 | ![](results/iqn/Enduro_rew.png) | `python3 atari_iqn.py --task "EnduroNoFrameskip-v4"` |
| QbertNoFrameskip-v4 | 15342.5 | ![](results/iqn/Qbert_rew.png) | `python3 atari_iqn.py --task "QbertNoFrameskip-v4"` |
| MsPacmanNoFrameskip-v4 | 2915 | ![](results/iqn/MsPacman_rew.png) | `python3 atari_iqn.py --task "MsPacmanNoFrameskip-v4"` |
| SeaquestNoFrameskip-v4 | 4874 | ![](results/iqn/Seaquest_rew.png) | `python3 atari_iqn.py --task "SeaquestNoFrameskip-v4"` |
| SpaceInvadersNoFrameskip-v4 | 1498.5 | ![](results/iqn/SpaceInvaders_rew.png) | `python3 atari_iqn.py --task "SpaceInvadersNoFrameskip-v4"` |
Add Implicit Quantile Network (#371)
2021-05-28 18:44:23 -07:00
Add Fully-parameterized Quantile Function (#376)
2021-06-14 20:59:02 -07:00
# FQF (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ |
| PongNoFrameskip-v4 | 20.7 | ![](results/fqf/Pong_rew.png) | `python3 atari_fqf.py --task "PongNoFrameskip-v4" --batch-size 64` |
| BreakoutNoFrameskip-v4 | 517.3 | ![](results/fqf/Breakout_rew.png) | `python3 atari_fqf.py --task "BreakoutNoFrameskip-v4" --n-step 1` |
| EnduroNoFrameskip-v4 | 2240.5 | ![](results/fqf/Enduro_rew.png) | `python3 atari_fqf.py --task "EnduroNoFrameskip-v4"` |
| QbertNoFrameskip-v4 | 16172.5 | ![](results/fqf/Qbert_rew.png) | `python3 atari_fqf.py --task "QbertNoFrameskip-v4"` |
| MsPacmanNoFrameskip-v4 | 2429 | ![](results/fqf/MsPacman_rew.png) | `python3 atari_fqf.py --task "MsPacmanNoFrameskip-v4"` |
| SeaquestNoFrameskip-v4 | 10775 | ![](results/fqf/Seaquest_rew.png) | `python3 atari_fqf.py --task "SeaquestNoFrameskip-v4"` |
| SpaceInvadersNoFrameskip-v4 | 2482 | ![](results/fqf/SpaceInvaders_rew.png) | `python3 atari_fqf.py --task "SpaceInvadersNoFrameskip-v4"` |
Add Rainbow DQN (#386) - add RainbowPolicy - add `set_beta` method in prio_buffer - add NoisyLinear in utils/network
2021-08-29 08:34:59 -07:00
# Rainbow (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
| --------------------------- | ----------- | ------------------------------------- | ------------------------------------------------------------ |
| PongNoFrameskip-v4 | 21 | ![](results/rainbow/Pong_rew.png) | `python3 atari_rainbow.py --task "PongNoFrameskip-v4" --batch-size 64` |
| BreakoutNoFrameskip-v4 | 684.6 | ![](results/rainbow/Breakout_rew.png) | `python3 atari_rainbow.py --task "BreakoutNoFrameskip-v4" --n-step 1` |
| EnduroNoFrameskip-v4 | 1625.9 | ![](results/rainbow/Enduro_rew.png) | `python3 atari_rainbow.py --task "EnduroNoFrameskip-v4"` |
| QbertNoFrameskip-v4 | 16192.5 | ![](results/rainbow/Qbert_rew.png) | `python3 atari_rainbow.py --task "QbertNoFrameskip-v4"` |
| MsPacmanNoFrameskip-v4 | 3101 | ![](results/rainbow/MsPacman_rew.png) | `python3 atari_rainbow.py --task "MsPacmanNoFrameskip-v4"` |
| SeaquestNoFrameskip-v4 | 2126 | ![](results/rainbow/Seaquest_rew.png) | `python3 atari_rainbow.py --task "SeaquestNoFrameskip-v4"` |
| SpaceInvadersNoFrameskip-v4 | 1794.5 | ![](results/rainbow/SpaceInvaders_rew.png) | `python3 atari_rainbow.py --task "SpaceInvadersNoFrameskip-v4"` |
Reference in New Issue Copy Permalink