Upload with huggingface_hub

.summary/0/events.out.tfevents.1657179944.andrew-gpu

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b29c5422f43a8e85abff5051e5cb4c82cf05792df7cd619f5bdfdf0789d8950
+size 1808093

README.md

@@ -0,0 +1,24 @@
+---
+library_name: sample-factory
+tags:
+- deep-reinforcement-learning
+- reinforcement-learning
+- sample-factory
+model-index:
+- name: APPO
+  results:
+  - metrics:
+    - type: mean_reward
+      value: 6262.48 +/- 0.00
+      name: mean_reward
+    task:
+      type: reinforcement-learning
+      name: reinforcement-learning
+    dataset:
+      name: mujoco_halfcheetah
+      type: mujoco_halfcheetah
+---
+
+A(n) **APPO** model trained on the **mujoco_halfcheetah** environment.
+This model was trained using Sample Factory 2.0: https://github.com/alex-petrenko/sample-factory
+    

cfg.json

@@ -0,0 +1,113 @@
+{
+  "help": false,
+  "algo": "APPO",
+  "env": "mujoco_halfcheetah",
+  "experiment": "17_mujoco_all_envs_see_4444_env_mujoco_halfcheetah",
+  "train_dir": "./train_dir",
+  "experiments_root": "mujoco_all_envs/mujoco_all_envs_crl_4",
+  "device": "cpu",
+  "seed": 4444,
+  "num_policies": 1,
+  "async_rl": false,
+  "serial_mode": false,
+  "batched_sampling": false,
+  "num_batches_to_accumulate": 2,
+  "worker_num_splits": 2,
+  "policy_workers_per_policy": 1,
+  "max_policy_lag": 10000,
+  "num_workers": 8,
+  "num_envs_per_worker": 8,
+  "batch_size": 1024,
+  "num_batches_per_epoch": 4,
+  "num_epochs": 2,
+  "rollout": 64,
+  "recurrence": 1,
+  "shuffle_minibatches": false,
+  "gamma": 0.99,
+  "reward_scale": 1,
+  "reward_clip": 1000.0,
+  "value_bootstrap": false,
+  "normalize_returns": true,
+  "exploration_loss_coeff": 0.0,
+  "value_loss_coeff": 1.3,
+  "kl_loss_coeff": 0.1,
+  "exploration_loss": "entropy",
+  "gae_lambda": 0.95,
+  "ppo_clip_ratio": 0.2,
+  "ppo_clip_value": 1.0,
+  "with_vtrace": false,
+  "vtrace_rho": 1.0,
+  "vtrace_c": 1.0,
+  "optimizer": "adam",
+  "adam_eps": 1e-06,
+  "adam_beta1": 0.9,
+  "adam_beta2": 0.999,
+  "max_grad_norm": 3.5,
+  "learning_rate": 0.00295,
+  "lr_schedule": "linear_decay",
+  "lr_schedule_kl_threshold": 0.008,
+  "obs_subtract_mean": 0.0,
+  "obs_scale": 1.0,
+  "normalize_input": true,
+  "decorrelate_experience_max_seconds": 10,
+  "decorrelate_envs_on_one_worker": true,
+  "actor_worker_gpus": [],
+  "set_workers_cpu_affinity": true,
+  "force_envs_single_thread": true,
+  "default_niceness": 0,
+  "experiment_summaries_interval": 3,
+  "stats_avg": 100,
+  "train_for_env_steps": 10000000,
+  "train_for_seconds": 10000000000,
+  "save_every_sec": 15,
+  "keep_checkpoints": 3,
+  "load_checkpoint_kind": "latest",
+  "save_milestones_sec": -1,
+  "save_best_every_sec": 5,
+  "save_best_metric": "reward",
+  "save_best_after": 100000,
+  "benchmark": false,
+  "encoder_type": "mlp",
+  "encoder_subtype": "mlp_mujoco",
+  "encoder_custom": null,
+  "encoder_extra_fc_layers": 0,
+  "hidden_size": 64,
+  "nonlinearity": "tanh",
+  "policy_initialization": "torch_default",
+  "policy_init_gain": 1.0,
+  "actor_critic_share_weights": true,
+  "adaptive_stddev": false,
+  "initial_stddev": 1.0,
+  "use_rnn": false,
+  "rnn_type": "gru",
+  "rnn_num_layers": 1,
+  "env_gpu_actions": false,
+  "env_frameskip": 1,
+  "env_framestack": 4,
+  "pixel_format": "CHW",
+  "with_wandb": true,
+  "wandb_user": null,
+  "wandb_project": "sample_factory",
+  "wandb_group": null,
+  "wandb_job_type": "SF",
+  "wandb_tags": [
+    "mujoco",
+    "runner_crl_4"
+  ],
+  "command_line": "--algo=APPO --with_wandb=True --wandb_tags mujoco runner_crl_4 --seed=4444 --env=mujoco_halfcheetah --experiment=17_mujoco_all_envs_see_4444_env_mujoco_halfcheetah --train_dir=./train_dir --experiments_root=mujoco_all_envs/mujoco_all_envs_crl_4",
+  "cli_args": {
+    "algo": "APPO",
+    "env": "mujoco_halfcheetah",
+    "experiment": "17_mujoco_all_envs_see_4444_env_mujoco_halfcheetah",
+    "train_dir": "./train_dir",
+    "experiments_root": "mujoco_all_envs/mujoco_all_envs_crl_4",
+    "seed": 4444,
+    "with_wandb": true,
+    "wandb_tags": [
+      "mujoco",
+      "runner_crl_4"
+    ]
+  },
+  "git_hash": "4e69a02b57fff18dbbf96054866d0d759f70c5fa",
+  "git_repo_name": "https://github.com/andrewzhang505/sample-factory.git"
+}

checkpoint_p0/best_000018192_9314304_reward_6076.824.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6df998865903a0f2e767d8695070b501bf29faa722c5025357b1baf0fd0542d
+size 80130

checkpoint_p0/checkpoint_000019368_9916416.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:862f4bb059a4a40e1580072dce7d55e27bfd328cdc7806a47f3aaa327ed59113
+size 80130

checkpoint_p0/checkpoint_000019464_9965568.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b2efd37896ce4b9d85774ab8720cbcd0b77c9145acd97019365cf0c4227676e
+size 80130

checkpoint_p0/checkpoint_000019536_10002432.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:857e35efd6d1d1be24b30b09afe9fc848124ee8b224f8206f769eff87aa0389e
+size 80130

git.diff

@@ -0,0 +1,328 @@
+diff --git a/sample_factory/algo/learning/learner.py b/sample_factory/algo/learning/learner.py
+index 178d2ab..20bb937 100644
+--- a/sample_factory/algo/learning/learner.py
++++ b/sample_factory/algo/learning/learner.py
+@@ -110,6 +110,20 @@ class KlAdaptiveSchedulerPerEpoch(KlAdaptiveScheduler):
+     def invoke_after_each_epoch(self):
+         return True
+ 
++class LinearDecayScheduler(LearningRateScheduler):
++    def __init__(self, cfg):
++        num_updates = cfg.train_for_env_steps // cfg.batch_size * cfg.num_epochs
++        self.linear_decay = LinearDecay([(0, cfg.learning_rate), (num_updates, 0)])
++        self.step = 0
++    
++    def invoke_after_each_minibatch(self):
++        return True
++
++    def update(self, current_lr, recent_kls):
++        self.step += 1
++        lr = self.linear_decay.at(self.step)
++        return lr
++
+ 
+ def get_lr_scheduler(cfg) -> LearningRateScheduler:
+     if cfg.lr_schedule == "constant":
+@@ -118,6 +132,8 @@ def get_lr_scheduler(cfg) -> LearningRateScheduler:
+         return KlAdaptiveSchedulerPerMinibatch(cfg)
+     elif cfg.lr_schedule == "kl_adaptive_epoch":
+         return KlAdaptiveSchedulerPerEpoch(cfg)
++    elif cfg.lr_schedule == "linear_decay":
++        return LinearDecayScheduler(cfg)
+     else:
+         raise RuntimeError(f"Unknown scheduler {cfg.lr_schedule}")
+ 
+diff --git a/sample_factory/envs/mujoco/mujoco_params.py b/sample_factory/envs/mujoco/mujoco_params.py
+index ef0b486..cb4b977 100644
+--- a/sample_factory/envs/mujoco/mujoco_params.py
++++ b/sample_factory/envs/mujoco/mujoco_params.py
+@@ -1,117 +1,155 @@
++# def mujoco_override_defaults(env, parser):
++#     parser.set_defaults(
++#         batched_sampling=False,
++#         num_workers=8,
++#         num_envs_per_worker=16,
++#         worker_num_splits=2,
++#         train_for_env_steps=1000000,
++#         encoder_type="mlp",
++#         encoder_subtype="mlp_mujoco",
++#         hidden_size=64,
++#         encoder_extra_fc_layers=0,
++#         env_frameskip=1,
++#         nonlinearity="tanh",
++#         batch_size=64,
++#         kl_loss_coeff=0.1,
++#         use_rnn=False,
++#         adaptive_stddev=False,
++#         policy_initialization="torch_default",
++#         reward_scale=0.01,
++#         rollout=8,
++#         max_grad_norm=0.0,
++#         ppo_epochs=10,
++#         num_batches_per_epoch=32,
++#         ppo_clip_ratio=0.2,
++#         value_loss_coeff=2.0,
++#         exploration_loss_coeff=0.0,
++#         learning_rate=3e-3,
++#         lr_schedule="constant",
++#         shuffle_minibatches=True,
++#         gamma=0.99,
++#         gae_lambda=0.95,
++#         with_vtrace=False,
++#         recurrence=1,
++#         value_bootstrap=False,
++#         normalize_input=True,
++#         experiment_summaries_interval=3,
++#         save_every_sec=15,
++#         serial_mode=False,
++#         async_rl=False,
++#     )
++
++#     # environment specific overrides
++#     env_name = "_".join(env.split("_")[1:]).lower()
++
++#     if env_name == "halfcheetah":
++#         parser.set_defaults(
++#             reward_scale=0.1,
++#             learning_rate=3e-3,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-2,
++#             normalize_input=False,
++#             num_batches_per_epoch=1,
++#         )
++#     if env_name == "humanoid":
++#         parser.set_defaults(
++#             learning_rate=3e-4,
++#         )
++#     if env_name == "hopper":
++#         parser.set_defaults(
++#             reward_scale=0.1,
++#             learning_rate=3e-3,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-2,
++#             # normalize_input=False,
++#             # num_batches_per_epoch=1,
++#             # normalize_returns=True,
++#             # hidden_size=128,
++#         )
++#     if env_name == "doublependulum":
++#         parser.set_defaults(
++#             reward_scale=0.01,
++#             learning_rate=3e-3,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-2,
++#         )
++#     if env_name == "pendulum":
++#         parser.set_defaults(
++#             # reward_scale=0.01,
++#             learning_rate=3e-4,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-3,
++#         )
++#     if env_name == "reacher":
++#         parser.set_defaults(
++#             reward_scale=0.1,
++#             learning_rate=3e-3,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-2,
++#             normalize_input=False,
++#             num_batches_per_epoch=1,
++#         )
++#     if env_name == "swimmer":
++#         parser.set_defaults(
++#             reward_scale=1,
++#             # learning_rate=3e-3,
++#             # lr_schedule="kl_adaptive_epoch",
++#             # lr_schedule_kl_threshold=3e-2,
++#             # gamma=0.9995,
++#             rollout=128,
++#             batch_size=128,
++#         )
++#     if env_name == "walker":
++#         parser.set_defaults(
++#             reward_scale=0.1,
++#             learning_rate=3e-3,
++#             lr_schedule="kl_adaptive_epoch",
++#             lr_schedule_kl_threshold=3e-2,
++#         )
++
+ def mujoco_override_defaults(env, parser):
+     parser.set_defaults(
+         batched_sampling=False,
+         num_workers=8,
+-        num_envs_per_worker=16,
++        num_envs_per_worker=8,
+         worker_num_splits=2,
+-        train_for_env_steps=1000000,
++        train_for_env_steps=10000000,
+         encoder_type="mlp",
+         encoder_subtype="mlp_mujoco",
+         hidden_size=64,
+         encoder_extra_fc_layers=0,
+         env_frameskip=1,
+         nonlinearity="tanh",
+-        batch_size=64,
++        batch_size=1024,
+         kl_loss_coeff=0.1,
+-
+         use_rnn=False,
+         adaptive_stddev=False,
+         policy_initialization="torch_default",
+-        reward_scale=0.01,
+-        rollout=8,
+-        max_grad_norm=0.0,
+-        ppo_epochs=10,
+-        num_batches_per_epoch=32,
++        reward_scale=1,
++        rollout=64,
++        max_grad_norm=3.5,
++        num_epochs=2,
++        num_batches_per_epoch=4,
+         ppo_clip_ratio=0.2,
+-        value_loss_coeff=2.0,
++        value_loss_coeff=1.3,
+         exploration_loss_coeff=0.0,
+-        learning_rate=3e-3,
+-        lr_schedule="constant",
+-        shuffle_minibatches=True,
++        learning_rate=0.00295,
++        lr_schedule="linear_decay",
++        shuffle_minibatches=False,
+         gamma=0.99,
+         gae_lambda=0.95,
+         with_vtrace=False,
+         recurrence=1,
+         value_bootstrap=False,
+         normalize_input=True,
++        normalize_returns=True,
+         experiment_summaries_interval=3,
+         save_every_sec=15,
+-        
+         serial_mode=False,
+         async_rl=False,
+     )
+ 
+-    # environment specific overrides
+-    env_name = "_".join(env.split("_")[1:]).lower()
+-
+-    if env_name == "halfcheetah":
+-        parser.set_defaults(
+-            reward_scale=0.1,
+-            learning_rate=3e-3,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-2,
+-            normalize_input=False,
+-            num_batches_per_epoch=1,
+-        )
+-    if env_name == "humanoid":
+-        parser.set_defaults(
+-            learning_rate=3e-4,
+-        )    
+-    if env_name == "hopper":
+-        parser.set_defaults(
+-            reward_scale=0.1,
+-            learning_rate=3e-3,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-2,
+-            # normalize_input=False,
+-            # num_batches_per_epoch=1,
+-            # normalize_returns=True,
+-            # hidden_size=128,
+-        )
+-    if env_name == "doublependulum":
+-        parser.set_defaults(
+-            reward_scale=0.01,
+-            learning_rate=3e-3,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-2,
+-        )
+-    if env_name == "pendulum":
+-        parser.set_defaults(
+-            # reward_scale=0.01,
+-            learning_rate=3e-4,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-3,
+-        )
+-    if env_name == "reacher":
+-        parser.set_defaults(
+-            reward_scale=0.1,
+-            learning_rate=3e-3,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-2,
+-            normalize_input=False,
+-            num_batches_per_epoch=1,
+-        )
+-    if env_name == "swimmer":
+-        parser.set_defaults(
+-            reward_scale=1,
+-            learning_rate=3e-4,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-3,
+-            # normalize_input=False,
+-            # num_batches_per_epoch=1,
+-            normalize_returns=True,
+-            hidden_size=128,
+-        )
+-    if env_name == "walker":
+-        parser.set_defaults(
+-            reward_scale=0.1,
+-            learning_rate=3e-3,
+-            lr_schedule="kl_adaptive_epoch",
+-            lr_schedule_kl_threshold=3e-2,
+-            # normalize_returns=True,
+-            # normalize_input=False,
+-            # num_batches_per_epoch=1,
+-        )
++    
+ 
+ 
+ # noinspection PyUnusedLocal
+diff --git a/sample_factory/model/model_utils.py b/sample_factory/model/model_utils.py
+index df6c82c..d8226d8 100644
+--- a/sample_factory/model/model_utils.py
++++ b/sample_factory/model/model_utils.py
+@@ -276,7 +276,7 @@ class MlpEncoder(EncoderBase):
+         self.init_fc_blocks(fc_encoder_layer)
+ 
+     def forward(self, obs_dict):
+-        x = self.mlp_head(obs_dict['obs'].float())
++        x = self.mlp_head(obs_dict["obs"].float())
+         x = self.forward_fc_blocks(x)
+         return x
+ 
+diff --git a/sample_factory/runner/runs/mujoco_all_envs.py b/sample_factory/runner/runs/mujoco_all_envs.py
+index 3ac67ce..5cbaa1a 100644
+--- a/sample_factory/runner/runs/mujoco_all_envs.py
++++ b/sample_factory/runner/runs/mujoco_all_envs.py
+@@ -8,12 +8,12 @@ _params = ParamGrid(
+             [
+                 "mujoco_ant",
+                 "mujoco_halfcheetah",
+-                "mujoco_hopper",
++                # "mujoco_hopper",
+                 "mujoco_humanoid",
+-                "mujoco_doublependulum",
+-                "mujoco_pendulum",
+-                "mujoco_reacher",
+-                "mujoco_swimmer",
++                # "mujoco_doublependulum",
++                # "mujoco_pendulum",
++                # "mujoco_reacher",
++                # "mujoco_swimmer",
+                 "mujoco_walker",
+             ],
+         ),
+@@ -23,11 +23,11 @@ _params = ParamGrid(
+ _experiments = [
+     Experiment(
+         "mujoco_all_envs",
+-        "python -m sample_factory_examples.mujoco_examples.train_mujoco --algo=APPO --with_wandb=True --wandb_tags mujoco runner_4",
++        "python -m sample_factory_examples.mujoco_examples.train_mujoco --algo=APPO --with_wandb=True --wandb_tags mujoco runner_crl_4",
+         _params.generate_params(randomize=False),
+     ),
+ ]
+ 
+ 
+ RUN_DESCRIPTION = RunDescription("mujoco_all_envs", experiments=_experiments)
+-# python -m sample_factory.runner.run --run=mujoco_all_envs --runner=processes --max_parallel=8  --pause_between=1 --experiments_per_gpu=10000 --num_gpus=1 --experiment_suffix=4
++# python -m sample_factory.runner.run --run=mujoco_all_envs --runner=processes --max_parallel=4  --pause_between=1 --experiments_per_gpu=32 --num_gpus=1 --experiment_suffix=crl_3