Fully Integrate Minecraft Dict Environments

.circleci/config.yml

@@ -6,7 +6,7 @@ jobs:
     docker:
       # this is also used in gcp_cluster_template.yaml (if you change this file,
       # remember to update all your cluster configs too)
-      - image: humancompatibleai/il-representations:2021.02.22
+      - image: humancompatibleai/il-representations:2021.03.16
     steps:
       - checkout
       - run:

Dockerfile

@@ -21,6 +21,7 @@ RUN apt-get update -q \
     wget \
     xpra \
     xserver-xorg-dev \
+    libopenmpi-dev \
     libxrandr2 \
     libxss1 \
     libxcursor1 \

cloud/gcp_cluster_template.yaml

@@ -32,7 +32,7 @@ docker:
   # the image below is taken from .circleci/config.yml
   # (if you change this file or that file, then remember to change the other
   # too)
-  image: "humancompatibleai/il-representations:2021.02.22"
+  image: "humancompatibleai/il-representations:2021.03.16"
   container_name: "il-rep-ray-tune"
   # Set to true if to always force-pull the latest image version (no cache).
   pull_before_run: False

requirements.txt

@@ -48,3 +48,8 @@ git+git://github.com/HumanCompatibleAI/stable-baselines3@ad902dd5f9d4afeef347897
 magical-il~=0.0.1a4
 dm_control~=0.0.319497192
 git+git://github.com/denisyarats/dmc2gym@6e34d8acf18e92f0ea0a38ecee9564bdf2549076#egg=dmc2gym
+
+# # NOTE TO SELF: Developing locally with updated RB, need to change that here once that PR is merged
+# # Target RB branch: ilr_wrappers
+# # [Also need to make branch un-private somehow, possibly giving CircleCI a checkout key?...]
+# git+ssh://git@github.com/HumanCompatibleAI/realistic-benchmarks@master#egg=realistic-benchmarks

src/il_representations/algos/encoders.py

@@ -14,6 +14,7 @@
 
 from gym import spaces
 from il_representations.algos.utils import independent_multivariate_normal
+from il_representations.utils import ForkedPdb
 import functools
 
 """
@@ -407,7 +408,6 @@ def infer_action_shape_info(action_space, action_embedding_dim):
     and the action_embedding_dim
     """
     # Machinery for turning raw actions into vectors.
-
     if isinstance(action_space, spaces.Discrete):
         # If actions are discrete, this is done via an Embedding.
         action_processor = nn.Embedding(num_embeddings=action_space.n, embedding_dim=action_embedding_dim)
@@ -439,6 +439,7 @@ def __init__(self, obs_space, representation_dim, action_space, learn_scale=Fals
 
         super().__init__(obs_space, representation_dim, obs_encoder_cls, learn_scale=learn_scale)
 
+        # TODO make this work better for more complex action types like those in Minecraft
         self.processed_action_dim, self.action_shape, self.action_processor = infer_action_shape_info(action_space,
                                                                                        action_embedding_dim)
         # Machinery for aggregating information from an arbitrary number of actions into a single vector,
@@ -476,7 +477,7 @@ def forward(self, x, traj_info, action=False):
         if self.action_encoder is not None:
             output, (hidden, cell) = self.action_encoder(processed_actions)
         else:
-            hidden = torch.mean(processed_actions, dim=1)
+            hidden = torch.mean(processed_actions.float(), dim=1)
 
         action_encoding_vector = torch.squeeze(hidden)
         assert action_encoding_vector.shape[0] == batch_dim, \

src/il_representations/algos/representation_learner.py

@@ -290,7 +290,6 @@ def learn(self, datasets, batches_per_epoch, n_epochs, n_trajs=None, callbacks=(
             for step, batch in enumerate(dataloader):
                 # Construct batch (currently just using Torch's default batch-creator)
                 contexts, targets, traj_ts_info, extra_context = self.unpack_batch(batch)
-
                 # Use an algorithm-specific augmentation strategy to augment either
                 # just context, or both context and targets
                 contexts, targets = self._prep_tensors(contexts), self._prep_tensors(targets)

src/il_representations/data/write_dataset.py

@@ -59,6 +59,7 @@ def write_frames(out_file_path, meta_dict, frame_dicts, n_traj=None):
     with wds.TarWriter(out_file_path, keep_meta=True, compress=True) \
           as writer:  # noqa: E207
         # first write _metadata.meta.pickle containing the benchmark config
+        meta_dict['frames'] = len(frame_dicts)
         writer.dwrite(key='_metadata', meta_pickle=meta_dict)
         # now write each frame in each trajectory
         for frame_num, frame_dict in enumerate(frame_dicts):

src/il_representations/envs/auto.py

@@ -1,6 +1,7 @@
 """Code for automatically loading data, creating vecenvs, etc. based on
 Sacred configuration."""
 
+from functools import partial
 import glob
 import logging
 import os
@@ -18,10 +19,10 @@
 from il_representations.envs.dm_control_envs import load_dataset_dm_control
 from il_representations.envs.magical_envs import (get_env_name_magical,
                                                   load_dataset_magical)
-from il_representations.envs.minecraft_envs import (MinecraftVectorWrapper,
-                                                    get_env_name_minecraft,
+from il_representations.envs.minecraft_envs import (get_env_name_minecraft,
                                                     load_dataset_minecraft)
 from il_representations.scripts.utils import update as dict_update
+from il_representations.envs.utils import wrap_env
 
 ERROR_MESSAGE = "no support for benchmark_name={benchmark_name!r}"
 
@@ -53,6 +54,7 @@ def benchmark_is_available(benchmark_name):
         # we check whether minecraft is installed by importing minerl
         try:
             import minerl  # noqa: F401
+            import realistic_benchmarks
             return True, None
         except ImportError as ex:
             return False, "MineRL not installed, cannot use Minecraft " \
@@ -62,7 +64,7 @@ def benchmark_is_available(benchmark_name):
 
 
 @env_cfg_ingredient.capture
-def load_dict_dataset(benchmark_name, n_traj=None):
+def load_dict_dataset(benchmark_name, n_traj, frames_per_traj):
     """Load a dict-type dataset. Also see load_wds_datasets, which instead
     lods a set of datasets that have been stored in a webdataset-compatible
     format."""
@@ -73,7 +75,7 @@ def load_dict_dataset(benchmark_name, n_traj=None):
     elif benchmark_name == 'atari':
         dataset_dict = load_dataset_atari(n_traj=n_traj)
     elif benchmark_name == 'minecraft':
-        dataset_dict = load_dataset_minecraft(n_traj=n_traj)
+        dataset_dict = load_dataset_minecraft(n_traj=n_traj, frames_per_traj=frames_per_traj)
     else:
         raise NotImplementedError(ERROR_MESSAGE.format(**locals()))
 
@@ -110,9 +112,10 @@ def _get_venv_opts(n_envs, venv_parallel, parallel_workers):
     return n_envs, venv_parallel, parallel_workers
 
 
+
 @env_cfg_ingredient.capture
 def load_vec_env(benchmark_name, dm_control_full_env_names,
-                 dm_control_frame_stack, minecraft_max_env_steps):
+                 dm_control_frame_stack, minecraft_max_env_steps, minecraft_wrappers):
     """Create a vec env for the selected benchmark task and wrap it with any
     necessary wrappers."""
     n_envs, venv_parallel, parallel_workers = _get_venv_opts()
@@ -161,7 +164,7 @@ def load_vec_env(benchmark_name, dm_control_full_env_names,
                             n_envs=1,  # TODO fix this eventually; currently hitting error
                                        # noted here: https://github.com/minerllabs/minerl/issues/177
                             parallel=venv_parallel,
-                            wrapper_class=MinecraftVectorWrapper,
+                            wrapper_class=partial(wrap_env, wrappers=minecraft_wrappers),
                             max_episode_steps=minecraft_max_env_steps)
     raise NotImplementedError(ERROR_MESSAGE.format(**locals()))
 

src/il_representations/envs/config.py

@@ -3,8 +3,15 @@
 vecenvs (`venv_opts_ingredient`), and for loading data
 (`env_data_ingredient`)."""
 import os
-
+import logging
 from sacred import Ingredient
+from il_representations.envs.utils import MinecraftPOVWrapper, Testing2500StepLimitWrapper
+try:
+    import realistic_benchmarks.wrappers as rb_wrappers
+except ImportError as e:
+    print(f"Hit error: {e}")
+    logging.info("Realistic Benchmarks is not installed; as a result much Minecraft functionality will not work")
+
 
 ALL_BENCHMARK_NAMES = {"atari", "magical", "dm_control", "minecraft"}
 
@@ -69,11 +76,49 @@ def env_cfg_defaults():
     # Minecraft-specific config variables
     # ###############################
     minecraft_max_env_steps = None
+    minecraft_wrappers = list()
+    _ = locals()
+    del _
+
+@env_cfg_ingredient.named_config
+def wrappers_frames_only_camera_disc_no_limit():
+    minecraft_wrappers = [rb_wrappers.CameraDiscretizationWrapper, rb_wrappers.ActionFlatteningWrapper,
+                          MinecraftPOVWrapper] #
+    _ = locals()
+    del _
+
+@env_cfg_ingredient.named_config
+def wrappers_frames_only_camera_disc():
+    minecraft_wrappers = [rb_wrappers.CameraDiscretizationWrapper, rb_wrappers.ActionFlatteningWrapper,
+                          MinecraftPOVWrapper, Testing2500StepLimitWrapper] #
+    _ = locals()
+    del _
 
+@env_cfg_ingredient.named_config
+def treechop_wrappers_with_frameskip():
+    minecraft_wrappers = [rb_wrappers.CameraDiscretizationWrapper, rb_wrappers.ActionFlatteningWrapper,
+                          MinecraftPOVWrapper, rb_wrappers.FrameSkip] #
     _ = locals()
     del _
 
 
+@env_cfg_ingredient.named_config
+def wrappers_frames_only_obfuscated():
+    minecraft_wrappers = [rb_wrappers.ActionFlatteningWrapper,
+                          MinecraftPOVWrapper, Testing2500StepLimitWrapper] #
+    _ = locals()
+    del _
+
+@env_cfg_ingredient.named_config
+def wrappers_obs_flatten_camera_disc():
+    minecraft_wrappers = [rb_wrappers.CameraDiscretizationWrapper, rb_wrappers.ActionFlatteningWrapper,
+                          rb_wrappers.ObservationFlatteningWrapper, Testing2500StepLimitWrapper]
+    _ = locals()
+    del _
+
+
+
+
 # see venv_opts_defaults docstring for description of this ingredient
 venv_opts_ingredient = Ingredient('venv_opts')