Fixing pipeline on master (HumanCompatibleAI#706)

* Properly remove directory-based demonstrations.

* Update testdata cartpole expert to newest sb3 version.

* Invalidate vindows dependencies cache.

* Pin version of setuptools and wheel

* Pin setuptools for readthedocs

* Pin setuptools for readthedocs for real

* Add todos

* Fix Windows Error on deletion

* Fix lint error

* Install the CPU-only version of pytorch to save cache memory.

---------

Co-authored-by: Adam Gleave <[email protected]>
Co-authored-by: Mohammad Taufeeque <[email protected]>

.circleci/config.yml

@@ -138,7 +138,7 @@ commands:
       # Download and cache dependencies
       - restore_cache:
           keys:
-            - v9win-dependencies-{{ checksum "setup.py" }}-{{ checksum "ci/build_and_activate_venv.ps1" }}
+            - v10win-dependencies-{{ checksum "setup.py" }}-{{ checksum "ci/build_and_activate_venv.ps1" }}
 
       - run:
           name: install python and binary dependencies
@@ -168,7 +168,7 @@ commands:
       - save_cache:
           paths:
             - .\venv
-          key: v9win-dependencies-{{ checksum "setup.py" }}-{{ checksum "ci/build_and_activate_venv.ps1" }}
+          key: v10win-dependencies-{{ checksum "setup.py" }}-{{ checksum "ci/build_and_activate_venv.ps1" }}
 
       - run:
           name: install imitation

ci/build_and_activate_venv.ps1

@@ -9,6 +9,8 @@ If ($venv -eq $null) {
 
 virtualenv -p python3.8 $venv
 & $venv\Scripts\activate
-# Note: We need to install setuptools==66.1.1 to allow installing gym==0.21.0.
-python -m pip install --upgrade pip setuptools==66.1.1
+# Note: We need to install these versions of setuptools and wheel to allow installing gym==0.21.0 on Windows.
+# See https://github.com/freqtrade/freqtrade/issues/8376
+# TODO(GH#707): remove pin once upgraded Gym
+python -m pip install --upgrade pip wheel==0.38.4 setuptools==65.5.1
 pip install ".[docs,parallel,test]"

ci/build_and_activate_venv.sh

@@ -22,4 +22,12 @@ virtualenv -p ${python_version} ${venv}
 source ${venv}/bin/activate
 # Note: We need to install setuptools==66.1.1 to allow installing gym==0.21.0.
 python -m pip install --upgrade pip setuptools==66.1.1
+
+# If platform is linux, install pytorch CPU version.
+# This will prevent installing the CUDA version in the pip install ".[docs,parallel,test]" command.
+# The CUDA version is a couple of gigabytes larger than the CPU version.
+# Since we don't need the CUDA version for testing, we can save some time by not installing it.
+if [[ "$OSTYPE" == "linux-gnu"* ]]; then
+  pip install torch --index-url https://download.pytorch.org/whl/cpu
+fi
 pip install ".[docs,parallel,test]"

docs/requirements.txt

@@ -0,0 +1,2 @@
+# specific version needed for gym==0.21.0
+setuptools==65.5.0

readthedocs.yml

@@ -12,6 +12,8 @@ build:
 
 python:
   install:
+    # TODO(GH#707): remove docs/requirements.txt once Gym upgraded
+    - requirements: docs/requirements.txt
     - method: pip
       path: .
       extra_requirements:

src/imitation/testing/expert_trajectories.py

@@ -2,6 +2,7 @@
 import math
 import pathlib
 import pickle
+import shutil
 import warnings
 from os import PathLike
 from pathlib import Path
@@ -73,10 +74,10 @@ def lazy_generate_expert_trajectories(
     environment_cache_path = pathlib.Path(cache_path) / hfsb3.EnvironmentName(env_id)
     environment_cache_path.mkdir(parents=True, exist_ok=True)
 
-    trajectories_path = environment_cache_path / "rollout.npz"
+    trajectories_path = environment_cache_path / "rollout"
 
     # Note: we cast to str here because FileLock doesn't support pathlib.Path.
-    with FileLock(str(environment_cache_path / "rollout.npz.lock")):
+    with FileLock(str(environment_cache_path / "rollout.lock")):
         try:
             trajectories = data_serialize.load_with_rewards(trajectories_path)
         except (FileNotFoundError, pickle.PickleError) as e:  # pragma: no cover
@@ -96,7 +97,13 @@ def lazy_generate_expert_trajectories(
         return trajectories[:num_trajectories]
     else:  # pragma: no cover
         # If it is not enough, just throw away the cache and generate more.
-        trajectories_path.unlink()
+        if trajectories_path.is_dir():
+            # rmtree won't remove directory on Windows
+            # until the last handle to the directory is closed
+            del trajectories
+            shutil.rmtree(trajectories_path)
+        else:
+            trajectories_path.unlink()
         return lazy_generate_expert_trajectories(
             cache_path,
             env_id,