[Feature] Make torchrl runnable without functorch and with gym==0.13 (p…

…ytorch#386)
ai4co · Oct 28, 2022 · af104c3 · af104c3
1 parent 6f9faaa
commit af104c3
Show file tree

Hide file tree

Showing 38 changed files with 2,063 additions and 356 deletions.
diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -501,7 +501,7 @@ jobs:
       - run:
           name: Install torchrl
 #          command: bash .circleci/unittest/linux_stable/scripts/install.sh
-          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux/scripts/install.sh
+          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_stable/scripts/install.sh
       - run:
           name: Run tests
           command: bash .circleci/unittest/linux_stable/scripts/run_test.sh
@@ -516,6 +516,53 @@ jobs:
       - store_test_results:
           path: test-results
 
+  unittest_linux_olddeps_gpu:
+    <<: *binary_common
+    machine:
+      image: ubuntu-2004-cuda-11.4:202110-01
+    resource_class: gpu.nvidia.medium
+    environment:
+      image_name: "pytorch/manylinux-cuda113"
+      TAR_OPTIONS: --no-same-owner
+      PYTHON_VERSION: << parameters.python_version >>
+      CU_VERSION: << parameters.cu_version >>
+
+    steps:
+      - checkout
+      - designate_upload_channel
+      - run:
+          name: Generate cache key
+          # This will refresh cache on Sundays, nightly build should generate new cache.
+          command: echo "$(date +"%Y-%U")" > .circleci-weekly
+      - restore_cache:
+          keys:
+            - env-v2-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_olddeps/scripts_gym_0_13/environment.yml" }}-{{ checksum ".circleci-weekly" }}
+      - run:
+          name: Setup
+          command: docker run -e PYTHON_VERSION -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux_olddeps/scripts_gym_0_13/setup_env.sh
+      - save_cache:
+          key: env-v2-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_olddeps/scripts_gym_0_13/environment.yml" }}-{{ checksum ".circleci-weekly" }}
+          paths:
+            - conda
+            - env
+      - run:
+          name: Install torchrl, run tests, upload codecov
+          command: |
+            docker run -t --env=CUDA_VISIBLE_DEVICES="" --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_olddeps/scripts_gym_0_13/batch_scripts.sh
+#            docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_olddeps/scripts_gym_0_13/batch_scripts.sh
+#      - run:
+#          name: Run tests
+#          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_olddeps/scripts_gym_0_13/run_test.sh
+#      - run:
+#          name: Codecov upload
+#          command: |
+#            docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" <(curl -s https://codecov.io/bash) -Z -F linux-stable-cpu
+#      - run:
+#          name: Post process
+#          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_olddeps/scripts_gym_0_13/post_process.sh
+      - store_test_results:
+          path: test-results
+
   unittest_macos_cpu:
     <<: *binary_common
     macos:
@@ -787,3 +834,8 @@ workflows:
           cu_version: cu113
           name: unittest_linux_stable_gpu_py3.10
           python_version: '3.10'
+
+      - unittest_linux_olddeps_gpu:
+          cu_version: cu113
+          name: unittest_linux_olddeps_gpu_py3.9
+          python_version: '3.9'
diff --git a/.circleci/unittest/linux/scripts/run_test.sh b/.circleci/unittest/linux/scripts/run_test.sh
@@ -1,5 +1,11 @@
 #!/usr/bin/env bash
 
+# this code is supposed to run on CPU
+# rendering with the combination of packages we have here in headless mode
+# is hard to nail.
+# IMPORTANT: As a consequence, we can't guarantee TorchRL compatibility with
+# rendering with this version of gym / mujoco-py.
+
 set -e
 
 eval "$(./conda/bin/conda shell.bash hook)"
@@ -19,6 +25,6 @@ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$lib_dir
 export MKL_THREADING_LAYER=GNU
 
 coverage run -m pytest test/smoke_test.py -v --durations 20
-coverage run -m pytest test/smoke_test_deps.py -v --durations 20
+coverage run -m pytest test/smoke_test_deps.py -v --durations 20 -k 'test_gym or test_dm_control_pixels or test_dm_control or test_tb'
 coverage run -m pytest --instafail -v --durations 20
 coverage xml
diff --git a/.circleci/unittest/linux_olddeps/scripts_gym_0_13/batch_scripts.sh b/.circleci/unittest/linux_olddeps/scripts_gym_0_13/batch_scripts.sh
@@ -0,0 +1,10 @@
+#!/usr/bin/env bash
+
+# Runs a batch of scripts in a row to allow docker run to keep installed libraries
+# and env variables across runs.
+
+DIR="$(cd "$(dirname "$0")" && pwd)"
+
+$DIR/install.sh
+$DIR/run_test.sh
+$DIR/post_process.sh
diff --git a/.circleci/unittest/linux_olddeps/scripts_gym_0_13/environment.yml b/.circleci/unittest/linux_olddeps/scripts_gym_0_13/environment.yml
@@ -0,0 +1,26 @@
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - pip
+  - protobuf
+  - pip:
+    - hypothesis
+    - future
+    - cloudpickle
+    - gym==0.13
+    - pygame
+    - gym[atari]
+    - moviepy
+    - tqdm
+    - pytest
+    - pytest-cov
+    - pytest-mock
+    - pytest-instafail
+    - expecttest
+    - pyyaml
+    - scipy
+    - hydra-core
+    - dm_control -e git+https://github.com/deepmind/dm_control.git@c053360edea6170acfd9c8f65446703307d9d352#egg={dm_control}
+    - patchelf
+    - pyopengl==3.1.4
diff --git a/.circleci/unittest/linux_olddeps/scripts_gym_0_13/install.sh b/.circleci/unittest/linux_olddeps/scripts_gym_0_13/install.sh
@@ -0,0 +1,45 @@
+#!/usr/bin/env bash
+
+unset PYTORCH_VERSION
+# For unittest, nightly PyTorch is used as the following section,
+# so no need to set PYTORCH_VERSION.
+# In fact, keeping PYTORCH_VERSION forces us to hardcode PyTorch version in config.
+
+set -e
+
+eval "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env
+
+#apt-get update -y && apt-get install git wget gcc g++ -y
+
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    cudatoolkit="cpuonly"
+    version="cpu"
+else
+    if [[ ${#CU_VERSION} -eq 4 ]]; then
+        CUDA_VERSION="${CU_VERSION:2:1}.${CU_VERSION:3:1}"
+    elif [[ ${#CU_VERSION} -eq 5 ]]; then
+        CUDA_VERSION="${CU_VERSION:2:2}.${CU_VERSION:4:1}"
+    fi
+    echo "Using CUDA $CUDA_VERSION as determined by CU_VERSION ($CU_VERSION)"
+    version="$(python -c "print('.'.join(\"${CUDA_VERSION}\".split('.')[:2]))")"
+    cudatoolkit="cudatoolkit=${version}"
+fi
+
+case "$(uname -s)" in
+    Darwin*) os=MacOSX;;
+    *) os=Linux
+esac
+
+# submodules
+git submodule sync && git submodule update --init --recursive
+
+printf "Installing PyTorch with %s\n" "${CU_VERSION}"
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cpuonly -c pytorch -y
+else
+    conda install pytorch==1.10.0 torchvision==0.11.0 torchaudio==0.10.0 cudatoolkit=11.3 -c pytorch -c conda-forge -y
+fi
+
+printf "* Installing torchrl\n"
+python setup.py develop
diff --git a/.circleci/unittest/linux_olddeps/scripts_gym_0_13/post_process.sh b/.circleci/unittest/linux_olddeps/scripts_gym_0_13/post_process.sh
@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+
+set -e
+
+eval "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env