[Feature] MPPI Planner

.circleci/config.yml

@@ -453,6 +453,51 @@ jobs:
       - store_test_results:
           path: test-results
 
+  unittest_linux_jumanji_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-v3-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_libs/scripts_jumanji/environment.yml" }}-{{ checksum ".circleci-weekly" }}
+      - run:
+          name: Setup
+          command: .circleci/unittest/linux_libs/scripts_jumanji/setup_env.sh
+      - save_cache:
+          key: env-v3-linux-{{ arch }}-py<< parameters.python_version >>-{{ checksum ".circleci/unittest/linux_libs/scripts_jumanji/environment.yml" }}-{{ checksum ".circleci-weekly" }}
+          paths:
+            - conda
+            - env
+      - run:
+          name: Install torchrl
+          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD -e UPLOAD_CHANNEL -e CU_VERSION "${image_name}" .circleci/unittest/linux_libs/scripts_jumanji/install.sh
+      - run:
+          name: Run tests
+          command: bash .circleci/unittest/linux_libs/scripts_jumanji/run_test.sh
+      - run:
+          name: Codecov upload
+          command: |
+            bash <(curl -s https://codecov.io/bash) -Z -F linux-jumanji
+      - run:
+          name: Post Process
+          command: docker run -t --gpus all -v $PWD:$PWD -w $PWD "${image_name}" .circleci/unittest/linux_libs/scripts_jumanji/post_process.sh
+      - store_test_results:
+          path: test-results
+
   unittest_linux_gym_gpu:
     <<: *binary_common
     machine:
@@ -879,6 +924,10 @@ workflows:
           cu_version: cu113
           name: unittest_linux_habitat_gpu_py3.8
           python_version: '3.8'
+      - unittest_linux_jumanji_gpu:
+          cu_version: cu113
+          name: unittest_linux_jumanji_gpu_py3.8
+          python_version: '3.8'
       - unittest_linux_gym_gpu:
           cu_version: cu113
           name: unittest_linux_gym_gpu_py3.8

.circleci/unittest/linux_examples/scripts/run_test.sh

@@ -26,24 +26,76 @@ 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 python examples/ddpg/ddpg.py \
-  total_frames=14 \
+coverage run examples/ddpg/ddpg.py \
+  total_frames=48 \
   init_random_frames=10 \
   batch_size=10 \
   frames_per_batch=16 \
   num_workers=2 \
   env_per_collector=1 \
   collector_devices=cuda:0 \
   optim_steps_per_batch=1 \
-  record_video=True
-coverage run -m python examples/dqn/dqn.py \
-  total_frames=14 \
+  record_video=True \
+  record_frames=4 \
+  buffer_size=120
+coverage run examples/dqn/dqn.py \
+  total_frames=48 \
   init_random_frames=10 \
   batch_size=10 \
   frames_per_batch=16 \
   num_workers=2 \
   env_per_collector=1 \
   collector_devices=cuda:0 \
   optim_steps_per_batch=1 \
-  record_video=True
+  record_video=True \
+  record_frames=4 \
+  buffer_size=120
+coverage run examples/redq/redq.py \
+  total_frames=48 \
+  init_random_frames=10 \
+  batch_size=10 \
+  frames_per_batch=16 \
+  num_workers=2 \
+  env_per_collector=1 \
+  collector_devices=cuda:0 \
+  optim_steps_per_batch=1 \
+  record_video=True \
+  record_frames=4 \
+  buffer_size=120
+coverage run examples/sac/sac.py \
+  total_frames=48 \
+  init_random_frames=10 \
+  batch_size=10 \
+  frames_per_batch=16 \
+  num_workers=2 \
+  env_per_collector=1 \
+  collector_devices=cuda:0 \
+  optim_steps_per_batch=1 \
+  record_video=True \
+  record_frames=4 \
+  buffer_size=120
+coverage run examples/ppo/ppo.py \
+  total_frames=48 \
+  batch_size=10 \
+  frames_per_batch=16 \
+  num_workers=2 \
+  env_per_collector=1 \
+  collector_devices=cuda:0 \
+  optim_steps_per_batch=1 \
+  record_video=True \
+  record_frames=4 \
+  lr_scheduler=
+coverage run examples/dreamer/dreamer.py \
+  total_frames=48 \
+  init_random_frames=10 \
+  batch_size=10 \
+  frames_per_batch=200 \
+  num_workers=2 \
+  env_per_collector=1 \
+  collector_devices=cuda:0 \
+  optim_steps_per_batch=1 \
+  record_video=True \
+  record_frames=4 \
+  buffer_size=120 \
+  rssm_hidden_dim=17
 coverage xml -i

.circleci/unittest/linux_libs/scripts_jumanji/environment.yml

@@ -0,0 +1,18 @@
+channels:
+  - pytorch
+  - defaults
+dependencies:
+  - pip
+  - pip:
+    - hypothesis
+    - future
+    - cloudpickle
+    - pytest
+    - pytest-cov
+    - pytest-mock
+    - pytest-instafail
+    - expecttest
+    - pyyaml
+    - scipy
+    - hydra-core
+    - jumanji

.circleci/unittest/linux_libs/scripts_jumanji/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 "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env
+
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    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]))")"
+fi
+
+# submodules
+git submodule sync && git submodule update --init --recursive
+
+printf "Installing PyTorch with %s\n" "${CU_VERSION}"
+if [ "${CU_VERSION:-}" == cpu ] ; then
+    # conda install -y pytorch torchvision cpuonly -c pytorch-nightly
+    # use pip to install pytorch as conda can frequently pick older release
+#    conda install -y pytorch cpuonly -c pytorch-nightly
+    pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cpu --force-reinstall
+else
+    pip3 install --pre torch --extra-index-url https://download.pytorch.org/whl/nightly/cu116 --force-reinstall
+fi
+
+# smoke test
+python -c "import functorch"
+
+printf "* Installing torchrl\n"
+pip3 install -e .
+
+# smoke test
+python -c "import torchrl"

.circleci/unittest/linux_libs/scripts_jumanji/post_process.sh

@@ -0,0 +1,6 @@
+#!/usr/bin/env bash
+
+set -e
+
+eval "$(./conda/bin/conda shell.bash hook)"
+conda activate ./env