Simplify installation instructions (#2)

* Add initial conda env and GitHub Actions workflow

* Add PyTorch cpuonly
Without cpuonly the pinned version of openfold fails to
install because it attempts to parse the CUDA version.
Only newer versions of openfold check for CPU environments

* Skip installing openfold
Even with cpuonly, cudatoolkit still installed. Likely due to known
PyG packaging problems reported in their GitHub issues.

* Install fair-esm outside conda
Can't use pip with --no-deps otherwise
https://stackoverflow.com/questions/75164677/pip-install-with-no-deps-in-conda-environment-yml

* Temporarily skip dllogger

* Install cudatoolkit from nvidia channel

* Try older pip to workaround functools lru_cache
Not perfectly compatible with Python 3.7

* Test ProteinDT import
Add back dllogger

* Test importing code that exists

.github/workflows/test.yml

@@ -0,0 +1,45 @@
+name: Tests
+on:
+- push
+- pull_request
+
+jobs:
+  # Installs the conda environment and ProteinDT package
+  install:
+    name: Test ProteinDT installation
+    runs-on: ${{ matrix.os }}
+    strategy:
+      fail-fast: false
+      matrix:
+        # Could also test on the beta M1 macOS or other runners
+        # https://docs.github.com/en/actions/using-github-hosted-runners/about-github-hosted-runners/about-github-hosted-runners#standard-github-hosted-runners-for-public-repositories
+        os:
+        - ubuntu-latest
+    steps:
+    - name: Checkout repository
+      uses: actions/checkout@v4
+    # Use mamba instead of conda
+    - name: Install conda environment
+      uses: conda-incubator/setup-miniconda@v2
+      with:
+        activate-environment: ProteinDT
+        environment-file: environment.yml
+        auto-activate-base: false
+        miniforge-variant: Mambaforge
+        miniforge-version: 'latest'
+        use-mamba: true
+      # Installs fair-esm package into the activated conda environment without dependencies
+    - name: Install fair-esm
+      shell: bash --login {0}
+      run: pip install fair-esm[esmfold]==2.0.0 --no-dependencies # Override deepspeed==0.5
+      # Installs ProteinDT package into the activated conda environment
+    - name: Install ProteinDT
+      shell: bash --login {0}
+      run: pip install .
+      # Log conda environment contents
+    - name: Log conda environment
+      shell: bash --login {0}
+      run: conda list
+    - name: Test ProteinDT import
+      shell: bash --login {0}
+      run: python -c 'from ProteinDT.models.model_ProteinText import ProteinTextModel'

environment.yml

@@ -0,0 +1,32 @@
+name: ProteinDT
+channels:
+  - conda-forge
+dependencies:
+  - python=3.7
+  - pip=18
+  - numpy
+  - networkx
+  - scikit-learn
+  - pytorch::pytorch=1.10
+  - nvidia::cudatoolkit # pyg installs cudatoolkit even if cpuonly is requested
+  - transformers
+  - lxml
+  - lmdb # for TAPE
+  - seqeval
+  - openai # for baseline ChatGPT
+  - accelerate # for baseline Galactica
+  - matplotlib # for visualization
+  - h5py # for binding editing
+  - biopython
+  - pyg::pyg=2.0
+  - pyg::pytorch-scatter
+  - pyg::pytorch-sparse
+  - pyg::pytorch-cluster
+  - dm-tree # for ESM folding
+  - omegaconf
+  - ml-collections
+  - einops
+  - mdtraj
+  - pip:
+      - git+https://github.com/NVIDIA/dllogger@0540a43971f4a8a16693a9de9de73c1072020769
+#      - git+https://github.com/aqlaboratory/openfold@4b41059694619831a7db195b7e0988fc4ff3a307