job script and image for bacalhau

Author: devinaconley

README.md

@@ -26,7 +26,7 @@ Copy `.env.template` to `.env` and define variables
 
 Update brownie configuration as needed for RPC access
 
-For Golem support, also install the [yagna](https://docs.golem.network/docs/creators/python/examples/tools/yagna-installation-for-requestors) CLI and service
+For Golem support, also install the [yagna](https://docs.golem.network/docs/quickstarts/python-quickstart) CLI and service
 
 
 ## Run
@@ -40,7 +40,7 @@ Once the trained model has been uploaded to IPFS, define the `IPFS_MODEL_HASH` v
 
 Activate appropriate virtual environment (e.g. for ocean)
 ```
-python activate mlweb3-ocean
+conda activate mlweb3-ocean
 ```
 
 Deploy model to web3 infrastructure

environment.yml

@@ -25,5 +25,6 @@ dependencies:
   - zlib
   - zstd
   - pip:
+    - numpy<2
     - python-dotenv
 prefix: /home/devin/miniconda3/envs/mlweb3

etc/requirements/environment-bacalhau.yml

@@ -0,0 +1,30 @@
+name: mlweb3-bacalhau
+channels:
+  - pytorch
+  - conda-forge
+  - defaults
+dependencies:
+  - bzip2
+  - ca-certificates
+  - pip
+  - python=3.10
+  - pytorch=1.12.1
+  - pytorch-mutex
+  - readline
+  - requests
+  - six
+  - sqlite
+  - tk
+  - torchvision=0.13.1
+  - typing_extensions
+  - tzdata
+  - urllib3
+  - wheel
+  - xz
+  - zlib
+  - zstd
+  - pip:
+    - python-dotenv
+    # ---- bacalhau ----
+    - bacalhau-sdk
+prefix: /home/devin/miniconda3/envs/mlweb3-bacalhau

etc/requirements/environment-golem.yml

@@ -29,5 +29,6 @@ dependencies:
     # ---- golem ----
     - flask~=3.0.1
     - gvmkit-build
+    - numpy<2
     - yapapi~=0.12.0
 prefix: /home/devin/miniconda3/envs/mlweb3-golem

mlweb3/bacalhau/Dockerfile

@@ -0,0 +1,23 @@
+# dockerfile for mnist classification job on bacalhau
+FROM nvidia/cuda:11.3.1-cudnn8-runtime
+
+# patch: https://github.com/NVIDIA/nvidia-container-toolkit/issues/258
+RUN rm /etc/apt/sources.list.d/cuda.list
+RUN rm /etc/apt/sources.list.d/nvidia-ml.list
+
+# install python
+RUN apt-get update && \
+    apt-get install -y python3 python3-pip && \
+    rm -rf /var/lib/apt/lists/*
+RUN python3 --version
+
+# install packages
+RUN pip install --no-cache-dir torch torchvision --index-url https://download.pytorch.org/whl/cu113
+
+# app
+WORKDIR /mlweb3
+COPY mlweb3/bacalhau/job.py job.py
+COPY mlweb3/model.py mlweb3/model.py
+COPY etc/models/cnn_mnist.pth etc/models/cnn_mnist.pth
+
+CMD python3 job.py

mlweb3/bacalhau/__init__.py

@@ -0,0 +1,7 @@
+"""
+deployment logic for bacalhau
+"""
+
+
+def deploy():
+    print('no bacalhau deployment needed.')

mlweb3/bacalhau/deployment.py

@@ -0,0 +1,7 @@
+"""
+inference logic for bacalhau
+"""
+
+
+def predict():
+    raise NotImplemented('bacalhau predict not implemented.')

mlweb3/bacalhau/inference.py

@@ -0,0 +1,49 @@
+"""
+inference logic for bacalhau job
+"""
+
+# lib
+import os
+import torch
+from torch.utils.data import DataLoader
+from torchvision import datasets
+from torchvision.transforms import ToTensor, Compose
+
+# src
+from mlweb3.model import SimpleCNN
+
+
+def main():
+    device = 'cuda' if torch.cuda.is_available() else 'cpu'
+    print(f'using device {device}')
+
+    # load model
+    print(f'loading SimpleCNN...')
+    model = SimpleCNN()
+    model.load_state_dict(torch.load('./etc/models/cnn_mnist.pth', map_location=torch.device(device)))
+    model.eval()
+    model = model.to(device)
+
+    # get data
+    os.makedirs('./etc/mnist', exist_ok=True)
+    data = DataLoader(
+        datasets.MNIST('./etc/mnist', train=False, download=True, transform=ToTensor()),
+        batch_size=64
+    )
+
+    # do inference
+    correct, total = 0, 0
+    predictions = []
+    with torch.no_grad():
+        for X, y in data:
+            X, y = X.to(device), y.to(device)
+            pred = model(X)
+            correct += (pred.argmax(1) == y).sum().item()
+            total += len(X)
+            predictions.extend(pred.argmax(1).cpu().numpy().tolist())
+
+    print(f'test:\n  correct: {correct}\n  total: {total}\n  accuracy: {correct / total:>0.4f}')
+
+
+if __name__ == '__main__':
+    main()