Add optional grain dependency & test (#9)

.github/workflows/test.yaml

@@ -12,6 +12,8 @@ on:
 permissions:
   contents: read  # to fetch code
 
+# TODO(jakevdp): add testing on macOS-11 and windows-2019 when grain supports them,
+# or alternatively run a subset of tests without grain.
 jobs:
   built-latest:
     name: Latest packages (${{ matrix.os }} Python ${{ matrix.python-version }})
@@ -20,9 +22,6 @@ jobs:
       matrix:
         os: ["ubuntu-latest"]
         python-version: ["3.10", "3.11", "3.12"]
-        include:
-        - os: windows-2019
-          python-version: "3.11"
 
     steps:
     - uses: actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # ratchet:actions/checkout@v4
@@ -35,8 +34,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install .[dev,tfds]
-        pip install -U jax flax optax orbax tensorflow tensorflow_datasets
+        pip install -U jax flax grain optax orbax tensorflow tensorflow_datasets pytest pytest-xdist
     - name: Run tests
       run: |
         pytest -n auto jax_ml_stack
@@ -60,7 +58,7 @@ jobs:
     - name: Install dependencies
       run: |
         python -m pip install --upgrade pip
-        pip install .[dev,tfds]
+        pip install .[dev,tfds,grain]
     - name: Run tests
       run: |
         pytest -n auto jax_ml_stack

jax_ml_stack/tests/test_nnx_with_tfds.py

@@ -16,6 +16,7 @@
 import functools
 import unittest
 from flax import nnx
+import grain.python as grain
 import numpy as np
 import optax
 import tensorflow_datasets as tfds
@@ -69,6 +70,52 @@ def loss_fn(model, batch):
       _, grads = grad_fn(model, batch)
       optimizer.update(grads)
 
+  def test_nnx_with_tfds_plus_grain(self):
+    data_source = tfds.data_source('mnist', split='train')
+
+    sampler = grain.IndexSampler(
+        num_records=5,
+        num_epochs=1,
+        shard_options=grain.NoSharding(),
+        shuffle=True,
+        seed=0,
+    )
+
+    class DownSample(grain.MapTransform):
+      shape: tuple[int, int]
+
+      def __init__(self, shape: tuple[int, int]):
+        self.shape = shape
+
+      def map(self, element: dict[str, np.ndarray]) -> dict[str, np.ndarray]:
+        image = element['image']
+        element['image_scaled'] = image - image.mean()
+        return element
+
+    operations = [DownSample((16, 16))]
+
+    loader = grain.DataLoader(
+        data_source=data_source,
+        operations=operations,
+        sampler=sampler,
+        worker_count=0,  # Scale to multiple workers in multiprocessing
+    )
+
+    model = CNN(rngs=nnx.Rngs(0))
+    optimizer = nnx.Optimizer(model, optax.adamw(learning_rate=0.005))
+
+    def loss_fn(model, batch):
+      logits = model(batch['image_scaled'])
+      loss = optax.softmax_cross_entropy_with_integer_labels(
+          logits=logits, labels=np.ravel(batch['label'])
+      ).mean()
+      return loss, logits
+
+    for batch in loader:
+      grad_fn = nnx.value_and_grad(loss_fn, has_aux=True)
+      _, grads = grad_fn(model, batch)
+      optimizer.update(grads)
+
 
 if __name__ == '__main__':
   unittest.main()

pyproject.toml

@@ -17,6 +17,7 @@ keywords = []
 dependencies = [
     "jax==0.4.31",
     "flax==0.8.5",
+    "grain==0.2.0",
     "optax==0.2.3",
     "orbax==0.1.9",
 ]
@@ -33,11 +34,18 @@ dev = [
     "pytest",
     "pytest-xdist",
 ]
+
+# TensorFlow datasets is an extra because it has a large install footprint.
 tfds = [
     "tensorflow==2.17.0",
     "tensorflow_datasets==4.9.6",
 ]
 
+# Grain is an extra because as of v0.2.0 it has no OSX wheels.
+grain = [
+    "grain==0.2.0",
+]
+
 [tool.pyink]
 # Formatting configuration to follow Google style-guide
 line-length = 80