TileDB-ML support for carrara

.github/workflows/ci.yml

@@ -9,10 +9,8 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python-version: ["3.9"]
-        ml-deps:
-          - "torch==2.1.2+cpu torchvision==0.16.2+cpu torchdata==0.7.1 tensorflow-cpu==2.15.1 "
-          - "torch==2.3.1+cpu torchvision==0.18.1+cpu torchdata==0.8.0 'tensorflow-cpu<2.16'"
+        python-version: ["3.12"]
+
     env:
       run_coverage: ${{ github.ref == 'refs/heads/master' }}
 
@@ -28,14 +26,13 @@ jobs:
       uses: actions/cache@v4
       with:
         path: ~/.cache/pip
-        key: ${{ runner.os }}:ml-deps=[${{ matrix.ml-deps }}]
+        key: ${{ runner.os }} # :ml-deps=[${{ matrix.ml-deps }}]
 
     - name: Install dependencies
       run: |
-        pip install --upgrade pip
-        pip install --extra-index-url https://download.pytorch.org/whl/cpu ${{ matrix.ml-deps }}
-        pip install pytest-mock pytest-cov scikit-learn==1.0.2
-        pip install -e .[cloud]
+        pip install --upgrade pip setuptools wheel
+        pip install pytest-mock pytest-cov
+        pip install -e .[full]
 
     - name: Run pre-commit hooks
       run: |
@@ -45,6 +42,7 @@ jobs:
     - name: Run notebook examples
       run: |
         pip install pytest-xdist nbmake matplotlib idx2numpy "numpy<2"
+        rm -rf examples/data/openml
         pytest --disable-warnings --nbmake examples/{models,readers}
         # Run tiledb-cloud in parallel
         if [[ "${{ secrets.TILEDB_API_TOKEN }}" != "" ]]; then

examples/cloud/serverless_training/pytorch/README.md

@@ -6,7 +6,7 @@ Firstly, we ingest all training images and labels in TileDB arrays and register
 and finally, we serverless-ly get some predictions using the trained model. In case you want to run the example, you will need a TileDB-Cloud account as described
 [here](https://docs.tiledb.com/cloud/tutorials/start-here). After signing up, you should export your username and password 
 as environmental variables (**TILEDB_USER_NAME**, **TILEDB_PASSWD**), in order to run ingestion, model training and prediction UDFs. Moreover,
-please add your TileDB namespace and your **S3** bucket in each script.
+please add your TileDB teamspace and your **S3** bucket in each script.
 
 # Steps
 

examples/cloud/serverless_training/pytorch/data_ingestion.py

@@ -2,20 +2,25 @@
 from typing import Any
 
 import numpy as np
-import tiledb.cloud
+import tiledb.client
 
 # Your TileDB username and password, exported as environmental variables
 TILEDB_USER_NAME = os.environ.get("TILEDB_USER_NAME")
 TILEDB_PASSWD = os.environ.get("TILEDB_PASSWD")
 
-# Your TileDB namespace
-TILEDB_NAMESPACE = "your_tiledb_namespace"
+# Your TileDB workspace/teamspace
+TILEDB_WORKSPACE = "your_tiledb_WORKSPACE"
+TILEDB_TEAMSPACE = "your_tiledb_TEAMSPACE"
 
 # Your S3 bucket
 S3_BUCKET = "your_s3_bucket"
 
-IMAGES_URI = "tiledb://{}/s3://{}/mnist_images".format(TILEDB_NAMESPACE, S3_BUCKET)
-LABELS_URI = "tiledb://{}/s3://{}/mnist_labels".format(TILEDB_NAMESPACE, S3_BUCKET)
+IMAGES_URI = (
+    f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/s3://{S3_BUCKET}/mnist_images"
+)
+LABELS_URI = (
+    f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/s3://{S3_BUCKET}/mnist_labels"
+)
 
 
 # Let's define an ingestion function
@@ -62,8 +67,11 @@ def mnist_ingest(ingestion_func: Any) -> None:
     ingestion_func(data=labels, batch_size=64, uri=LABELS_URI)
 
 
-tiledb.cloud.login(username=TILEDB_USER_NAME, password=TILEDB_PASSWD)
+tiledb.client.configure(
+    username=TILEDB_USER_NAME, password=TILEDB_PASSWD, workspace=TILEDB_WORKSPACE
+)
+tiledb.client.login()
 
-tiledb.cloud.udf.exec(mnist_ingest, ingestion_func=ingest_in_tiledb)
+tiledb.client.udf.exec(mnist_ingest, ingestion_func=ingest_in_tiledb)
 
-print(tiledb.cloud.last_udf_task().logs)
+print(tiledb.client.last_udf_task().logs)

examples/cloud/serverless_training/pytorch/model_load_and_predict.py

@@ -1,21 +1,28 @@
 import os
 from typing import List
 
-import tiledb.cloud
+import tiledb.client
 
 # Your TileDB username and password, exported as environmental variables
 TILEDB_USER_NAME = os.environ.get("TILEDB_USER_NAME")
 TILEDB_PASSWD = os.environ.get("TILEDB_PASSWD")
 
 # Your TileDB namespace
-TILEDB_NAMESPACE = "your_tiledb_namespace"
+TILEDB_WORKSPACE = "your_tiledb_WORKSPACE"
+TILEDB_TEAMSPACE = "your_tiledb_TEAMSPACE"
 
 # Your S3 bucket
 S3_BUCKET = "your_s3_bucket"
 
-IMAGES_URI = "tiledb://{}/s3://{}/mnist_images".format(TILEDB_NAMESPACE, S3_BUCKET)
-LABELS_URI = "tiledb://{}/s3://{}/mnist_labels".format(TILEDB_NAMESPACE, S3_BUCKET)
-MODEL_URI = "tiledb://{}/s3://{}/mnist_model".format(TILEDB_NAMESPACE, S3_BUCKET)
+IMAGES_URI = (
+    f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/s3://{S3_BUCKET}/mnist_images"
+)
+LABELS_URI = (
+    f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/s3://{S3_BUCKET}/mnist_labels"
+)
+MODEL_URI = (
+    f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/s3://{S3_BUCKET}/mnist_model"
+)
 
 IO_BATCH_SIZE = 20000
 
@@ -70,8 +77,11 @@ def forward(self, x: torch.Tensor) -> Any:
     return [np.argmax(pred) for pred in output.numpy()]
 
 
-tiledb.cloud.login(username=TILEDB_USER_NAME, password=TILEDB_PASSWD)
+tiledb.client.configure(
+    username=TILEDB_USER_NAME, password=TILEDB_PASSWD, workspace=TILEDB_WORKSPACE
+)
+tiledb.client.login()
 
-predictions = tiledb.cloud.udf.exec(predict)
+predictions = tiledb.client.udf.exec(predict)
 
 print(predictions)

examples/cloud/serverless_training/pytorch/model_training.py

@@ -1,20 +1,20 @@
 import os
 
-import tiledb.cloud
+import tiledb.client
+
+from tiledb.ml.readers.types import ArrayParams
 
 # Your TileDB username and password, exported as environmental variables
 TILEDB_USER_NAME = os.environ.get("TILEDB_USER_NAME")
 TILEDB_PASSWD = os.environ.get("TILEDB_PASSWD")
 
 # Your TileDB namespace
-TILEDB_NAMESPACE = "your_tiledb_namespace"
-
-# Your S3 bucket
-S3_BUCKET = "your_s3_bucket"
+TILEDB_WORKSPACE = "TileDB-Inc."
+TILEDB_TEAMSPACE = "tiledb-ml-ts"
 
-IMAGES_URI = "tiledb://{}/s3://{}/mnist_images".format(TILEDB_NAMESPACE, S3_BUCKET)
-LABELS_URI = "tiledb://{}/s3://{}/mnist_labels".format(TILEDB_NAMESPACE, S3_BUCKET)
-MODEL_URI = "tiledb://{}/s3://{}/mnist_model".format(TILEDB_NAMESPACE, S3_BUCKET)
+IMAGES_URI = f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/mnist_images"
+LABELS_URI = f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/mnist_labels"
+MODEL_URI = f"tiledb://{TILEDB_WORKSPACE}/{TILEDB_TEAMSPACE}/mnist_model"
 
 # The size of each slice from a image and label TileDB arrays.
 IO_BATCH_SIZE = 20000
@@ -49,8 +49,20 @@ def forward(self, x: torch.Tensor) -> Any:
             logits = self.linear_relu_stack(x)
             return logits
 
+    def do_random_noise(img: np.ndarray, mag: float = 0.1) -> np.ndarray:
+        noise = np.random.uniform(-1, 1, img.shape) * mag
+        img = img + noise
+        img = np.clip(img, 0, 1)
+        return img
+
     with tiledb.open(IMAGES_URI) as x, tiledb.open(LABELS_URI) as y:
-        train_loader = PyTorchTileDBDataLoader(x, y, batch_size=IO_BATCH_SIZE)
+        train_loader = PyTorchTileDBDataLoader(
+            ArrayParams(x, fn=do_random_noise),
+            ArrayParams(y),
+            batch_size=IO_BATCH_SIZE,
+            num_workers=0,
+            shuffle_buffer_size=256,
+        )
 
         net = Net(shape=(28, 28))
         criterion = nn.CrossEntropyLoss()
@@ -95,7 +107,7 @@ def forward(self, x: torch.Tensor) -> Any:
 
         model = PyTorchTileDBModel(
             uri="mnist_model",
-            namespace=TILEDB_NAMESPACE,
+            teamspace=TILEDB_TEAMSPACE,
             model=net,
             optimizer=optimizer,
         )
@@ -104,8 +116,11 @@ def forward(self, x: torch.Tensor) -> Any:
         model.save()
 
 
-tiledb.cloud.login(username=TILEDB_USER_NAME, password=TILEDB_PASSWD)
+# tiledb.client.configure(
+#     username=TILEDB_USER_NAME, password=TILEDB_PASSWD, workspace=TILEDB_WORKSPACE
+# )
+tiledb.client.login()
 
-tiledb.cloud.udf.exec(train)
+tiledb.client.udf.exec(train)
 
-print(tiledb.cloud.last_udf_task().logs)
+print(tiledb.client.last_udf_task().logs)