Add missing training support to onert Python API (experimental module)

This commit integrate previously omitted modifications for training support in the onert Python API.
- Expose new experimental training functionalities by updating the package’s public API:
  - Modified `__init__.py` to include the `experimental` submodule.
  - Added the experimental module, which imports train components and exposes `TrainSession`, `traininfo`, `DataLoader`, `optimizer`, `losses`, and `metrics`.
- Implemented a flexible DataLoader in the experimental training module:
  - Supports input from file paths or NumPy arrays.
  - Handles loading of both .npy and raw binary files with configurable shapes and data types.
  - Includes batching logic and a split method for training/validation separation.
- Improved training compiler behavior in `TrainingCompiler.cc`:
  - Adjusted the shape validation to accept unspecified dimensions (using `ir::Shape::kUnspecifiedDim`) in addition to dimensions of value 1.

ONE-DCO-1.0-Signed-off-by: ragmani <ragmani0216@gmail.com>

Author: ragmani

runtime/onert/api/python/package/__init__.py

@@ -1,8 +1,11 @@
 # Define the public API of the onert package
-__all__ = ["infer", "tensorinfo"]
+__all__ = ["infer", "tensorinfo", "experimental"]
 
 # Import and expose the infer module's functionalities
 from . import infer
 
 # Import and expose tensorinfo
 from .common import tensorinfo as tensorinfo
+
+# Import and expose the experimental module's functionalities
+from . import experimental

runtime/onert/api/python/package/experimental/__init__.py

@@ -0,0 +1,3 @@
+__all__ = ["train"]
+
+from . import train

runtime/onert/api/python/package/experimental/train/__init__.py

@@ -0,0 +1,8 @@
+from .session import TrainSession as session
+from onert.native.libnnfw_api_pybind import traininfo
+from .dataloader import DataLoader
+from . import optimizer
+from . import losses
+from . import metrics
+
+__all__ = ["session", "traininfo", "DataLoader", "optimizer", "losses", "metrics"]

runtime/onert/api/python/package/experimental/train/dataloader.py

@@ -0,0 +1,231 @@
+import os
+import numpy as np
+
+
+class DataLoader:
+    """
+    A flexible DataLoader to manage training and validation data.
+    Automatically detects whether inputs are paths or NumPy arrays.
+    """
+    def __init__(self,
+                 input_dataset,
+                 expected_dataset,
+                 batch_size,
+                 input_shape=None,
+                 expected_shape=None,
+                 dtype=np.float32):
+        """
+        Initialize the DataLoader.
+
+        Args:
+            input_dataset (list of np.ndarray): List of input arrays where each array's first dimension is the batch dimension.
+            expected_dataset (list of np.ndarray): List of expected arrays where each array's first dimension is the batch dimension.
+            batch_size (int): Number of samples per batch.
+            input_shape (tuple, optional): Shape of the input data if raw format is used.
+            expected_shape (tuple, optional): Shape of the expected data if raw format is used.
+            dtype (type, optional): Data type of the raw file (default: np.float32).
+        """
+        self.batch_size = batch_size
+        self.inputs = self._process_dataset(input_dataset, input_shape, dtype)
+        self.expecteds = self._process_dataset(expected_dataset, expected_shape, dtype)
+        self.batched_inputs = []
+
+        # Verify data consistency
+        self.num_samples = self.inputs[0].shape[0]  # Batch dimension
+        if self.num_samples != self.expecteds[0].shape[0]:
+            raise ValueError(
+                "Input data and expected data must have the same number of samples.")
+
+        # Precompute batches
+        self.batched_inputs, self.batched_expecteds = self._create_batches()
+
+    def _process_dataset(self, data, shape, dtype=np.float32):
+        """
+        Process a dataset or file path.
+
+        Args:
+            data (str or np.ndarray): Path to file or NumPy arrays.
+            shape (tuple, optional): Shape of the data if raw format is used.
+            dtype (type, optional): Data type for raw files.
+
+        Returns:
+            list of np.ndarray: Loaded or passed data as NumPy arrays.
+        """
+        if isinstance(data, list):
+            # Check if all elements in the list are NumPy arrays
+            if all(isinstance(item, np.ndarray) for item in data):
+                return data
+            raise ValueError("All elements in the list must be NumPy arrays.")
+        if isinstance(data, np.ndarray):
+            # If it's already a NumPy array and is not a list of arrays
+            if len(data.shape) > 1:
+                # If the array has multiple dimensions, split it into a list of arrays
+                return [data[i] for i in range(data.shape[0])]
+            else:
+                # If it's a single array, wrap it into a list
+                return [data]
+        elif isinstance(data, str):
+            # If it's a string, assume it's a file path
+            return [self._load_data(data, shape, dtype)]
+        else:
+            raise ValueError("Data must be a NumPy array or a valid file path.")
+
+    def _load_data(self, file_path, shape, dtype=np.float32):
+        """
+        Load data from a file, supporting both .npy and raw formats.
+
+        Args:
+            file_path (str): Path to the file to load.
+            shape (tuple, optional): Shape of the data if raw format is used.
+            dtype (type, optional): Data type of the raw file (default: np.float32).
+
+        Returns:
+            np.ndarray: Loaded data as a NumPy array.
+        """
+        _, ext = os.path.splitext(file_path)
+
+        if ext == ".npy":
+            # Load .npy file
+            return np.load(file_path)
+        elif ext in [".bin", ".raw"]:
+            # Load raw binary file
+            if shape is None:
+                raise ValueError(f"Shape must be provided for raw file: {file_path}")
+            return self._load_raw(file_path, shape, dtype)
+        else:
+            raise ValueError(f"Unsupported file format: {ext}")
+
+    def _load_raw(self, file_path, shape, dtype):
+        """
+        Load raw binary data.
+
+        Args:
+            file_path (str): Path to the raw binary file.
+            shape (tuple): Shape of the data to reshape into.
+            dtype (type): Data type of the binary file.
+
+        Returns:
+            np.ndarray: Loaded data as a NumPy array.
+        """
+        # Calculate the expected number of elements based on the provided shape
+        expected_elements = np.prod(shape)
+
+        # Calculate the expected size of the raw file in bytes
+        expected_size = expected_elements * np.dtype(dtype).itemsize
+
+        # Get the actual size of the raw file
+        actual_size = os.path.getsize(file_path)
+
+        # Check if the sizes match
+        if actual_size != expected_size:
+            raise ValueError(
+                f"Raw file size ({actual_size} bytes) does not match the expected size "
+                f"({expected_size} bytes) based on the provided shape {shape} and dtype {dtype}."
+            )
+
+        # Read and load the raw data
+        with open(file_path, "rb") as f:
+            data = f.read()
+            array = np.frombuffer(data, dtype=dtype)
+            if array.size != expected_elements:
+                raise ValueError(
+                    f"Raw data size does not match the expected shape: {shape}. "
+                    f"Expected {expected_elements} elements, got {array.size} elements.")
+            return array.reshape(shape)
+
+    def _create_batches(self):
+        """
+        Precompute batches for inputs and expected outputs.
+
+        Returns:
+            tuple: Lists of batched inputs and batched expecteds.
+        """
+        batched_inputs = []
+        batched_expecteds = []
+
+        for batch_start in range(0, self.num_samples, self.batch_size):
+            batch_end = min(batch_start + self.batch_size, self.num_samples)
+
+            # Collect batched inputs
+            inputs_batch = [
+                input_array[batch_start:batch_end] for input_array in self.inputs
+            ]
+            if batch_end - batch_start < self.batch_size:
+                # Resize the last batch to match batch_size
+                inputs_batch = [
+                    np.resize(batch, (self.batch_size, *batch.shape[1:]))
+                    for batch in inputs_batch
+                ]
+
+            batched_inputs.append(inputs_batch)
+
+            # Collect batched expecteds
+            expecteds_batch = [
+                expected_array[batch_start:batch_end] for expected_array in self.expecteds
+            ]
+            if batch_end - batch_start < self.batch_size:
+                # Resize the last batch to match batch_size
+                expecteds_batch = [
+                    np.resize(batch, (self.batch_size, *batch.shape[1:]))
+                    for batch in expecteds_batch
+                ]
+
+            batched_expecteds.append(expecteds_batch)
+
+        return batched_inputs, batched_expecteds
+
+    def __iter__(self):
+        """
+        Make the DataLoader iterable.
+
+        Returns:
+            self
+        """
+        self.index = 0
+        return self
+
+    def __next__(self):
+        """
+        Return the next batch of data.
+
+        Returns:
+            tuple: (inputs, expecteds) for the next batch.
+        """
+        if self.index >= len(self.batched_inputs):
+            raise StopIteration
+
+        # Retrieve precomputed batch
+        input_batch = self.batched_inputs[self.index]
+        expected_batch = self.batched_expecteds[self.index]
+
+        self.index += 1
+        return input_batch, expected_batch
+
+    def split(self, validation_split):
+        """
+        Split the data into training and validation sets.
+
+        Args:
+            validation_split (float): Ratio of validation data. Must be between 0.0 and 1.0.
+
+        Returns:
+            tuple: Two DataLoader instances, one for training and one for validation.
+        """
+        if not (0.0 <= validation_split <= 1.0):
+            raise ValueError("Validation split must be between 0.0 and 1.0.")
+
+        split_index = int(len(self.inputs[0]) * (1.0 - validation_split))
+
+        train_inputs = [input_array[:split_index] for input_array in self.inputs]
+        val_inputs = [input_array[split_index:] for input_array in self.inputs]
+        train_expecteds = [
+            expected_array[:split_index] for expected_array in self.expecteds
+        ]
+        val_expecteds = [
+            expected_array[split_index:] for expected_array in self.expecteds
+        ]
+
+        train_loader = DataLoader(train_inputs, train_expecteds, self.batch_size)
+        val_loader = DataLoader(val_inputs, val_expecteds, self.batch_size)
+
+        return train_loader, val_loader

runtime/onert/core/src/compiler/train/TrainingCompiler.cc

@@ -181,7 +181,7 @@ std::shared_ptr<CompilerArtifact> TrainingCompiler::compile(void)
       auto &input = trainable_subg->operands().at(ind);
       auto new_shape = input.info().shape();
       // TODO Consider batch size index
-      if (new_shape.dim(0) != 1)
+      if (new_shape.dim(0) != 1 && new_shape.dim(0) != ir::Shape::kUnspecifiedDim)
         throw std::runtime_error("the first dim is not 1. It is not supported yet.");
       new_shape.dim(0) = _training_info.batchSize();
       input.info().shape(new_shape);