Merge pull request #1089 from AndreaCossu/master

AvalancheDatasetType removal, dataset collate_fn used in dataloading during training and eval, sequential benchmark

Author: AntonioCarta

avalanche/benchmarks/classic/stream51.py

@@ -21,8 +21,6 @@
 import math
 import os
 
-from avalanche.benchmarks.utils import AvalancheDatasetType
-
 _mu = [0.485, 0.456, 0.406]
 _std = [0.229, 0.224, 0.225]
 _default_stream51_transform = transforms.Compose(
@@ -286,8 +284,7 @@ def CLStream51(
         task_labels=[0 for _ in range(num_tasks)],
         complete_test_set_only=scenario == "instance",
         train_transform=train_transform,
-        eval_transform=eval_transform,
-        dataset_type=AvalancheDatasetType.CLASSIFICATION,
+        eval_transform=eval_transform
     )
 
     return benchmark_obj

avalanche/benchmarks/datasets/torchaudio_wrapper.py

@@ -0,0 +1,91 @@
+################################################################################
+# Copyright (c) 2022 ContinualAI.                                              #
+# Copyrights licensed under the MIT License.                                   #
+# See the accompanying LICENSE file for terms.                                 #
+#                                                                              #
+# Author(s): Andrea Cossu                                                      #
+# E-mail: contact@continualai.org                                              #
+# Website: www.continualai.org                                                 #
+################################################################################
+
+""" This module conveniently wraps TorchAudio Datasets for using a clean and
+comprehensive Avalanche API."""
+
+try:
+    import torchaudio
+except ImportError:
+    raise ModuleNotFoundError(
+        "TorchAudio package is required to load its dataset. "
+        "You can install it as extra dependency with "
+        "`pip install avalanche-lib[extra]`")
+from torchaudio.datasets import SPEECHCOMMANDS
+from avalanche.benchmarks.utils import AvalancheDataset
+from avalanche.benchmarks.datasets import default_dataset_location
+import torch
+
+
+def speech_commands_collate(batch):
+    tensors, targets, t_labels = [], [], []
+    for waveform, label, rate, sid, uid, t_label in batch:
+        tensors += [waveform]
+        targets += [torch.tensor(label)]
+        t_labels += [torch.tensor(t_label)]
+    tensors = [item.t() for item in tensors]
+    tensors = torch.nn.utils.rnn.pad_sequence(tensors,
+                                              batch_first=True,
+                                              padding_value=0.)
+    if len(tensors.size()) == 2:  # no MFCC, add feature dimension
+        tensors = tensors.unsqueeze(-1)
+    targets = torch.stack(targets)
+    t_labels = torch.stack(t_labels)
+    return tensors, targets, t_labels
+
+
+class SpeechCommandsData(SPEECHCOMMANDS):
+    def __init__(self, root, url, download, subset, mfcc_preprocessing):
+        super().__init__(root=root, download=download,
+                         subset=subset, url=url)
+        self.labels_names = ['backward', 'bed', 'bird', 'cat', 'dog', 'down',
+                             'eight', 'five', 'follow', 'forward', 'four',
+                             'go', 'happy', 'house', 'learn', 'left',
+                             'marvin', 'nine', 'no', 'off', 'on', 'one',
+                             'right', 'seven', 'sheila', 'six', 'stop',
+                             'three', 'tree', 'two', 'up', 'visual',
+                             'wow', 'yes', 'zero']
+        self.mfcc_preprocessing = mfcc_preprocessing
+
+    def __getitem__(self, item):
+        wave, rate, label, speaker_id, ut_number = super().__getitem__(item)
+        label = self.labels_names.index(label)
+        wave = wave.squeeze(0)  # (T,)
+        if self.mfcc_preprocessing is not None:
+            assert rate == self.mfcc_preprocessing.sample_rate
+            # (T, MFCC)
+            wave = self.mfcc_preprocessing(wave).permute(1, 0)
+        return wave, label, rate, speaker_id, ut_number
+
+
+def SpeechCommands(root=default_dataset_location(''),
+                   url='speech_commands_v0.02',
+                   download=True, subset=None,
+                   mfcc_preprocessing=None):
+    """
+    root: dataset root location
+    url: version name of the dataset
+    download: automatically download the dataset, if not present
+    subset: one of 'training', 'validation', 'testing'
+    mfcc_preprocessing: an optional torchaudio.transforms.MFCC instance
+        to preprocess each audio. Warning: this may slow down the execution
+        since preprocessing is applied on-the-fly each time a sample is
+        retrieved from the dataset.
+    """
+    dataset = SpeechCommandsData(root=root, download=download,
+                                 subset=subset, url=url,
+                                 mfcc_preprocessing=mfcc_preprocessing)
+    labels = [datapoint[1] for datapoint in dataset]
+    return AvalancheDataset(dataset,
+                            collate_fn=speech_commands_collate,
+                            targets=labels)
+
+
+__all__ = ['SpeechCommands']

avalanche/benchmarks/generators/benchmark_generators.py

@@ -49,7 +49,6 @@
 from avalanche.benchmarks.utils.avalanche_dataset import (
     SupportedDataset,
     AvalancheDataset,
-    AvalancheDatasetType,
     AvalancheSubset,
 )
 
@@ -224,14 +223,14 @@ class "34" will be mapped to "1", class "11" to "2" and so on.
         train_dataset,
         transform_groups=transform_groups,
         initial_transform_group="train",
-        dataset_type=AvalancheDatasetType.CLASSIFICATION,
+        targets_adapter=int
     )
 
     test_dataset = AvalancheDataset(
         test_dataset,
         transform_groups=transform_groups,
         initial_transform_group="eval",
-        dataset_type=AvalancheDatasetType.CLASSIFICATION,
+        targets_adapter=int
     )
 
     return NCScenario(
@@ -348,14 +347,14 @@ def ni_benchmark(
         seq_train_dataset,
         transform_groups=transform_groups,
         initial_transform_group="train",
-        dataset_type=AvalancheDatasetType.CLASSIFICATION,
+        targets_adapter=int
     )
 
     seq_test_dataset = AvalancheDataset(
         seq_test_dataset,
         transform_groups=transform_groups,
         initial_transform_group="eval",
-        dataset_type=AvalancheDatasetType.CLASSIFICATION,
+        targets_adapter=int
     )
 
     return NIScenario(

avalanche/benchmarks/generators/scenario_generators.py

@@ -38,8 +38,7 @@
 from avalanche.benchmarks.utils import concat_datasets_sequentially
 from avalanche.benchmarks.utils.avalanche_dataset import (
     SupportedDataset,
-    as_classification_dataset,
-    AvalancheDatasetType,
+    as_classification_dataset
 )
 
 
@@ -325,8 +324,7 @@ def dataset_scenario(
     test_dataset_list: Sequence[SupportedDataset],
     task_labels: Sequence[int],
     *,
-    complete_test_set_only: bool = False,
-    dataset_type: AvalancheDatasetType = AvalancheDatasetType.UNDEFINED
+    complete_test_set_only: bool = False
 ) -> GenericCLScenario:
     """
     This helper function is DEPRECATED in favor of `dataset_benchmark`.
@@ -363,10 +361,6 @@ def dataset_scenario(
         parameter must be list with a single element (the complete test set).
         Defaults to False, which means that ``train_dataset_list`` and
         ``test_dataset_list`` must contain the same amount of datasets.
-    :param dataset_type: The type of the dataset. Defaults to None, which
-        means that the type will be obtained from the input datasets. If input
-        datasets are not instances of :class:`AvalancheDataset`, the type
-        UNDEFINED will be used.
 
     :returns: A properly initialized :class:`GenericCLScenario` instance.
     """
@@ -380,8 +374,7 @@ def dataset_scenario(
         train_dataset_list=train_dataset_list,
         test_dataset_list=test_dataset_list,
         task_labels=task_labels,
-        complete_test_set_only=complete_test_set_only,
-        dataset_type=dataset_type,
+        complete_test_set_only=complete_test_set_only
     )
 
 
@@ -482,8 +475,7 @@ def paths_scenario(
     train_transform=None,
     train_target_transform=None,
     eval_transform=None,
-    eval_target_transform=None,
-    dataset_type: AvalancheDatasetType = AvalancheDatasetType.UNDEFINED
+    eval_target_transform=None
 ) -> GenericCLScenario:
     """
     This helper function is DEPRECATED in favor of `paths_benchmark`.
@@ -545,7 +537,6 @@ def paths_scenario(
         comprehensive list of possible transformations). Defaults to None.
     :param eval_target_transform: The transformation to apply to test
         patterns targets. Defaults to None.
-    :param dataset_type: The type of the dataset. Defaults to UNDEFINED.
 
     :returns: A properly initialized :class:`GenericCLScenario` instance.
     """
@@ -563,8 +554,7 @@ def paths_scenario(
         train_transform=train_transform,
         train_target_transform=train_target_transform,
         eval_transform=eval_transform,
-        eval_target_transform=eval_target_transform,
-        dataset_type=dataset_type,
+        eval_target_transform=eval_target_transform
     )
 
 
@@ -577,8 +567,7 @@ def tensors_scenario(
     train_transform=None,
     train_target_transform=None,
     eval_transform=None,
-    eval_target_transform=None,
-    dataset_type: AvalancheDatasetType = AvalancheDatasetType.UNDEFINED
+    eval_target_transform=None
 ) -> GenericCLScenario:
     """
     This helper function is DEPRECATED in favor of `tensors_benchmark`.
@@ -635,7 +624,6 @@ def tensors_scenario(
         comprehensive list of possible transformations). Defaults to None.
     :param eval_target_transform: The transformation to apply to test
         patterns targets. Defaults to None.
-    :param dataset_type: The type of the dataset. Defaults to UNDEFINED.
 
     :returns: A properly initialized :class:`GenericCLScenario` instance.
     """
@@ -653,8 +641,7 @@ def tensors_scenario(
         train_transform=train_transform,
         train_target_transform=train_target_transform,
         eval_transform=eval_transform,
-        eval_target_transform=eval_target_transform,
-        dataset_type=dataset_type,
+        eval_target_transform=eval_target_transform
     )
 
 
@@ -669,8 +656,7 @@ def tensor_scenario(
     train_transform=None,
     train_target_transform=None,
     eval_transform=None,
-    eval_target_transform=None,
-    dataset_type: AvalancheDatasetType = AvalancheDatasetType.UNDEFINED
+    eval_target_transform=None
 ) -> GenericCLScenario:
     """
     This helper function is DEPRECATED in favor of `tensors_benchmark`.
@@ -731,7 +717,6 @@ def tensor_scenario(
         comprehensive list of possible transformations). Defaults to None.
     :param eval_target_transform: The transformation to apply to test
         patterns targets. Defaults to None.
-    :param dataset_type: The type of the dataset. Defaults to UNDEFINED.
 
     :returns: A properly initialized :class:`GenericCLScenario` instance.
     """
@@ -782,8 +767,7 @@ def tensor_scenario(
         train_transform=train_transform,
         train_target_transform=train_target_transform,
         eval_transform=eval_transform,
-        eval_target_transform=eval_target_transform,
-        dataset_type=dataset_type,
+        eval_target_transform=eval_target_transform
     )