Add iemocap variants (#2778)

Summary:
add ability to load only improvised or only scripted utterances.

Pull Request resolved: https://github.com/pytorch/audio/pull/2778

Reviewed By: nateanl

Differential Revision: D40511865

Pulled By: carolineechen

fbshipit-source-id: e1fe3908ac2aa306ad30c242ddd25762b2268539

test/torchaudio_unittest/datasets/iemocap_test.py

@@ -4,7 +4,7 @@ import random
 from torchaudio.datasets import iemocap
 from torchaudio_unittest.common_utils import get_whitenoise, save_wav, TempDirMixin, TorchaudioTestCase
 
-LABELS = ["neu", "hap", "ang", "sad", "exc", "xxx"]
+LABELS = ["neu", "hap", "ang", "sad", "exc", "fru", "xxx"]
 SAMPLE_RATE = 16000
 
 
@@ -21,8 +21,6 @@ def _save_wav(filepath: str, seed: int):
 
 def _save_label(label_folder: str, filename: str, wav_stem: str):
     label = random.choice(LABELS)
-    if label == "exc":
-        label = "hap"
     line = f"[xxx]\t{wav_stem}\t{label}\t[yyy]"
     filepath = os.path.join(label_folder, filename)
 
@@ -40,12 +38,12 @@ def _get_samples(dataset_dir: str, session: int):
     os.makedirs(wav_folder, exist_ok=True)
     os.makedirs(label_folder, exist_ok=True)
 
-    samples = []
     wav_stems = []
     for i in range(5):
         for g in ["F", "M"]:
+            for utt in ["impro", "script"]:
                 speaker = f"Ses0{session}{g}"
-            subfolder = f"{speaker}_impro0{i}"
+                subfolder = f"{speaker}_{utt}0{i}"
                 subfolder_path = os.path.join(wav_folder, subfolder)
                 os.makedirs(subfolder_path, exist_ok=True)
 
@@ -53,6 +51,9 @@ def _get_samples(dataset_dir: str, session: int):
                     wav_stem = f"{subfolder}_F00{j}"
                     wav_stems.append(wav_stem)
 
+    all_samples = []
+    impro_samples = []
+    script_samples = []
     wav_stems = sorted(wav_stems)
     for wav_stem in wav_stems:
         subfolder = wav_stem[:-5]
@@ -64,31 +65,43 @@ def _get_samples(dataset_dir: str, session: int):
         if label == "xxx":
             continue
         sample = (wav, SAMPLE_RATE, wav_stem, label, speaker)
-        samples.append(sample)
+        all_samples.append(sample)
 
-    return samples
+        if "impro" in subfolder:
+            impro_samples.append(sample)
+        else:
+            script_samples.append(sample)
+
+    return all_samples, script_samples, impro_samples
 
 
 def get_mock_dataset(dataset_dir: str):
     os.makedirs(dataset_dir, exist_ok=True)
 
-    samples = []
+    all_samples = []
+    script_samples = []
+    impro_samples = []
     for session in range(1, 4):
-        samples += _get_samples(dataset_dir, session)
-    return samples
+        samples = _get_samples(dataset_dir, session)
+        all_samples += samples[0]
+        script_samples += samples[1]
+        impro_samples += samples[2]
+    return all_samples, script_samples, impro_samples
 
 
 class TestIemocap(TempDirMixin, TorchaudioTestCase):
     root_dir = None
     backend = "default"
 
-    samples = []
+    all_samples = []
+    script_samples = []
+    impro_samples = []
 
     @classmethod
     def setUpClass(cls):
         cls.root_dir = cls.get_base_temp_dir()
         dataset_dir = os.path.join(cls.root_dir, "IEMOCAP")
-        cls.samples = get_mock_dataset(dataset_dir)
+        cls.all_samples, cls.script_samples, cls.impro_samples = get_mock_dataset(dataset_dir)
 
     def _testIEMOCAP(self, dataset, samples):
         num_samples = 0
@@ -98,6 +111,14 @@ class TestIemocap(TempDirMixin, TorchaudioTestCase):
 
         assert num_samples == len(samples)
 
-    def testIEMOCAPDataset(self):
+    def testIEMOCAPFullDataset(self):
         dataset = iemocap.IEMOCAP(self.root_dir)
-        self._testIEMOCAP(dataset, self.samples)
+        self._testIEMOCAP(dataset, self.all_samples)
+
+    def testIEMOCAPScriptedDataset(self):
+        dataset = iemocap.IEMOCAP(self.root_dir, utterance_type="scripted")
+        self._testIEMOCAP(dataset, self.script_samples)
+
+    def testIEMOCAPImprovisedDataset(self):
+        dataset = iemocap.IEMOCAP(self.root_dir, utterance_type="improvised")
+        self._testIEMOCAP(dataset, self.impro_samples)

torchaudio/datasets/iemocap.py

 import os
 import re
 from pathlib import Path
-from typing import Tuple, Union
+from typing import Optional, Tuple, Union
 
 from torch import Tensor
 from torch.utils.data import Dataset
@@ -28,12 +28,16 @@ class IEMOCAP(Dataset):
     Args:
         root (str or Path): Root directory where the dataset's top level directory is found
         sessions (Tuple[int]): Tuple of sessions (1-5) to use. (Default: ``(1, 2, 3, 4, 5)``)
+        utterance_type (str or None, optional): Which type(s) of utterances to include in the dataset.
+            Options: ("scripted", "improvised", ``None``). If ``None``, both scripted and improvised
+            data are used.
     """
 
     def __init__(
         self,
         root: Union[str, Path],
         sessions: Tuple[str] = (1, 2, 3, 4, 5),
+        utterance_type: Optional[str] = None,
     ):
         root = Path(root)
         self._path = root / "IEMOCAP"
@@ -41,6 +45,9 @@ class IEMOCAP(Dataset):
         if not os.path.isdir(self._path):
             raise RuntimeError("Dataset not found.")
 
+        if utterance_type not in ["scripted", "improvised", None]:
+            raise ValueError("utterance_type must be one of ['scripted', 'improvised', or None]")
+
         all_data = []
         self.data = []
         self.mapping = {}
@@ -57,7 +64,12 @@ class IEMOCAP(Dataset):
 
             # add labels
             label_dir = session_dir / "dialog" / "EmoEvaluation"
-            label_paths = label_dir.glob("*.txt")
+            query = "*.txt"
+            if utterance_type == "scripted":
+                query = "*script*.txt"
+            elif utterance_type == "improvised":
+                query = "*impro*.txt"
+            label_paths = label_dir.glob(query)
 
             for label_path in label_paths:
                 with open(label_path, "r") as f:
@@ -67,11 +79,9 @@ class IEMOCAP(Dataset):
                         line = re.split("[\t\n]", line)
                         wav_stem = line[1]
                         label = line[2]
-                        if label == "exc":
-                            label = "hap"
                         if wav_stem not in all_data:
                             continue
-                        if label not in ["neu", "hap", "ang", "sad"]:
+                        if label not in ["neu", "hap", "ang", "sad", "exc", "fru"]:
                             continue
                         self.mapping[wav_stem] = {}
                         self.mapping[wav_stem]["label"] = label
@@ -99,7 +109,7 @@ class IEMOCAP(Dataset):
             str:
                 File name
             str:
-                Label (one of ``"neu"``, ``"hap"``, ``"ang"``, ``"sad"``)
+                Label (one of ``"neu"``, ``"hap"``, ``"ang"``, ``"sad"``, ``"exc"``, ``"fru"``)
             str:
                 Speaker
         """
@@ -125,7 +135,7 @@ class IEMOCAP(Dataset):
             str:
                 File name
             str:
-                Label (one of ``"neu"``, ``"hap"``, ``"ang"``, ``"sad"``)
+                Label (one of ``"neu"``, ``"hap"``, ``"ang"``, ``"sad"``, ``"exc"``, ``"fru"``)
             str:
                 Speaker
         """