Fix _init_hubert_pretrain_model (#2886)

Summary:
address https://github.com/pytorch/audio/issues/2885

In `_init_hubert_pretrain_model ` method which initialize the hubert pretrain models, `kaiming_normal_` should be applied on `ConvLayerBlock` instead of `LayerNorm` layer. This PR fixes it and adds more unit tests.

Pull Request resolved: https://github.com/pytorch/audio/pull/2886

Reviewed By: hwangjeff

Differential Revision: D41713801

Pulled By: nateanl

fbshipit-source-id: ed199baf7504d06bbf2d31c522ae708a75426a2d

test/torchaudio_unittest/models/wav2vec2/model_test.py

+import math
 import os
 from typing import Tuple
 
@@ -7,6 +8,9 @@ from parameterized import parameterized
 from torchaudio.models.wav2vec2 import (
     hubert_base,
     hubert_large,
+    hubert_pretrain_base,
+    hubert_pretrain_large,
+    hubert_pretrain_xlarge,
     hubert_xlarge,
     wav2vec2_base,
     wav2vec2_large,
@@ -37,6 +41,15 @@ factory_funcs = parameterized.expand(
     name_func=_name_func,
 )
 
+factory_funcs_hubert_pretrain = parameterized.expand(
+    [
+        (hubert_pretrain_base,),
+        (hubert_pretrain_large,),
+        (hubert_pretrain_xlarge,),
+    ],
+    name_func=_name_func,
+)
+
 
 class TestWav2Vec2Model(TorchaudioTestCase):
     def _smoke_test(self, model, device, dtype):
@@ -263,3 +276,127 @@ class TestWav2Vec2Model(TorchaudioTestCase):
     def test_quantize_torchscript(self, factory_func):
         """Quantized Wav2Vec2Model should be scriptable"""
         self._test_quantize_torchscript(factory_func(aux_num_out=32))
+
+
+def _compute_label_frame(audio_frame: int) -> int:
+    """Compute number of frames in the label tensor based on
+    the number of frames in the audio tensor."""
+    kernel_size = 25
+    stride = 20
+    sample_rate = 16  # 16 per millisecond
+    label_frame = math.floor((audio_frame - kernel_size * sample_rate) / (stride * sample_rate)) + 1
+    return label_frame
+
+
+class TestHuBERTPretrainModel(TorchaudioTestCase):
+    def _smoke_test(self, model, device, dtype):
+        model = model.to(device=device, dtype=dtype)
+        model = model.eval()
+
+        batch_size, num_frames = 3, 1024
+
+        waveforms = torch.randn(batch_size, num_frames, device=device, dtype=dtype)
+        labels = torch.randint(
+            low=0,
+            high=100,
+            size=[
+                batch_size,
+                _compute_label_frame(num_frames),
+            ],
+            device=device,
+        )
+        lengths = torch.randint(
+            low=0,
+            high=num_frames,
+            size=[
+                batch_size,
+            ],
+            device=device,
+        )
+
+        model(waveforms, labels, lengths)
+
+    @parameterized.expand([(torch.float32,), (torch.float64,)])
+    def test_cpu_smoke_test(self, dtype):
+        model = hubert_pretrain_base()
+        self._smoke_test(model, torch.device("cpu"), dtype)
+
+    @parameterized.expand([(torch.float32,), (torch.float64,)])
+    @skipIfNoCuda
+    def test_cuda_smoke_test(self, dtype):
+        model = hubert_pretrain_base()
+        self._smoke_test(model, torch.device("cuda"), dtype)
+
+    def _feature_extractor_test(self, model):
+        batch_size, num_frames = 3, 1024
+
+        model = model.wav2vec2
+        model.eval()
+        num_layers = len(model.encoder.transformer.layers)
+
+        waveforms = torch.randn(batch_size, num_frames)
+        lengths = torch.randint(
+            low=0,
+            high=num_frames,
+            size=[
+                batch_size,
+            ],
+        )
+
+        # Not providing num_layers returns all the intermediate features from
+        # tranformer layers
+        all_features, lengths_ = model.extract_features(waveforms, lengths, num_layers=None)
+        assert len(all_features) == num_layers
+        for features in all_features:
+            assert features.ndim == 3
+            assert features.shape[0] == batch_size
+        assert lengths_.shape == torch.Size([batch_size])
+
+        # Limiting the number of layers to `l`.
+        for l in range(1, num_layers + 1):
+            features, lengths_ = model.extract_features(waveforms, lengths, num_layers=l)
+            assert len(features) == l
+            for i in range(l):
+                self.assertEqual(all_features[i], features[i])
+            assert lengths_.shape == torch.Size([batch_size])
+
+    @factory_funcs_hubert_pretrain
+    def test_extract_feature(self, factory_func):
+        """`extract_features` method does not fail"""
+        self._feature_extractor_test(factory_func())
+
+    def _test_quantize_smoke_test(self, model):
+        model.eval()
+        batch_size, num_frames = 3, 1024
+
+        # Remove the weight normalization forward hook
+        model.wav2vec2.encoder.transformer.pos_conv_embed.__prepare_scriptable__()
+        quantized = tq.quantize_dynamic(model, qconfig_spec={torch.nn.Linear}, dtype=torch.qint8)
+
+        # A lazy way to check that Modules are different
+        assert str(quantized) != str(model), "Dynamic quantization did not modify the module."
+
+        waveforms = torch.randn(batch_size, num_frames)
+        labels = torch.randint(
+            low=0,
+            high=100,
+            size=[
+                batch_size,
+                _compute_label_frame(num_frames),
+            ],
+        )
+        lengths = torch.randint(
+            low=0,
+            high=num_frames,
+            size=[
+                batch_size,
+            ],
+        )
+
+        _, _, _ = quantized(waveforms, labels, lengths)
+
+    @factory_funcs_hubert_pretrain
+    @skipIfNoQengine
+    def test_quantize(self, factory_func):
+        """HuBERTPretrainModel should support basic quantization"""
+        self._test_quantize_smoke_test(factory_func())

torchaudio/models/wav2vec2/model.py

@@ -683,7 +683,7 @@ def hubert_xlarge(
 
 
 def _init_hubert_pretrain_model(module):
-    if isinstance(module, components.LayerNorm):
+    if isinstance(module, components.ConvLayerBlock):
         torch.nn.init.kaiming_normal_(module.conv.weight)
     elif isinstance(module, components.ConvolutionalPositionalEmbedding):
         # normalize the weight to normal distribution.