Merge pull request #777 from PaddlePaddle/feat

feat (D,T) to (T, D); time warp

deepspeech/frontend/augmentor/spec_augment.py

@@ -12,7 +12,11 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """Contains the volume perturb augmentation model."""
+import random
+
 import numpy as np
+from PIL import Image
+from PIL.Image import BICUBIC
 
 from deepspeech.frontend.augmentor.base import AugmentorBase
 from deepspeech.utils.log import Log
@@ -42,7 +46,8 @@ class SpecAugmentor(AugmentorBase):
                  adaptive_number_ratio=0,
                  adaptive_size_ratio=0,
                  max_n_time_masks=20,
-                 replace_with_zero=True):
+                 replace_with_zero=True,
+                 warp_mode='PIL'):
         """SpecAugment class.
         Args:
             rng (random.Random): random generator object.
@@ -56,11 +61,15 @@ class SpecAugmentor(AugmentorBase):
             adaptive_size_ratio (float): adaptive size ratio for time masking
             max_n_time_masks (int): maximum number of time masking
             replace_with_zero (bool): pad zero on mask if true else use mean
+            warp_mode (str):  "PIL" (default, fast, not differentiable) 
+                 or "sparse_image_warp" (slow, differentiable)
         """
         super().__init__()
         self._rng = rng
+        self.inplace = True
         self.replace_with_zero = replace_with_zero
 
+        self.mode = warp_mode
         self.W = W
         self.F = F
         self.T = T
@@ -126,24 +135,80 @@ class SpecAugmentor(AugmentorBase):
     def __repr__(self):
         return f"specaug: F-{F}, T-{T}, F-n-{n_freq_masks}, T-n-{n_time_masks}"
 
-    def time_warp(xs, W=40):
-        return xs
+    def time_warp(self, x, mode='PIL'):
+        """time warp for spec augment
+        move random center frame by the random width ~ uniform(-window, window)
+
+        Args:
+            x (np.ndarray): spectrogram (time, freq)
+            mode (str): PIL or sparse_image_warp
+
+        Raises:
+            NotImplementedError: [description]
+            NotImplementedError: [description]
+
+        Returns:
+            np.ndarray: time warped spectrogram (time, freq)
+        """
+        window = max_time_warp = self.W
+        if mode == "PIL":
+            t = x.shape[0]
+            if t - window <= window:
+                return x
+            # NOTE: randrange(a, b) emits a, a + 1, ..., b - 1
+            center = random.randrange(window, t - window)
+            warped = random.randrange(center - window, center +
+                                      window) + 1  # 1 ... t - 1
+
+            left = Image.fromarray(x[:center]).resize((x.shape[1], warped),
+                                                      BICUBIC)
+            right = Image.fromarray(x[center:]).resize((x.shape[1], t - warped),
+                                                       BICUBIC)
+            if self.inplace:
+                x[:warped] = left
+                x[warped:] = right
+                return x
+            return np.concatenate((left, right), 0)
+        elif mode == "sparse_image_warp":
+            raise NotImplementedError('sparse_image_warp')
+        else:
+            raise NotImplementedError(
+                "unknown resize mode: " + mode +
+                ", choose one from (PIL, sparse_image_warp).")
+
+    def mask_freq(self, x, replace_with_zero=False):
+        """freq mask
 
-    def mask_freq(self, xs, replace_with_zero=False):
-        n_bins = xs.shape[0]
+        Args:
+            x (np.ndarray): spectrogram (time, freq)
+            replace_with_zero (bool, optional): Defaults to False.
+
+        Returns:
+            np.ndarray: freq mask spectrogram (time, freq)
+        """
+        n_bins = x.shape[1]
         for i in range(0, self.n_freq_masks):
             f = int(self._rng.uniform(low=0, high=self.F))
             f_0 = int(self._rng.uniform(low=0, high=n_bins - f))
             assert f_0 <= f_0 + f
-            if self.replace_with_zero:
-                xs[f_0:f_0 + f, :] = 0
+            if replace_with_zero:
+                x[:, f_0:f_0 + f] = 0
             else:
-                xs[f_0:f_0 + f, :] = xs.mean()
+                x[:, f_0:f_0 + f] = x.mean()
             self._freq_mask = (f_0, f_0 + f)
-        return xs
+        return x
+
+    def mask_time(self, x, replace_with_zero=False):
+        """time mask
 
-    def mask_time(self, xs, replace_with_zero=False):
-        n_frames = xs.shape[1]
+        Args:
+            x (np.ndarray): spectrogram (time, freq)
+            replace_with_zero (bool, optional): Defaults to False.
+
+        Returns:
+            np.ndarray: time mask spectrogram (time, freq)
+        """
+        n_frames = x.shape[0]
 
         if self.adaptive_number_ratio > 0:
             n_masks = int(n_frames * self.adaptive_number_ratio)
@@ -161,26 +226,26 @@ class SpecAugmentor(AugmentorBase):
             t = min(t, int(n_frames * self.p))
             t_0 = int(self._rng.uniform(low=0, high=n_frames - t))
             assert t_0 <= t_0 + t
-            if self.replace_with_zero:
-                xs[:, t_0:t_0 + t] = 0
+            if replace_with_zero:
+                x[t_0:t_0 + t, :] = 0
             else:
-                xs[:, t_0:t_0 + t] = xs.mean()
+                x[t_0:t_0 + t, :] = x.mean()
             self._time_mask = (t_0, t_0 + t)
-        return xs
+        return x
 
     def __call__(self, x, train=True):
         if not train:
             return x
         return self.transform_feature(x)
 
-    def transform_feature(self, xs: np.ndarray):
+    def transform_feature(self, x: np.ndarray):
         """
         Args:
-            xs (FloatTensor): `[F, T]`
+            x (np.ndarray): `[T, F]`
         Returns:
-            xs (FloatTensor): `[F, T]`
+            x (np.ndarray): `[T, F]`
         """
-        xs = self.time_warp(xs)
-        xs = self.mask_freq(xs)
-        xs = self.mask_time(xs)
-        return xs
+        x = self.time_warp(x, self.mode)
+        x = self.mask_freq(x, self.replace_with_zero)
+        x = self.mask_time(x, self.replace_with_zero)
+        return x

deepspeech/frontend/featurizer/audio_featurizer.py

@@ -167,32 +167,6 @@ class AudioFeaturizer(object):
             raise ValueError("Unknown specgram_type %s. "
                              "Supported values: linear." % self._specgram_type)
 
-    def _compute_linear_specgram(self,
-                                 samples,
-                                 sample_rate,
-                                 stride_ms=10.0,
-                                 window_ms=20.0,
-                                 max_freq=None,
-                                 eps=1e-14):
-        """Compute the linear spectrogram from FFT energy."""
-        if max_freq is None:
-            max_freq = sample_rate / 2
-        if max_freq > sample_rate / 2:
-            raise ValueError("max_freq must not be greater than half of "
-                             "sample rate.")
-        if stride_ms > window_ms:
-            raise ValueError("Stride size must not be greater than "
-                             "window size.")
-        stride_size = int(0.001 * sample_rate * stride_ms)
-        window_size = int(0.001 * sample_rate * window_ms)
-        specgram, freqs = self._specgram_real(
-            samples,
-            window_size=window_size,
-            stride_size=stride_size,
-            sample_rate=sample_rate)
-        ind = np.where(freqs <= max_freq)[0][-1] + 1
-        return np.log(specgram[:ind, :] + eps)
-
     def _specgram_real(self, samples, window_size, stride_size, sample_rate):
         """Compute the spectrogram for samples from a real signal."""
         # extract strided windows
@@ -217,26 +191,65 @@ class AudioFeaturizer(object):
         freqs = float(sample_rate) / window_size * np.arange(fft.shape[0])
         return fft, freqs
 
+    def _compute_linear_specgram(self,
+                                 samples,
+                                 sample_rate,
+                                 stride_ms=10.0,
+                                 window_ms=20.0,
+                                 max_freq=None,
+                                 eps=1e-14):
+        """Compute the linear spectrogram from FFT energy.
+
+        Args:
+            samples ([type]): [description]
+            sample_rate ([type]): [description]
+            stride_ms (float, optional): [description]. Defaults to 10.0.
+            window_ms (float, optional): [description]. Defaults to 20.0.
+            max_freq ([type], optional): [description]. Defaults to None.
+            eps ([type], optional): [description]. Defaults to 1e-14.
+
+        Raises:
+            ValueError: [description]
+            ValueError: [description]
+
+        Returns:
+            np.ndarray: log spectrogram, (time, freq)
+        """
+        if max_freq is None:
+            max_freq = sample_rate / 2
+        if max_freq > sample_rate / 2:
+            raise ValueError("max_freq must not be greater than half of "
+                             "sample rate.")
+        if stride_ms > window_ms:
+            raise ValueError("Stride size must not be greater than "
+                             "window size.")
+        stride_size = int(0.001 * sample_rate * stride_ms)
+        window_size = int(0.001 * sample_rate * window_ms)
+        specgram, freqs = self._specgram_real(
+            samples,
+            window_size=window_size,
+            stride_size=stride_size,
+            sample_rate=sample_rate)
+        ind = np.where(freqs <= max_freq)[0][-1] + 1
+        # (freq, time)
+        spec = np.log(specgram[:ind, :] + eps)
+        return np.transpose(spec)
+
     def _concat_delta_delta(self, feat):
         """append delat, delta-delta feature.
 
         Args:
-            feat (np.ndarray): (D, T)
+            feat (np.ndarray): (T, D)
 
         Returns:
-            np.ndarray: feat with delta-delta, (3*D, T)
+            np.ndarray: feat with delta-delta, (T, 3*D)
         """
-        feat = np.transpose(feat)
         # Deltas
         d_feat = delta(feat, 2)
         # Deltas-Deltas
         dd_feat = delta(feat, 2)
-        # transpose
-        feat = np.transpose(feat)
-        d_feat = np.transpose(d_feat)
-        dd_feat = np.transpose(dd_feat)
         # concat above three features
-        concat_feat = np.concatenate((feat, d_feat, dd_feat))
+        concat_feat = np.concatenate((feat, d_feat, dd_feat), axis=1)
         return concat_feat
 
     def _compute_mfcc(self,
@@ -292,7 +305,6 @@ class AudioFeaturizer(object):
             ceplifter=22,
             useEnergy=True,
             winfunc='povey')
-        mfcc_feat = np.transpose(mfcc_feat)
         if delta_delta:
             mfcc_feat = self._concat_delta_delta(mfcc_feat)
         return mfcc_feat
@@ -346,8 +358,6 @@ class AudioFeaturizer(object):
             remove_dc_offset=True,
             preemph=0.97,
             wintype='povey')
-
-        fbank_feat = np.transpose(fbank_feat)
         if delta_delta:
             fbank_feat = self._concat_delta_delta(fbank_feat)
         return fbank_feat

deepspeech/frontend/normalizer.py

@@ -40,21 +40,21 @@ class CollateFunc(object):
         number = 0
         for item in batch:
             audioseg = AudioSegment.from_file(item['feat'])
-            feat = self.feature_func(audioseg)  #(D, T)
+            feat = self.feature_func(audioseg)  #(T, D)
 
-            sums = np.sum(feat, axis=1)
+            sums = np.sum(feat, axis=0)
             if mean_stat is None:
                 mean_stat = sums
             else:
                 mean_stat += sums
 
-            square_sums = np.sum(np.square(feat), axis=1)
+            square_sums = np.sum(np.square(feat), axis=0)
             if var_stat is None:
                 var_stat = square_sums
             else:
                 var_stat += square_sums
 
-            number += feat.shape[1]
+            number += feat.shape[0]
         return number, mean_stat, var_stat
 
 
@@ -120,7 +120,7 @@ class FeatureNormalizer(object):
         """Normalize features to be of zero mean and unit stddev.
 
         :param features: Input features to be normalized.
-        :type features: ndarray, shape (D, T)
+        :type features: ndarray, shape (T, D)
         :param eps:  added to stddev to provide numerical stablibity.
         :type eps: float
         :return: Normalized features.
@@ -131,8 +131,8 @@ class FeatureNormalizer(object):
     def _read_mean_std_from_file(self, filepath, eps=1e-20):
         """Load mean and std from file."""
         mean, istd = load_cmvn(filepath, filetype='json')
-        self._mean = np.expand_dims(mean, axis=-1)
-        self._istd = np.expand_dims(istd, axis=-1)
+        self._mean = np.expand_dims(mean, axis=0)
+        self._istd = np.expand_dims(istd, axis=0)
 
     def write_to_file(self, filepath):
         """Write the mean and stddev to the file.

deepspeech/io/collator.py

@@ -242,7 +242,6 @@ class SpeechCollator():
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
-        specgram = specgram.transpose([1, 0])
         return specgram, transcript_part
 
     def __call__(self, batch):
@@ -250,7 +249,7 @@ class SpeechCollator():
 
         Args:
             batch ([List]): batch is (audio, text)
-                audio (np.ndarray) shape (D, T)
+                audio (np.ndarray) shape (T, D)
                 text (List[int] or str): shape (U,)
 
         Returns:

deepspeech/io/collator_st.py

@@ -217,6 +217,34 @@ class SpeechCollator():
         return self._local_data.tar2object[tarpath].extractfile(
             self._local_data.tar2info[tarpath][filename])
 
+    @property
+    def manifest(self):
+        return self._manifest
+
+    @property
+    def vocab_size(self):
+        return self._speech_featurizer.vocab_size
+
+    @property
+    def vocab_list(self):
+        return self._speech_featurizer.vocab_list
+
+    @property
+    def vocab_dict(self):
+        return self._speech_featurizer.vocab_dict
+
+    @property
+    def text_feature(self):
+        return self._speech_featurizer.text_feature
+
+    @property
+    def feature_size(self):
+        return self._speech_featurizer.feature_size
+
+    @property
+    def stride_ms(self):
+        return self._speech_featurizer.stride_ms
+
     def process_utterance(self, audio_file, translation):
         """Load, augment, featurize and normalize for speech data.
 
@@ -244,7 +272,6 @@ class SpeechCollator():
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
-        specgram = specgram.transpose([1, 0])
         return specgram, translation_part
 
     def __call__(self, batch):
@@ -252,7 +279,7 @@ class SpeechCollator():
 
         Args:
             batch ([List]): batch is (audio, text)
-                audio (np.ndarray) shape (D, T)
+                audio (np.ndarray) shape (T, D)
                 text (List[int] or str): shape (U,)
 
         Returns:
@@ -296,34 +323,6 @@ class SpeechCollator():
         text_lens = np.array(text_lens).astype(np.int64)
         return utts, padded_audios, audio_lens, padded_texts, text_lens
 
-    @property
-    def manifest(self):
-        return self._manifest
-
-    @property
-    def vocab_size(self):
-        return self._speech_featurizer.vocab_size
-
-    @property
-    def vocab_list(self):
-        return self._speech_featurizer.vocab_list
-
-    @property
-    def vocab_dict(self):
-        return self._speech_featurizer.vocab_dict
-
-    @property
-    def text_feature(self):
-        return self._speech_featurizer.text_feature
-
-    @property
-    def feature_size(self):
-        return self._speech_featurizer.feature_size
-
-    @property
-    def stride_ms(self):
-        return self._speech_featurizer.stride_ms
-
 
 class TripletSpeechCollator(SpeechCollator):
     def process_utterance(self, audio_file, translation, transcript):
@@ -355,7 +354,6 @@ class TripletSpeechCollator(SpeechCollator):
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
-        specgram = specgram.transpose([1, 0])
         return specgram, translation_part, transcript_part
 
     def __call__(self, batch):
@@ -363,7 +361,7 @@ class TripletSpeechCollator(SpeechCollator):
 
         Args:
             batch ([List]): batch is (audio, text)
-                audio (np.ndarray) shape (D, T)
+                audio (np.ndarray) shape (T, D)
                 text (List[int] or str): shape (U,)
 
         Returns:
@@ -524,49 +522,19 @@ class KaldiPrePorocessedCollator(SpeechCollator):
         :rtype: tuple of (2darray, list)
         """
         specgram = kaldiio.load_mat(audio_file)
-        specgram = specgram.transpose([1, 0])
         assert specgram.shape[
-            0] == self._feat_dim, 'expect feat dim {}, but got {}'.format(
-                self._feat_dim, specgram.shape[0])
+            1] == self._feat_dim, 'expect feat dim {}, but got {}'.format(
+                self._feat_dim, specgram.shape[1])
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
 
-        specgram = specgram.transpose([1, 0])
         if self._keep_transcription_text:
             return specgram, translation
         else:
             text_ids = self._text_featurizer.featurize(translation)
             return specgram, text_ids
 
-    @property
-    def manifest(self):
-        return self._manifest
-
-    @property
-    def vocab_size(self):
-        return self._text_featurizer.vocab_size
-
-    @property
-    def vocab_list(self):
-        return self._text_featurizer.vocab_list
-
-    @property
-    def vocab_dict(self):
-        return self._text_featurizer.vocab_dict
-
-    @property
-    def text_feature(self):
-        return self._text_featurizer
-
-    @property
-    def feature_size(self):
-        return self._feat_dim
-
-    @property
-    def stride_ms(self):
-        return self._stride_ms
-
 
 class TripletKaldiPrePorocessedCollator(KaldiPrePorocessedCollator):
     def process_utterance(self, audio_file, translation, transcript):
@@ -583,15 +551,13 @@ class TripletKaldiPrePorocessedCollator(KaldiPrePorocessedCollator):
         :rtype: tuple of (2darray, (list, list))
         """
         specgram = kaldiio.load_mat(audio_file)
-        specgram = specgram.transpose([1, 0])
         assert specgram.shape[
-            0] == self._feat_dim, 'expect feat dim {}, but got {}'.format(
-                self._feat_dim, specgram.shape[0])
+            1] == self._feat_dim, 'expect feat dim {}, but got {}'.format(
+                self._feat_dim, specgram.shape[1])
 
         # specgram augment
         specgram = self._augmentation_pipeline.transform_feature(specgram)
 
-        specgram = specgram.transpose([1, 0])
         if self._keep_transcription_text:
             return specgram, translation, transcript
         else:
@@ -604,7 +570,7 @@ class TripletKaldiPrePorocessedCollator(KaldiPrePorocessedCollator):
 
         Args:
             batch ([List]): batch is (audio, text)
-                audio (np.ndarray) shape (D, T)
+                audio (np.ndarray) shape (T, D)
                 translation (List[int] or str): shape (U,)
                 transcription (List[int] or str): shape (V,)
 

requirements.txt

@@ -1,6 +1,7 @@
 coverage
 gpustat
 kaldiio
+Pillow
 pre-commit
 pybind11
 resampy==0.2.2