cleaning

features_server.py

@@ -482,13 +482,12 @@ class FeaturesServer(object):
                                                                  label=label,
                                                                  start=start, stop=stop,
                                                                  global_cmvn=self.global_cmvn)
-
         # Post-process the features and return the features and vad label
         if global_cmvn:
             feat, label = self.post_processing(feat, label, global_mean, global_std)
         else:
             feat, label = self.post_processing(feat, label)
-        if mask is not None:
+        if self.mask is not None:
             feat = feat[:, self.mask]
 
         return feat, label

frontend/io.py

@@ -586,7 +586,6 @@ def read_hdf5_segment(file_handler,
         stop = dataset_length
     pad_end = stop - dataset_length if stop > dataset_length else 0
     stop = min(stop, dataset_length)
-
     global_cmvn = global_cmvn and not (start is None or stop is None)
 
     # Get the data between start and stop
@@ -601,7 +600,10 @@ def read_hdf5_segment(file_handler,
             if "/".join((show, data_id)) in h5f:
                 dataset_id = show + '/{}'.format(data_id)
                 if compression == 'none':
-                    feat.append(_read_segment(h5f, dataset_id, start, stop))
+                    data = _read_segment(h5f, dataset_id, start, stop)
+                    if data.ndim ==1:
+                        data = data[:, numpy.newaxis]
+                    feat.append(data)
                 elif compression == 'htk':
                     feat.append(_read_segment_htk(h5f, dataset_id, start, stop))
                 else:
@@ -995,8 +997,9 @@ def _read_dataset(h5f, dataset_id):
         data = data[:, numpy.newaxis]
     return data
 
-def _read_segment(h5f, dataset_id, e, s):
-    return h5f[dataset_id][s:e, :]
+def _read_segment(h5f, dataset_id, s, e):
+    data = h5f[dataset_id][s:e]
+    return data
 
 def _read_dataset_htk(h5f, dataset_id):
     (A, B) = h5f[dataset_id + "comp"].value

nnet/feed_forward.py

@@ -423,8 +423,6 @@ class FForwardNetwork():
             features, _ = dnn_features_server.load(seg)
             stat_features, labels = features_server.load(seg)
 
-            #s0 = self.forward(torch.from_numpy(
-            #    dnn_features_server.get_context(feat=features)[0]).type(torch.FloatTensor).to(device))[labels]
             s0 = self.forward(torch.from_numpy(
                 dnn_features_server.get_context(feat=features)[0][labels]).type(torch.FloatTensor).to(device))
             stat_features = stat_features[labels, :]
@@ -473,7 +471,6 @@ class FForwardNetwork():
         """
         model.cpu()
         for idx in seg_indices:
-            print("Compute statistics for {}".format(segset[idx]))
             logging.debug('Compute statistics for {}'.format(segset[idx]))
 
             show = segset[idx]
@@ -520,11 +517,10 @@ class FForwardNetwork():
         ss.stat1 = numpy.zeros((idmap.leftids.shape[0], ndim * feature_size), dtype=numpy.float32)
 
         self.model.cpu()
-        for idx in numpy.arange(len(idmap.segset)):
-            print("Compute statistics for {}".format(idmap.segset[idx]))
-            logging.debug('Compute statistics for {}'.format(idmap.segset[idx]))
+        for idx in numpy.arange(len(idmap.rightids)):
+            logging.debug('Compute statistics for {}'.format(idmap.rightids[idx]))
 
-            show = idmap.segset[idx]
+            show = idmap.rightids[idx]
             channel = 0
             if features_server.features_extractor is not None \
                     and show.endswith(features_server.double_channel_extension[1]):
@@ -533,8 +529,10 @@ class FForwardNetwork():
             features, _ = dnn_features_server.load(show, channel=channel)
             stat_features = stat_features[labels, :]
 
-            s0 = self.model(torch.from_numpy(dnn_features_server.get_context(feat=features)[0]).type(torch.FloatTensor).cpu())[labels]
-            s0.cpu().data.numpy()
+            s0 = self.model(torch.from_numpy(
+                dnn_features_server.get_context(feat=features)[0][labels]).type(torch.FloatTensor).cpu())
+
+            s0 = s0.cpu().data.numpy()
             s1 = numpy.dot(stat_features.T, s0).T
 
             ss.stat0[idx, :] = s0.sum(axis=0)