aishell 3040

fluid/DeepASR/infer_by_ckpt.py

@@ -12,6 +12,7 @@ import paddle.fluid as fluid
 import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
 import data_utils.augmentor.trans_add_delta as trans_add_delta
 import data_utils.augmentor.trans_splice as trans_splice
+import data_utils.augmentor.trans_delay as trans_delay
 import data_utils.async_data_reader as reader
 from decoder.post_decode_faster import Decoder
 from data_utils.util import lodtensor_to_ndarray
@@ -35,7 +36,7 @@ def parse_args():
     parser.add_argument(
         '--frame_dim',
         type=int,
-        default=120 * 11,
+        default=80,
         help='Frame dimension of feature data. (default: %(default)d)')
     parser.add_argument(
         '--stacked_num',
@@ -155,7 +156,7 @@ def infer_from_ckpt(args):
     ltrans = [
         trans_add_delta.TransAddDelta(2, 2),
         trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
-        trans_splice.TransSplice()
+        trans_splice.TransSplice(), trans_delay.TransDelay(5)
     ]
 
     feature_t = fluid.LoDTensor()

fluid/DeepASR/model_utils/model.py

@@ -32,25 +32,31 @@ def stacked_lstmp_model(frame_dim,
 
     # network configuration
     def _net_conf(feature, label):
-        seq_conv1 = fluid.layers.sequence_conv(
+        conv1 = fluid.layers.conv2d(
             input=feature,
-            num_filters=1024,
+            num_filters=32,
             filter_size=3,
-            filter_stride=1,
-            bias_attr=True)
-        bn1 = fluid.layers.batch_norm(
-            input=seq_conv1,
-            act="sigmoid",
-            is_test=not is_train,
-            momentum=0.9,
-            epsilon=1e-05,
-            data_layout='NCHW')
+            stride=1,
+            padding=1,
+            bias_attr=True,
+            act="relu")
 
-        stack_input = bn1
+        pool1 = fluid.layers.pool2d(
+            conv1, pool_size=3, pool_type="max", pool_stride=2, pool_padding=0)
+
+        #bn1 = fluid.layers.batch_norm(
+        #    input=pool1,
+        #    act="sigmoid",
+        #    is_test=not is_train,
+        #    momentum=0.9,
+        #    epsilon=1e-05,
+        #    data_layout='NCHW')
+
+        stack_input = pool1
         for i in range(stacked_num):
             fc = fluid.layers.fc(input=stack_input,
                                  size=hidden_dim * 4,
-                                 bias_attr=True)
+                                 bias_attr=None)
             proj, cell = fluid.layers.dynamic_lstmp(
                 input=fc,
                 size=hidden_dim * 4,
@@ -62,7 +68,7 @@ def stacked_lstmp_model(frame_dim,
                 proj_activation="tanh")
             bn = fluid.layers.batch_norm(
                 input=proj,
-                act="sigmoid",
+                #act="linear",
                 is_test=not is_train,
                 momentum=0.9,
                 epsilon=1e-05,
@@ -80,7 +86,10 @@ def stacked_lstmp_model(frame_dim,
 
     # data feeder
     feature = fluid.layers.data(
-        name="feature", shape=[-1, frame_dim], dtype="float32", lod_level=1)
+        name="feature",
+        shape=[-1, 3, 11, frame_dim],
+        dtype="float32",
+        lod_level=1)
     label = fluid.layers.data(
         name="label", shape=[-1, 1], dtype="int64", lod_level=1)
 

fluid/DeepASR/tools/profile.py

@@ -13,6 +13,7 @@ import _init_paths
 import data_utils.augmentor.trans_mean_variance_norm as trans_mean_variance_norm
 import data_utils.augmentor.trans_add_delta as trans_add_delta
 import data_utils.augmentor.trans_splice as trans_splice
+import data_utils.augmentor.trans_delay as trans_delay
 import data_utils.async_data_reader as reader
 from model_utils.model import stacked_lstmp_model
 from data_utils.util import lodtensor_to_ndarray
@@ -87,7 +88,7 @@ def parse_args():
     parser.add_argument(
         '--max_batch_num',
         type=int,
-        default=10,
+        default=11,
         help='Maximum number of batches for profiling. (default: %(default)d)')
     parser.add_argument(
         '--first_batches_to_skip',
@@ -146,10 +147,10 @@ def profile(args):
     ltrans = [
         trans_add_delta.TransAddDelta(2, 2),
         trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
-        trans_splice.TransSplice()
+        trans_splice.TransSplice(), trans_delay.TransDelay(5)
     ]
 
-    data_reader = reader.AsyncDataReader(args.feature_lst, args.label_lst)
+    data_reader = reader.AsyncDataReader(args.feature_lst, args.label_lst, -1)
     data_reader.set_transformers(ltrans)
 
     feature_t = fluid.LoDTensor()

fluid/DeepASR/train.py

@@ -34,7 +34,7 @@ def parse_args():
     parser.add_argument(
         '--frame_dim',
         type=int,
-        default=120 * 11,
+        default=80,
         help='Frame dimension of feature data. (default: %(default)d)')
     parser.add_argument(
         '--stacked_num',
@@ -54,7 +54,7 @@ def parse_args():
     parser.add_argument(
         '--class_num',
         type=int,
-        default=1749,
+        default=3040,
         help='Number of classes in label. (default: %(default)d)')
     parser.add_argument(
         '--pass_num',
@@ -158,6 +158,7 @@ def train(args):
     # program for test
     test_program = fluid.default_main_program().clone()
 
+    #optimizer = fluid.optimizer.Momentum(learning_rate=args.learning_rate, momentum=0.9)
     optimizer = fluid.optimizer.Adam(learning_rate=args.learning_rate)
     optimizer.minimize(avg_cost)
 
@@ -172,7 +173,7 @@ def train(args):
     ltrans = [
         trans_add_delta.TransAddDelta(2, 2),
         trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
-        trans_splice.TransSplice(), trans_delay.TransDelay(5)
+        trans_splice.TransSplice(5, 5), trans_delay.TransDelay(5)
     ]
 
     feature_t = fluid.LoDTensor()
@@ -221,6 +222,8 @@ def train(args):
                                                  args.minimum_batch_size)):
             # load_data
             (features, labels, lod, name_lst) = batch_data
+            features = np.reshape(features, (-1, 11, 3, args.frame_dim))
+            features = np.transpose(features, (0, 2, 1, 3))
             feature_t.set(features, place)
             feature_t.set_lod([lod])
             label_t.set(labels, place)