Merge pull request #938 from zhxfl/new_config

New config

fluid/DeepASR/data_utils/augmentor/tests/test_data_trans.py

@@ -8,6 +8,7 @@ import numpy as np
 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
 
 
 class TestTransMeanVarianceNorm(unittest.TestCase):
@@ -112,5 +113,24 @@ class TestTransSplict(unittest.TestCase):
                         self.assertAlmostEqual(feature[i][j * 10 + k], cur_val)
 
 
+class TestTransDelay(unittest.TestCase):
+    """unittest TransDelay
+    """
+
+    def test_perform(self):
+        label = np.zeros((10, 1), dtype="int64")
+        for i in xrange(10):
+            label[i][0] = i
+
+        trans = trans_delay.TransDelay(5)
+        (_, label, _) = trans.perform_trans((None, label, None))
+
+        for i in xrange(5):
+            self.assertAlmostEqual(label[i + 5][0], i)
+
+        for i in xrange(5):
+            self.assertAlmostEqual(label[i][0], 0)
+
+
 if __name__ == '__main__':
     unittest.main()

fluid/DeepASR/data_utils/augmentor/trans_delay.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import math
+
+
+class TransDelay(object):
+    """ Delay label, and copy first label value in the front. 
+        Attributes:
+            _delay_time : the delay frame num of label 
+    """
+
+    def __init__(self, delay_time):
+        """init construction
+            Args:
+                delay_time : the delay frame num of label
+        """
+        self._delay_time = delay_time
+
+    def perform_trans(self, sample):
+        """ 
+            Args:
+                sample(object):input sample, contain feature numpy and label numpy, sample name list
+            Returns:
+                (feature, label, name)
+        """
+        (feature, label, name) = sample
+
+        shape = label.shape
+        assert len(shape) == 2
+        label[self._delay_time:shape[0]] = label[0:shape[0] - self._delay_time]
+        for i in xrange(self._delay_time):
+            label[i][0] = label[self._delay_time][0]
+
+        return (feature, label, name)

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
@@ -36,7 +37,7 @@ 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',
@@ -179,7 +180,7 @@ 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 +32,23 @@ 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)
+
+        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 +60,6 @@ def stacked_lstmp_model(frame_dim,
                 proj_activation="tanh")
             bn = fluid.layers.batch_norm(
                 input=proj,
-                act="sigmoid",
                 is_test=not is_train,
                 momentum=0.9,
                 epsilon=1e-05,
@@ -80,7 +77,10 @@ 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

@@ -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 data_utils.util import lodtensor_to_ndarray
 from model_utils.model import stacked_lstmp_model
@@ -33,7 +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',
@@ -53,7 +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',
@@ -157,6 +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)
 
@@ -171,7 +173,7 @@ def train(args):
     ltrans = [
         trans_add_delta.TransAddDelta(2, 2),
         trans_mean_variance_norm.TransMeanVarianceNorm(args.mean_var),
-        trans_splice.TransSplice()
+        trans_splice.TransSplice(5, 5), trans_delay.TransDelay(5)
     ]
 
     feature_t = fluid.LoDTensor()
@@ -220,6 +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)