Change StaticRNN to fluid.layers.rnn

model_utils/network.py

@@ -59,50 +59,53 @@ def conv_bn_layer(input, filter_size, num_channels_in, num_channels_out, stride,
     return padding_reset
 
 
-def simple_rnn(input, size, param_attr=None, bias_attr=None, is_reverse=False):
+class RNNCell(fluid.layers.RNNCell):
-    '''A simple rnn layer.
+    '''A simple rnn cell.
-    :param input: input layer.
+    :param hidden_size: Dimension of RNN cells.
-    :type input: Variable
+    :type hidden_size: int
-    :param size: Dimension of RNN cells.
-    :type size: int
     :param param_attr: Parameter properties of hidden layer weights that 
                       can be learned
     :type param_attr: ParamAttr
     :param bias_attr: Bias properties of hidden layer weights that can be learned
     :type bias_attr: ParamAttr
-    :param is_reverse: Whether to calculate the inverse RNN
+    :param hidden_activation: Activation for hidden cell
-    :type is_reverse: bool
+    :type hidden_activation: Activation
-    :return: A simple RNN layer.
+    :param activation: Activation for output
-    :rtype: Variable
+    :type activation: Activation
+    :param name: Name of cell
+    :type name: string
     '''
-    if is_reverse:
-        input = fluid.layers.sequence_reverse(x=input)
 
-    pad_value = fluid.layers.assign(input=np.array([0.0], dtype=np.float32))
+    def __init__(self,
-    input, length = fluid.layers.sequence_pad(input, pad_value)
+                 hidden_size,
-    rnn = fluid.layers.StaticRNN()
+                 param_attr=None,
-    input = fluid.layers.transpose(input, [1, 0, 2])
+                 bias_attr=None,
-    with rnn.step():
+                 hidden_activation=None,
-        in_ = rnn.step_input(input)
+                 activation=None,
-        mem = rnn.memory(shape=[-1, size], batch_ref=in_)
+                 dtype="float32",
-        out = fluid.layers.fc(
+                 name="RNNCell"):
-            input=mem,
+        self.hidden_size = hidden_size
-            size=size,
+        self.param_attr = param_attr
-            act=None,
+        self.bias_attr = bias_attr
-            param_attr=param_attr,
+        self.hidden_activation = hidden_activation
-            bias_attr=bias_attr)
+        self.activation = activation or fluid.layers.brelu
-        out = fluid.layers.elementwise_add(out, in_)
+        self.name = name
-        out = fluid.layers.brelu(out)
+
-        rnn.update_memory(mem, out)
+    def call(self, inputs, states):
-        rnn.output(out)
+        new_hidden = fluid.layers.fc(
-
+            input=states,
-    out = rnn()
+            size=self.hidden_size,
-    out = fluid.layers.transpose(out, [1, 0, 2])
+            act=self.hidden_activation,
-    out = fluid.layers.sequence_unpad(x=out, length=length)
+            param_attr=self.param_attr,
-
+            bias_attr=self.bias_attr)
-    if is_reverse:
+        new_hidden = fluid.layers.elementwise_add(new_hidden, inputs)
-        out = fluid.layers.sequence_reverse(x=out)
+        new_hidden = self.activation(new_hidden)
-    return out
+
+        return new_hidden, new_hidden
+
+    @property
+    def state_shape(self):
+        return [self.hidden_size]
 
 
 def bidirectional_simple_rnn_bn_layer(name, input, size, share_weights):
@@ -138,19 +141,31 @@ def bidirectional_simple_rnn_bn_layer(name, input, size, share_weights):
             moving_mean_name=name + '_batch_norm_moving_mean',
             moving_variance_name=name + '_batch_norm_moving_variance')
         #forward and backword in time
-        forward_rnn = simple_rnn(
+        forward_cell = RNNCell(
-            input=input_proj_bn,
+            hidden_size=size,
-            size=size,
+            activation=fluid.layers.brelu,
             param_attr=fluid.ParamAttr(name=name + '_forward_rnn_weight'),
-            bias_attr=fluid.ParamAttr(name=name + '_forward_rnn_bias'),
+            bias_attr=fluid.ParamAttr(name=name + '_forward_rnn_bias'))
-            is_reverse=False)
 
-        reverse_rnn = simple_rnn(
+        pad_value = fluid.layers.assign(input=np.array([0.0], dtype=np.float32))
-            input=input_proj_bn,
+        input, length = fluid.layers.sequence_pad(input_proj_bn, pad_value)
-            size=size,
+        forward_rnn, _ = fluid.layers.rnn(
+            cell=forward_cell, inputs=input, time_major=False, is_reverse=False)
+        forward_rnn = fluid.layers.sequence_unpad(x=forward_rnn, length=length)
+
+        reverse_cell = RNNCell(
+            hidden_size=size,
+            activation=fluid.layers.brelu,
             param_attr=fluid.ParamAttr(name=name + '_reverse_rnn_weight'),
-            bias_attr=fluid.ParamAttr(name=name + '_reverse_rnn_bias'),
+            bias_attr=fluid.ParamAttr(name=name + '_reverse_rnn_bias'))
+        input, length = fluid.layers.sequence_pad(input_proj_bn, pad_value)
+        reverse_rnn, _ = fluid.layers.rnn(
+            cell=reverse_cell,
+            inputs=input,
+            sequence_length=length,
+            time_major=False,
             is_reverse=True)
+        reverse_rnn = fluid.layers.sequence_unpad(x=reverse_rnn, length=length)
 
     else:
         input_proj_forward = fluid.layers.fc(
@@ -183,18 +198,32 @@ def bidirectional_simple_rnn_bn_layer(name, input, size, share_weights):
             moving_mean_name=name + '_reverse_batch_norm_moving_mean',
             moving_variance_name=name + '_reverse_batch_norm_moving_variance')
         # forward and backward in time
-        forward_rnn = simple_rnn(
+        forward_cell = RNNCell(
-            input=input_proj_bn_forward,
+            hidden_size=size,
-            size=size,
+            activation=fluid.layers.brelu,
             param_attr=fluid.ParamAttr(name=name + '_forward_rnn_weight'),
-            bias_attr=fluid.ParamAttr(name=name + '_forward_rnn_bias'),
+            bias_attr=fluid.ParamAttr(name=name + '_forward_rnn_bias'))
-            is_reverse=False)
+
-        reverse_rnn = simple_rnn(
+        pad_value = fluid.layers.assign(input=np.array([0.0], dtype=np.float32))
-            input=input_proj_bn_backward,
+        input, length = fluid.layers.sequence_pad(input_proj_bn, pad_value)
-            size=size,
+        forward_rnn, _ = fluid.layers.rnn(
+            cell=forward_cell, inputs=input, time_major=False, is_reverse=False)
+        forward_rnn = fluid.layers.sequence_unpad(x=forward_rnn, length=length)
+
+        reverse_cell = RNNCell(
+            hidden_size=size,
+            activation=fluid.layers.brelu,
             param_attr=fluid.ParamAttr(name=name + '_reverse_rnn_weight'),
-            bias_attr=fluid.ParamAttr(name=name + '_reverse_rnn_bias'),
+            bias_attr=fluid.ParamAttr(name=name + '_reverse_rnn_bias'))
+        input, length = fluid.layers.sequence_pad(input_proj_bn, pad_value)
+        reverse_rnn, _ = fluid.layers.rnn(
+            cell=reverse_cell,
+            inputs=input,
+            sequence_length=length,
+            time_major=False,
             is_reverse=True)
+        reverse_rnn = fluid.layers.sequence_unpad(x=reverse_rnn, length=length)
+
     out = fluid.layers.concat(input=[forward_rnn, reverse_rnn], axis=1)
     return out