Fix basic_gru and docs of gru_unit and dynamic_gru (#19393)

* Fix docs of gru_unit and dynamic_gru.

Fix basic_gru in rnn_impl.py.

Add error messages for param_attr setting in layer_norm api.

Add int64 dtype for expand.

test=develop

* Reopen unit-tests of basic_gru/basic_lstm in rnn_impl.py.
test=develop

* Add unit test for layer_norm api.
test=develop

* Remove the deprecated gru doc fix. test=develop

* Fix basic_gru test coverage. test=develop

* Update API.spec. test=develop

* Update API.spec. test=develop

* Fix test_basic_gru coverage test. test=develop

* Update test_basic_gru in test_layers to use fluid.data
test=develop

* Update test_basic_gru for coverage. test=develop

paddle/fluid/API.spec

@@ -581,7 +581,7 @@ paddle.fluid.contrib.BasicGRUUnit.set_dict (ArgSpec(args=['self', 'stat_dict', '
 paddle.fluid.contrib.BasicGRUUnit.state_dict (ArgSpec(args=['self', 'destination', 'include_sublayers'], varargs=None, keywords=None, defaults=(None, True)), ('document', 'bb62a4e57bc58f171091fe78b1e7f7f3'))
 paddle.fluid.contrib.BasicGRUUnit.sublayers (ArgSpec(args=['self', 'include_sublayers'], varargs=None, keywords=None, defaults=(True,)), ('document', 'c960169ad78b8306f5dc16d47e609340'))
 paddle.fluid.contrib.BasicGRUUnit.train (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.basic_gru (ArgSpec(args=['input', 'init_hidden', 'hidden_size', 'num_layers', 'sequence_length', 'dropout_prob', 'bidirectional', 'batch_first', 'param_attr', 'bias_attr', 'gate_activation', 'activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(1, None, 0.0, False, True, None, None, None, None, 'float32', 'basic_gru')), ('document', '0afcbe4fbe1b8c35eda58b4efe48f9fd'))
+paddle.fluid.contrib.basic_gru (ArgSpec(args=['input', 'init_hidden', 'hidden_size', 'num_layers', 'sequence_length', 'dropout_prob', 'bidirectional', 'batch_first', 'param_attr', 'bias_attr', 'gate_activation', 'activation', 'dtype', 'name'], varargs=None, keywords=None, defaults=(1, None, 0.0, False, True, None, None, None, None, 'float32', 'basic_gru')), ('document', 'a9930834e44da910282894449a6e8a6d'))
 paddle.fluid.contrib.BasicLSTMUnit ('paddle.fluid.contrib.layers.rnn_impl.BasicLSTMUnit', ('document', '3d0b2e3172ce58e1304199efee066c99'))
 paddle.fluid.contrib.BasicLSTMUnit.__init__ (ArgSpec(args=['self', 'name_scope', 'hidden_size', 'param_attr', 'bias_attr', 'gate_activation', 'activation', 'forget_bias', 'dtype'], varargs=None, keywords=None, defaults=(None, None, None, None, 1.0, 'float32')), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
 paddle.fluid.contrib.BasicLSTMUnit.add_parameter (ArgSpec(args=['self', 'name', 'parameter'], varargs=None, keywords=None, defaults=None), ('document', 'e75f4ab651bed0c9129a9a2c10aaaa7d'))
@@ -599,7 +599,7 @@ paddle.fluid.contrib.BasicLSTMUnit.set_dict (ArgSpec(args=['self', 'stat_dict',
 paddle.fluid.contrib.BasicLSTMUnit.state_dict (ArgSpec(args=['self', 'destination', 'include_sublayers'], varargs=None, keywords=None, defaults=(None, True)), ('document', 'bb62a4e57bc58f171091fe78b1e7f7f3'))
 paddle.fluid.contrib.BasicLSTMUnit.sublayers (ArgSpec(args=['self', 'include_sublayers'], varargs=None, keywords=None, defaults=(True,)), ('document', 'c960169ad78b8306f5dc16d47e609340'))
 paddle.fluid.contrib.BasicLSTMUnit.train (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
-paddle.fluid.contrib.basic_lstm (ArgSpec(args=['input', 'init_hidden', 'init_cell', 'hidden_size', 'num_layers', 'sequence_length', 'dropout_prob', 'bidirectional', 'batch_first', 'param_attr', 'bias_attr', 'gate_activation', 'activation', 'forget_bias', 'dtype', 'name'], varargs=None, keywords=None, defaults=(1, None, 0.0, False, True, None, None, None, None, 1.0, 'float32', 'basic_lstm')), ('document', 'fe4d0c3c55a162b8cfe10b05fabb7ce4'))
+paddle.fluid.contrib.basic_lstm (ArgSpec(args=['input', 'init_hidden', 'init_cell', 'hidden_size', 'num_layers', 'sequence_length', 'dropout_prob', 'bidirectional', 'batch_first', 'param_attr', 'bias_attr', 'gate_activation', 'activation', 'forget_bias', 'dtype', 'name'], varargs=None, keywords=None, defaults=(1, None, 0.0, False, True, None, None, None, None, 1.0, 'float32', 'basic_lstm')), ('document', '2db201a8a2fb268097e220036dfe0b88'))
 paddle.fluid.contrib.ctr_metric_bundle (ArgSpec(args=['input', 'label'], varargs=None, keywords=None, defaults=None), ('document', 'b68d12366896c41065fc3738393da2aa'))
 paddle.fluid.data (ArgSpec(args=['name', 'shape', 'dtype', 'lod_level'], varargs=None, keywords=None, defaults=('float32', 0)), ('document', 'a44fce9b5c8919bf5937a1cc0fe484ca'))
 paddle.fluid.dygraph.Layer ('paddle.fluid.dygraph.layers.Layer', ('document', 'ec49de13e8d63aa20b51d24fc53a1d15'))

python/paddle/fluid/contrib/layers/rnn_impl.py

@@ -127,7 +127,7 @@ class BasicGRUUnit(Layer):
         r_hidden = r * pre_hidden
 
         candidate = layers.matmul(
-            layers.concat([input, pre_hidden], 1), self._candidate_weight)
+            layers.concat([input, r_hidden], 1), self._candidate_weight)
         candidate = layers.elementwise_add(candidate, self._candidate_bias)
 
         c = self._activation(candidate)
@@ -179,6 +179,11 @@ def basic_gru(input,
         dropout_prob(float|0.0): Dropout prob, dropout ONLY works after rnn output of earch layers, 
                              NOT between time steps
         bidirectional (bool|False): If it is bidirectional
+        batch_first (bool|True): The shape format of the input and output tensors. If true,
+            the shape format should be :attr:`[batch_size, seq_len, hidden_size]`. If false,
+            the shape format should be :attr:`[seq_len, batch_size, hidden_size]`. By default
+            this function accepts input and emits output in batch-major form to be consistent
+            with most of data format, though a bit less efficient because of extra transposes.
         param_attr(ParamAttr|None): The parameter attribute for the learnable
             weight matrix. Note:
             If it is set to None or one attribute of ParamAttr, gru_unit will
@@ -345,7 +350,7 @@ def basic_gru(input,
         last_hidden = fw_last_hidden
 
         if batch_first:
-            rnn_out = fluid.layser.transpose(rnn_out, [1, 0, 2])
+            rnn_out = layers.transpose(rnn_out, [1, 0, 2])
 
         return rnn_out, last_hidden
 
@@ -404,6 +409,11 @@ def basic_lstm(input,
         dropout_prob(float|0.0): Dropout prob, dropout ONLY work after rnn output of earch layers, 
                              NOT between time steps
         bidirectional (bool|False): If it is bidirectional
+        batch_first (bool|True): The shape format of the input and output tensors. If true,
+            the shape format should be :attr:`[batch_size, seq_len, hidden_size]`. If false,
+            the shape format should be :attr:`[seq_len, batch_size, hidden_size]`. By default
+            this function accepts input and emits output in batch-major form to be consistent
+            with most of data format, though a bit less efficient because of extra transposes.
         param_attr(ParamAttr|None): The parameter attribute for the learnable
             weight matrix. Note:
             If it is set to None or one attribute of ParamAttr, lstm_unit will

python/paddle/fluid/layers/nn.py

@@ -1519,6 +1519,7 @@ def gru_unit(input,
         attrs={
             'activation': 2,  # tanh
             'gate_activation': 1,  # sigmoid
+            'origin_mode': origin_mode
         })
 
     return updated_hidden, reset_hidden_pre, gate
@@ -4646,17 +4647,24 @@ def layer_norm(input,
     input_shape = input.shape
     param_shape = [reduce(lambda x, y: x * y, input_shape[begin_norm_axis:])]
     if scale:
+        assert param_attr is not False, "param_attr should not be False when using scale."
         scale = helper.create_parameter(
             attr=helper.param_attr,
             shape=param_shape,
             dtype=dtype,
             default_initializer=Constant(1.0))
         inputs['Scale'] = scale
+    else:
+        if param_attr:
+            warnings.warn("param_attr is only avaliable with scale is True.")
     if shift:
-        assert bias_attr is not False
+        assert bias_attr is not False, "bias_attr should not be False when using shift."
         bias = helper.create_parameter(
             attr=helper.bias_attr, shape=param_shape, dtype=dtype, is_bias=True)
         inputs['Bias'] = bias
+    else:
+        if bias_attr:
+            warnings.warn("bias_attr is only avaliable with shift is True.")
 
     # create output
     mean_out = helper.create_variable_for_type_inference(

python/paddle/fluid/tests/unittests/test_basic_gru_api.py

@@ -25,6 +25,8 @@ from paddle.fluid import framework
 
 import numpy as np
 
+np.set_seed(123)
+
 SIGMOID_THRESHOLD_MIN = -40.0
 SIGMOID_THRESHOLD_MAX = 13.0
 EXP_MAX_INPUT = 40.0
@@ -65,7 +67,7 @@ def gru_np(input,
         r_hidden = r * pre_hidden
 
         candidate = np.matmul(
-            np.concatenate([step_in, pre_hidden], 1), candidate_w)
+            np.concatenate([step_in, r_hidden], 1), candidate_w)
 
         candidate += candidate_b
         c = tanh(candidate)

python/paddle/fluid/tests/unittests/test_basic_gru_unit_op.py

@@ -25,6 +25,8 @@ from paddle.fluid import framework
 
 import numpy as np
 
+np.set_seed(123)
+
 SIGMOID_THRESHOLD_MIN = -40.0
 SIGMOID_THRESHOLD_MAX = 13.0
 EXP_MAX_INPUT = 40.0
@@ -53,7 +55,7 @@ def step(step_in, pre_hidden, gate_w, gate_b, candidate_w, candidate_b):
 
     r_hidden = r * pre_hidden
 
-    candidate = np.matmul(np.concatenate([step_in, pre_hidden], 1), candidate_w)
+    candidate = np.matmul(np.concatenate([step_in, r_hidden], 1), candidate_w)
 
     candidate += candidate_b
     c = tanh(candidate)

python/paddle/fluid/tests/unittests/test_basic_lstm_api.py

@@ -25,6 +25,8 @@ from paddle.fluid import framework
 
 import numpy as np
 
+np.set_seed(123)
+
 SIGMOID_THRESHOLD_MIN = -40.0
 SIGMOID_THRESHOLD_MAX = 13.0
 EXP_MAX_INPUT = 40.0

python/paddle/fluid/tests/unittests/test_basic_lstm_unit_op.py

@@ -25,6 +25,8 @@ from paddle.fluid import framework
 
 import numpy as np
 
+np.set_seed(123)
+
 SIGMOID_THRESHOLD_MIN = -40.0
 SIGMOID_THRESHOLD_MAX = 13.0
 EXP_MAX_INPUT = 40.0

python/paddle/fluid/tests/unittests/test_layer_norm_op.py

@@ -175,5 +175,38 @@ class TestLayerNormOp(unittest.TestCase):
         self.check_forward_backward(shape=[2, 3, 4, 5], begin_norm_axis=3)
 
 
+class TestLayerNormAPI(unittest.TestCase):
+    def test_case(self):
+        x = fluid.layers.data(
+            name='x',
+            shape=[64, 32, 256],
+            dtype='float32',
+            append_batch_size=False)
+        x = fluid.layers.layer_norm(
+            x,
+            scale=True,
+            shift=True,
+            begin_norm_axis=1,
+            epsilon=1e-05,
+            param_attr=None,
+            bias_attr=None)
+        x = fluid.layers.layer_norm(
+            x,
+            scale=False,
+            shift=False,
+            begin_norm_axis=1,
+            epsilon=1e-05,
+            param_attr=None,
+            bias_attr=None)
+        x = fluid.layers.layer_norm(
+            x,
+            scale=False,
+            shift=False,
+            begin_norm_axis=1,
+            epsilon=1e-05,
+            param_attr="scale",
+            bias_attr="shift")
+
+
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/tests/unittests/test_layers.py

@@ -3060,6 +3060,29 @@ class TestBook(LayerTest):
             evaluator = fluid.evaluator.EditDistance(predict, label)
             return evaluator.metrics
 
+    def test_basic_gru(self):
+        input_size = 128
+        hidden_size = 256
+        with self.static_graph():
+            input = fluid.data(
+                name="input", shape=[None, None, input_size], dtype='float32')
+            pre_hidden = fluid.data(
+                name="pre_hidden", shape=[None, hidden_size], dtype='float32')
+            sequence_length = fluid.data(
+                name="sequence_length", shape=[None], dtype='int32')
+
+            for bidirectional in [True, False]:
+                for batch_first in [True, False]:
+                    rnn_out, last_hidden = fluid.contrib.layers.basic_gru(
+                        input,
+                        pre_hidden,
+                        hidden_size=256,
+                        num_layers=2,
+                        sequence_length=sequence_length,
+                        dropout_prob=0.5,
+                        bidirectional=bidirectional,
+                        batch_first=batch_first)
+
 
 if __name__ == '__main__':
     unittest.main()