Fix/sequence pool (PaddlePaddle#5229)

* "modify layers.py"

* "fix pool interface"

* "add export type to layers"

* "fix based on comment"

python/paddle/v2/framework/layers.py

@@ -5,7 +5,8 @@
 
 __all__ = [
     'fc', 'data', 'cross_entropy', 'conv2d', 'pool2d', 'embedding', 'concat',
-    'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'accuracy'
+    'StaticRNN', 'cast', 'sequence_conv', 'sequence_pool', 'sums', 'cos_sim',
+    'batch_norm', 'accuracy'
 ]
 
 
@@ -165,18 +166,6 @@ def func(**kwargs):
 _create_op_func_('reshape')
 
 
-def cast(x, data_type, program=None):
-    helper = LayerHelper('cast', **locals())
-    out = helper.create_tmp_variable(dtype=data_type)
-    helper.append_op(
-        type='cast',
-        inputs={'X': [x]},
-        outputs={'Out': [out]},
-        attrs={'in_data_type': x.data_type,
-               'out_data_type': out.data_type})
-    return out
-
-
 def cast(x, data_type, program=None):
     helper = LayerHelper('cast', **locals())
     out = helper.create_tmp_variable(dtype=data_type)
@@ -191,9 +180,7 @@ def cast(x, data_type, program=None):
 
 def concat(input, axis, program=None, init_program=None):
     helper = LayerHelper('concat', **locals())
-    if not isinstance(input, list) and not isinstance(input, tuple):
-        input = [input]
-    out = helper.create_tmp_variable(dtype=input[0].data_type)
+    out = helper.create_tmp_variable(dtype=helper.input_dtype())
     helper.append_op(
         type='concat',
         inputs={'X': input},
@@ -202,6 +189,28 @@ def concat(input, axis, program=None, init_program=None):
     return out
 
 
+def sums(input, program=None, init_program=None):
+    helper = LayerHelper('sum', **locals())
+    out = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(type='sum', inputs={'X': [input]}, outputs={'Out': out})
+    return out
+
+
+def cos_sim(X, Y, program=None, init_program=None):
+    helper = LayerHelper('cos_sim', **locals())
+    out = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
+    xnorm = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
+    ynorm = helper.create_tmp_variable(dtype=helper.input_dtype("X"))
+    helper.append_op(
+        type='cos_sim',
+        inputs={'X': [X],
+                'Y': [Y]},
+        outputs={'Out': [out],
+                 'XNorm': [xnorm],
+                 'YNorm': [ynorm]})
+    return out, xnorm, ynorm
+
+
 def cross_entropy(input, label, **kwargs):
     helper = LayerHelper('cross_entropy', **kwargs)
     out = helper.create_tmp_variable(dtype=input.data_type)
@@ -254,9 +263,7 @@ def accuracy(input, label, k=1, **kwargs):
 
 def sequence_conv(input,
                   num_filters,
-                  name=None,
                   filter_size=3,
-                  act=None,
                   stride=1,
                   padding=None,
                   bias_attr=None,
@@ -270,7 +277,7 @@ def sequence_conv(input,
     helper = LayerHelper('sequence_conv', **locals())
     dtype = helper.input_dtype()
 
-    filter_shape = [num_filters, filter_size]
+    filter_shape = [filter_size * input.shape[1], num_filters]
     filter = helper.create_parameter(
         attr=helper.param_attr, shape=filter_shape, dtype=dtype)
     pre_bias = helper.create_tmp_variable(dtype)
@@ -279,15 +286,14 @@ def sequence_conv(input,
         type='sequence_conv',
         inputs={
             'X': [input],
-            'Filter': filter,
+            'Filter': [filter],
         },
         outputs={"Out": pre_bias},
         attrs={
             'context_stride': stride,
             'context_start': 0,
             'context_length': filter_size
         })
-
     pre_act = helper.append_bias_op(pre_bias)
     return helper.append_activation(pre_act)
 
@@ -344,31 +350,32 @@ def conv2d(input,
     return helper.append_activation(pre_act)
 
 
-def sequence_pool(input,
-                  pool_size,
-                  pool_type,
-                  pool_stride=1,
-                  pool_padding=0,
-                  global_pooling=False,
-                  program=None,
-                  init_program=None):
+def sequence_pool(input, pool_type, program=None, init_program=None):
     # FIXME(dzh) : want to unify the argument of python layer
     # function. So we ignore some unecessary attributes
 
-    ENUM_POOL_TYPE = set(["max", "avg", "sqrt", "last", "first"])
-    if pool_type not in ENUM_POOL_TYPE:
+    ENUM_POOL_TYPE = dict({
+        "AVERAGE": 0,
+        "SUM": 1,
+        "SQRT": 2,
+        "MAX": 3,
+        "LAST": 4,
+        "FIRST": 5
+    })
+    if pool_type.upper() not in ENUM_POOL_TYPE:
         raise ValueError("Unknown pool_type: '%s'. It can only be %s.",
-                         str(pool_type), " ".join(ENUM_POOL_TYPE))
+                         str(pool_type), " ".join(ENUM_POOL_TYPE.keys()))
 
     helper = LayerHelper('sequence_pool', **locals())
     dtype = helper.input_dtype()
     pool_out = helper.create_tmp_variable(dtype)
 
+    # FIXME(dzh): strategy
     helper.append_op(
         type="sequence_pool",
         inputs={"X": [input]},
-        outputs={"Out": pool_out},
-        attrs={"strategy": pool_type})
+        outputs={"Out": [pool_out]},
+        attrs={"strategy": ENUM_POOL_TYPE[pool_type.upper()]})
 
     return pool_out
 

python/paddle/v2/framework/nets.py

@@ -101,24 +101,19 @@ def __extend_list__(obj):
 def sequence_conv_pool(input,
                        num_filters,
                        filter_size,
-                       pool_size,
-                       pool_stride,
-                       act,
+                       pool_type="max",
                        program=None,
                        init_program=None):
     conv_out = layers.sequence_conv(
         input=input,
         num_filters=num_filters,
         filter_size=filter_size,
-        act=act,
         program=program,
         init_program=init_program)
 
     pool_out = layers.sequence_pool(
         input=conv_out,
-        pool_size=pool_size,
-        pool_type='max',
-        pool_stride=pool_stride,
+        pool_type=pool_type,
         program=program,
         init_program=init_program)
     return pool_out