From 8e52b34a0c2f14175724571b89814277cbf28d59 Mon Sep 17 00:00:00 2001
From: Yu Yang <yuyang18@baidu.com>
Date: Mon, 16 Oct 2017 16:08:01 -0700
Subject: [PATCH] Implement FC layer with helper (#4726)

* Implement FC layer with helper

* Update LayerHelper

* Add debug string for Python ProtoBuf

and Rename `Sync` to `Flush`

* Add check of ProtoBuf initialization

* Layer wrapper for FC

* Fix unittest

* Fix CI

* Add code generator

* AttributeChecker Better error log and speicalize bool

Since lots of types can be cast to bool

* Complete mlp, fit_a_line
---
 python/paddle/v2/framework/framework.py       |  14 +-
 python/paddle/v2/framework/layer_helper.py    | 160 ++++++++++++++++++
 python/paddle/v2/framework/layers.py          | 143 ++++++++++++++++
 .../paddle/v2/framework/tests/test_layers.py  |  43 +++++
 .../v2/framework/tests/test_operator_desc.py  |   2 +-
 5 files changed, 354 insertions(+), 8 deletions(-)
 create mode 100644 python/paddle/v2/framework/layer_helper.py
 create mode 100644 python/paddle/v2/framework/layers.py
 create mode 100644 python/paddle/v2/framework/tests/test_layers.py

diff --git a/python/paddle/v2/framework/framework.py b/python/paddle/v2/framework/framework.py
index d649e69d589..a17f988bf43 100644
--- a/python/paddle/v2/framework/framework.py
+++ b/python/paddle/v2/framework/framework.py
@@ -153,7 +153,8 @@ class OpProtoHolder(object):
             self.op_proto_map[proto.type] = proto
 
     def get_op_proto(self, type):
-        assert type in self.op_proto_map, "Operator \"%s\" has not been registered." % type
+        if type not in self.op_proto_map:
+            raise ValueError("Operator \"%s\" has not been registered." % type)
         return self.op_proto_map[type]
 
 
@@ -374,10 +375,10 @@ class Program(object):
             cls._instance = cls()
         return cls._instance
 
-    def __init__(self):
-        assert not hasattr(self.__class__,
-                           '_instance'), 'Do not call constructor directly!'
-        self.desc = core.ProgramDesc.instance()
+    def __init__(self, desc=None):
+        if desc is None:
+            desc = core.ProgramDesc.instance()
+        self.desc = desc
         self.blocks = [Block(self, 0)]
         self.current_block_idx = 0
 
@@ -428,7 +429,6 @@ class Parameter(Variable):
             if each < 0:
                 raise ValueError("Parameter shape should not be related with "
                                  "batch-size")
-
         Variable.__init__(self, block, shape=shape, dtype=dtype, **kwargs)
         self.trainable = kwargs.get('trainable', True)
         self.init_attr = kwargs.get('initialize_attr', {
@@ -441,7 +441,7 @@ class Parameter(Variable):
         self._append_initialize_ops_()
 
     def _append_initialize_ops_(self):
-        attr = copy.deepcopy(self.init_attr)
+        attr = self.init_attr
         op_type = attr.pop('type', None)
         block = self.block
         assert isinstance(block, Block)
diff --git a/python/paddle/v2/framework/layer_helper.py b/python/paddle/v2/framework/layer_helper.py
new file mode 100644
index 00000000000..26d3e04310b
--- /dev/null
+++ b/python/paddle/v2/framework/layer_helper.py
@@ -0,0 +1,160 @@
+from paddle.v2.framework.framework import Variable, OpProtoHolder, g_program
+import paddle.v2.framework.core as core
+import copy
+import itertools
+
+
+def unique_name(prefix):
+    uid = core.unique_integer()  # unique during whole process.
+    return "_".join([prefix, str(uid)])
+
+
+class LayerHelper(object):
+    def __init__(self, layer_type, **kwargs):
+        self.kwargs = kwargs
+        self.layer_type = layer_type
+        name = self.kwargs.get('name', None)
+        if name is None:
+            self.kwargs['name'] = unique_name(self.layer_type)
+
+    @property
+    def name(self):
+        return self.kwargs['name']
+
+    @property
+    def program(self):
+        prog = self.kwargs.get('program', None)
+        if prog is None:
+            return g_program
+        else:
+            return prog
+
+    def append_op(self, *args, **kwargs):
+        return self.program.current_block().append_op(*args, **kwargs)
+
+    def multiple_input(self, input_param_name='input'):
+        inputs = self.kwargs.get(input_param_name, [])
+        type_error = TypeError(
+            "Input of {0} layer should be Variable or sequence of Variable".
+            format(self.layer_type))
+        if isinstance(inputs, Variable):
+            inputs = [inputs]
+        elif not isinstance(inputs, list) and not isinstance(inputs, tuple):
+            raise type_error
+        else:
+            for each in inputs:
+                if not isinstance(each, Variable):
+                    raise type_error
+        return inputs
+
+    def input(self, input_param_name='input'):
+        inputs = self.multiple_input(input_param_name)
+        if len(inputs) != 1:
+            raise "{0} layer only takes one input".format(self.layer_type)
+        return inputs[0]
+
+    @property
+    def param_attr(self):
+        default = {
+            'name': None,
+            'init_attr': {
+                'type': 'uniform_random',
+                'min': -1.0,
+                'max': 1.0
+            }
+        }
+        actual = self.kwargs.get('param_attr', None)
+        return actual if actual is not None else default
+
+    def bias_attr(self, size, dtype):
+        bias_attr = self.kwargs.get('bias_attr', False)
+        if bias_attr is None or bias_attr:
+            bias_attr = {
+                'name': None,
+                'init_attr': {
+                    'type': 'fill_constant',
+                    'value': 0.0,
+                    'shape': [size],
+                    'dataType': dtype
+                }
+            }
+        return bias_attr
+
+    def multiple_param_attr(self, length):
+        param_attr = self.param_attr
+        if isinstance(param_attr, dict):
+            param_attr = [param_attr]
+
+        if len(param_attr) != 1 and len(param_attr) != length:
+            raise ValueError("parameter number mismatch")
+        elif len(param_attr) == 1 and length != 1:
+            tmp = [None] * length
+            for i in xrange(length):
+                tmp[i] = copy.deepcopy(param_attr[0])
+            param_attr = tmp
+        return param_attr
+
+    def iter_inputs_and_params(self, input_param_name='input'):
+        inputs = self.multiple_input(input_param_name)
+        param_attrs = self.multiple_param_attr(len(inputs))
+        for ipt, param_attr in itertools.izip(inputs, param_attrs):
+            yield ipt, param_attr
+
+    def input_dtype(self, input_param_name='input'):
+        inputs = self.multiple_input(input_param_name)
+        dtype = None
+        for each in inputs:
+            if dtype is None:
+                dtype = each.data_type
+            elif dtype != each.data_type:
+                raise ValueError("Data Type mismatch")
+        return dtype
+
+    def create_parameter(self, attr, shape, dtype, suffix='w'):
+        if attr['name'] is None:
+            attr['name'] = unique_name(".".join([self.name, suffix]))
+        return self.program.global_block().create_parameter(
+            name=attr['name'],
+            dtype=dtype,
+            shape=shape,
+            initialize_attr=attr['init_attr'])
+
+    def create_tmp_variable(self, dtype):
+        return self.program.current_block().create_var(
+            name=unique_name(".".join([self.name, 'tmp'])), dtype=dtype)
+
+    def create_global_variable(self, *args, **kwargs):
+        return self.program.global_block().create_var(*args, **kwargs)
+
+    def append_bias_op(self, input_var):
+        bias_attr = self.bias_attr(
+            self.kwargs['size'], dtype=input_var.data_type)
+        if not bias_attr:
+            return input_var
+        b = self.create_parameter(
+            attr=bias_attr,
+            shape=[self.kwargs['size']],
+            dtype=input_var.data_type,
+            suffix='b')
+        tmp = self.create_tmp_variable(dtype=input_var.data_type)
+        self.append_op(
+            type='elementwise_add',
+            inputs={'X': [input_var],
+                    'Y': [b]},
+            outputs={'Out': [tmp]})
+        return tmp
+
+    def append_activation(self, input_var):
+        act = self.kwargs.get('act', None)
+        if act is None:
+            return input_var
+        if isinstance(act, basestring):
+            act = {'type': act}
+        tmp = self.create_tmp_variable(dtype=input_var.data_type)
+        act_type = act.pop('type')
+        self.append_op(
+            type=act_type,
+            inputs={"X": [input_var]},
+            outputs={"Y": [tmp]},
+            attrs=act)
+        return tmp
diff --git a/python/paddle/v2/framework/layers.py b/python/paddle/v2/framework/layers.py
new file mode 100644
index 00000000000..44b587b116e
--- /dev/null
+++ b/python/paddle/v2/framework/layers.py
@@ -0,0 +1,143 @@
+from paddle.v2.framework.layer_helper import LayerHelper
+import paddle.v2.framework.core as core
+from paddle.v2.framework.framework import OpProtoHolder, Variable
+import re
+
+__all__ = ['fc_layer', 'data_layer', 'cross_entropy']
+
+
+def fc_layer(input,
+             size,
+             param_attr=None,
+             bias_attr=True,
+             name=None,
+             act=None,
+             num_flatten_dims=1,
+             program=None):
+    # create helper
+    helper = LayerHelper('fc', **locals())
+
+    dtype = helper.input_dtype()
+
+    # mul
+    mul_results = []
+    for input_var, param_attr in helper.iter_inputs_and_params():
+        input_shape = input_var.shape
+        param_shape = list(input_shape[num_flatten_dims:]) + [size]
+        w = helper.create_parameter(
+            attr=param_attr, shape=param_shape, dtype=dtype)
+        tmp = helper.create_tmp_variable(dtype)
+        helper.append_op(
+            type="mul",
+            inputs={
+                "X": input_var,
+                "Y": w,
+            },
+            outputs={"Out": tmp},
+            attrs={'x_num_col_dims': num_flatten_dims})
+        mul_results.append(tmp)
+
+    # sum
+    if len(mul_results) == 1:
+        pre_bias = mul_results[0]
+    else:
+        pre_bias = helper.create_tmp_variable(dtype)
+        helper.append_op(
+            type="sum", inputs={"X": mul_results}, outputs={"Out": pre_bias})
+    # add bias
+    pre_activation = helper.append_bias_op(pre_bias)
+    # add activation
+    return helper.append_activation(pre_activation)
+
+
+def data_layer(name,
+               shape,
+               data_type='float32',
+               type=core.VarDesc.VarType.LOD_TENSOR,
+               program=None):
+    helper = LayerHelper('data', **locals())
+    shape = [-1] + shape  # append batch size as -1
+    return helper.create_global_variable(
+        name=name, shape=shape, dtype=data_type, type=type)
+
+
+def _convert_(name):
+    s1 = re.sub('(.)([A-Z][a-z]+)', r'\1_\2', name)
+    return re.sub('([a-z0-9])([A-Z])', r'\1_\2', s1).lower()
+
+
+def _create_op_func_(op_type):
+    op_proto = OpProtoHolder.instance().get_op_proto(op_type)
+    if len(op_proto.outputs) != 1:
+        raise ValueError(
+            "Only one output operator can be automatically generated")
+
+    if op_proto.outputs[0].duplicable:
+        raise ValueError(
+            "Only not duplicable op can be automatically generated")
+
+    o_name = op_proto.outputs[0].name
+
+    def func(**kwargs):
+        helper = LayerHelper(op_type, **kwargs)
+        inputs = dict()
+        dtype = None
+        for ipt in op_proto.inputs:
+            name = _convert_(ipt.name)
+            val = kwargs.pop(name, [])
+            if not isinstance(val, list) and not isinstance(val, tuple):
+                val = [val]
+            for each in val:
+                if not isinstance(each, Variable):
+                    raise ValueError("input of {0} must be variable".format(
+                        op_type))
+
+                if dtype is None:
+                    dtype = each.data_type
+                elif dtype != each.data_type:
+                    raise ValueError(
+                        "operator {0} must input same dtype".format(op_type))
+            inputs[ipt.name] = val
+
+        out = helper.create_tmp_variable(dtype=dtype)
+        helper.append_op(
+            type=op_type, inputs=inputs, outputs={o_name: [out]}, attrs=kwargs)
+        return out
+
+    func.__name__ = op_type
+    globals()[op_type] = func
+    global __all__
+    __all__.append(op_type)
+
+
+_create_op_func_('mean')
+
+
+def cross_entropy(input, label, **kwargs):
+    helper = LayerHelper('cross_entropy', **kwargs)
+    out = helper.create_tmp_variable(dtype=input.data_type)
+    helper.append_op(
+        type='cross_entropy',
+        inputs={'X': [input],
+                'Label': [label]},
+        outputs={'Y': [out]},
+        attrs=kwargs)
+    return out
+
+
+def square_error_cost(input, label, **kwargs):
+    helper = LayerHelper('square_error_cost', **kwargs)
+    minus_out = helper.create_tmp_variable(dtype=input.data_type)
+    helper.append_op(
+        type='elementwise_sub',
+        inputs={'X': [input],
+                'Y': [label]},
+        outputs={'Out': [minus_out]})
+
+    square_out = helper.create_tmp_variable(dtype=input.data_type)
+    helper.append_op(
+        type='pow',
+        inputs={'X': [minus_out]},
+        outputs={'Y': [square_out]},
+        attrs={'factor': 2.0})
+    return square_out
diff --git a/python/paddle/v2/framework/tests/test_layers.py b/python/paddle/v2/framework/tests/test_layers.py
new file mode 100644
index 00000000000..1ef2591cca0
--- /dev/null
+++ b/python/paddle/v2/framework/tests/test_layers.py
@@ -0,0 +1,43 @@
+from paddle.v2.framework.layers import fc_layer, data_layer, cross_entropy, mean, square_error_cost
+from paddle.v2.framework.framework import Program, g_program
+import paddle.v2.framework.core as core
+import unittest
+
+
+class TestBook(unittest.TestCase):
+    def test_fit_a_line(self):
+        pd = core.ProgramDesc.__create_program_desc__()
+        program = Program(desc=pd)
+        x = data_layer(
+            name='x', shape=[13], data_type='float32', program=program)
+        y_predict = fc_layer(input=x, size=1, act=None, program=program)
+
+        y = data_layer(
+            name='y', shape=[1], data_type='float32', program=program)
+        cost = square_error_cost(input=y_predict, label=y, program=program)
+
+        avg_cost = mean(x=cost, program=program)
+        self.assertIsNotNone(avg_cost)
+        print str(program)
+
+    def test_recognize_digits_mlp(self):
+        pd = core.ProgramDesc.__create_program_desc__()
+        program = Program(desc=pd)
+
+        # Change g_program, so the rest layers use `g_program`
+        images = data_layer(
+            name='pixel', shape=[784], data_type='float32', program=program)
+        label = data_layer(
+            name='label', shape=[1], data_type='int32', program=program)
+        hidden1 = fc_layer(input=images, size=128, act='relu', program=program)
+        hidden2 = fc_layer(input=hidden1, size=64, act='relu', program=program)
+        predict = fc_layer(
+            input=hidden2, size=10, act='softmax', program=program)
+        cost = cross_entropy(input=predict, label=label, program=program)
+        avg_cost = mean(x=cost, program=program)
+        self.assertIsNotNone(avg_cost)
+        print str(program)
+
+
+if __name__ == '__main__':
+    unittest.main()
diff --git a/python/paddle/v2/framework/tests/test_operator_desc.py b/python/paddle/v2/framework/tests/test_operator_desc.py
index dfe39c98f7f..af4e980b8ed 100644
--- a/python/paddle/v2/framework/tests/test_operator_desc.py
+++ b/python/paddle/v2/framework/tests/test_operator_desc.py
@@ -16,7 +16,7 @@ class TestOperator(unittest.TestCase):
         try:
             block.append_op(type="no_such_op")
             self.assertFail()
-        except AssertionError as a_err:
+        except ValueError as a_err:
             self.assertEqual(a_err.message,
                              "Operator \"no_such_op\" has not been registered.")
 
-- 
GitLab