Update

1. Add init_program to hold initializers
2. bug fix

paddle/operators/uniform_random_op.cc

@@ -53,10 +53,10 @@ class UniformRandomOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE(
         ctx->Attrs().Get<float>("min") < ctx->Attrs().Get<float>("max"),
         "uniform_random's min must less then max");
-    auto& dims = ctx->Attrs().Get<std::vector<int>>("dims");
+    auto& shape = ctx->Attrs().Get<std::vector<int>>("shape");
     std::vector<int64_t> temp;
-    temp.reserve(dims.size());
-    for (auto dim : dims) {
+    temp.reserve(shape.size());
+    for (auto dim : shape) {
       temp.push_back(static_cast<int64_t>(dim));
     }
     ctx->SetOutputDim("Out", framework::make_ddim(temp));
@@ -78,7 +78,7 @@ class UniformRandomOpMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(Uniform random operator.
 Used to initialize tensor with uniform random generator.
 )DOC");
-    AddAttr<std::vector<int>>("dims", "the dimension of random tensor");
+    AddAttr<std::vector<int>>("shape", "the dimension of random tensor");
     AddAttr<float>("min", "Minimum value of uniform random").SetDefault(-1.0f);
     AddAttr<float>("max", "Maximun value of uniform random").SetDefault(1.0f);
     AddAttr<int>("seed",

python/paddle/v2/framework/framework.py

@@ -343,6 +343,8 @@ class Block(object):
     def create_parameter(self, *args, **kwargs):
         global_block = self.program.global_block()
         param = Parameter(global_block, *args, **kwargs)
+        if 'init_attr' in kwargs:
+            self._prepend_initialize_ops_(param, kwargs['init_attr'])
         return param
 
     def append_op(self, *args, **kwargs):
@@ -401,6 +403,17 @@ class Block(object):
         for index in range(len(self.ops)):
             assert self.ops[index].desc == ops_in_cpp[index]
 
+    def _prepend_initialize_ops_(self, param, init_attr):
+        op_type = init_attr['type']
+        init_attr['shape'] = param.shape
+        init_attr['data_type'] = int(param.data_type)
+        op = self.prepend_op(
+            type=op_type,
+            inputs=None,
+            outputs={'Out': [param]},
+            attrs=init_attr)
+        param.op = op
+
 
 class Program(object):
     def __init__(self):
@@ -475,27 +488,10 @@ class Parameter(Variable):
         Variable.__init__(
             self, block, persistable=True, shape=shape, dtype=dtype, **kwargs)
         self.trainable = kwargs.get('trainable', True)
-        self.init_attr = kwargs.get('initialize_attr', {
-            'type': 'uniform_random',
-            'min': -1.0,
-            'max': 1.0
-        })
 
         self.optimize_attr = kwargs.get('optimize_attr', {'learning_rate': 1.0})
-        self._append_initialize_ops_()
-
-    def _append_initialize_ops_(self):
-        attr = self.init_attr
-        op_type = attr.pop('type', None)
-        block = self.block
-        assert isinstance(block, Block)
-        shape = self.shape
-        attr['dims'] = shape
-        attr['data_type'] = int(self.data_type)
-        op = block.prepend_op(
-            type=op_type, inputs=None, outputs={'Out': [self]}, attrs=attr)
-        self.op = op
 
 
 # program is a global instance.
 g_program = Program()
+g_init_program = Program()

python/paddle/v2/framework/layer_helper.py

-from paddle.v2.framework.framework import Variable, OpProtoHolder, g_program
+from paddle.v2.framework.framework import Variable, OpProtoHolder, g_program, g_init_program
 import paddle.v2.framework.core as core
 import copy
 import itertools
@@ -29,6 +29,14 @@ class LayerHelper(object):
         else:
             return prog
 
+    @property
+    def init_program(self):
+        prog = self.kwargs.get('init_program', None)
+        if prog is None:
+            return g_init_program
+        else:
+            return prog
+
     def append_op(self, *args, **kwargs):
         return self.program.current_block().append_op(*args, **kwargs)
 
@@ -73,9 +81,9 @@ class LayerHelper(object):
                 'name': None,
                 'init_attr': {
                     'type': 'fill_constant',
-                    'value': 0.0,
-                    'shape': shape,
-                    'dataType': dtype
+                    'value': 0.0
+                    #'shape': shape,
+                    #'dataType': dtype
                 }
             }
         return bias_attr
@@ -113,11 +121,13 @@ class LayerHelper(object):
     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(
+        self.init_program.global_block().create_parameter(
             name=attr['name'],
             dtype=dtype,
             shape=shape,
-            initialize_attr=attr['init_attr'])
+            init_attr=attr['init_attr'])
+        return self.program.global_block().create_parameter(
+            name=attr['name'], dtype=dtype, shape=shape)
 
     def create_tmp_variable(self, dtype):
         return self.program.current_block().create_var(

python/paddle/v2/framework/layers.py

@@ -13,7 +13,8 @@ def fc(input,
        name=None,
        act=None,
        num_flatten_dims=1,
-       program=None):
+       program=None,
+       init_program=None):
     # create helper
     helper = LayerHelper('fc', **locals())
 
@@ -59,7 +60,8 @@ def data(name,
          data_type='float32',
          type=core.VarDesc.VarType.LOD_TENSOR,
          append_batch_size=True,
-         program=None):
+         program=None,
+         init_program=None):
     helper = LayerHelper('data', **locals())
     if append_batch_size:
         shape = [-1] + shape  # append batch size as -1
@@ -160,7 +162,8 @@ def conv2d(input,
            padding=None,
            bias_attr=None,
            param_attr=None,
-           program=None):
+           program=None,
+           init_program=None):
     helper = LayerHelper('conv2d', **locals())
     dtype = helper.input_dtype()
 
@@ -207,7 +210,8 @@ def pool2d(input,
            pool_stride=[1, 1],
            pool_padding=[0, 0],
            global_pooling=False,
-           program=None):
+           program=None,
+           init_program=None):
     if pool_type not in ["max", "avg"]:
         raise ValueError(
             "Unknown pool_type: '%s'. It can only be 'max' or 'avg'.",

python/paddle/v2/framework/tests/test_fit_a_line.py

@@ -8,24 +8,35 @@ from paddle.v2.framework.executor import Executor
 
 import numpy as np
 
+init_program = Program()
 program = Program()
-x = layers.data(name='x', shape=[13], data_type='float32', program=program)
-y_predict = layers.fc(input=x, size=1, act=None, program=program)
+x = layers.data(
+    name='x',
+    shape=[13],
+    data_type='float32',
+    program=program,
+    init_program=init_program)
+y_predict = layers.fc(input=x,
+                      size=1,
+                      act=None,
+                      program=program,
+                      init_program=init_program)
 
-y = layers.data(name='y', shape=[1], data_type='float32', program=program)
+y = layers.data(
+    name='y',
+    shape=[1],
+    data_type='float32',
+    program=program,
+    init_program=init_program)
 
-cost = layers.square_error_cost(input=y_predict, label=y, program=program)
-avg_cost = layers.mean(x=cost, program=program)
+cost = layers.square_error_cost(
+    input=y_predict, label=y, program=program, init_program=init_program)
+avg_cost = layers.mean(x=cost, program=program, init_program=init_program)
 
-sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.01)
+sgd_optimizer = optimizer.SGDOptimizer(learning_rate=0.005)
 opts = sgd_optimizer.minimize(avg_cost)
 
-print str(program)
-
-import pdb
-pdb.set_trace()
-
-BATCH_SIZE = 100
+BATCH_SIZE = 10
 
 train_reader = paddle.batch(
     paddle.reader.shuffle(
@@ -35,12 +46,15 @@ train_reader = paddle.batch(
 place = core.CPUPlace()
 exe = Executor(place)
 
-PASS_NUM = 200
+exe.run(init_program,
+        feed={},
+        fetch_list=[init_program.global_block().var('fc_0.w_1')])
+
+PASS_NUM = 10
 for pass_id in range(PASS_NUM):
     for data in train_reader():
         x_data = np.array(map(lambda x: x[0], data)).astype("float32")
         y_data = np.array(map(lambda x: x[1], data)).astype("float32")
-        #y_data = np.expand_dims(y_data, axis=1)
 
         tensor_x = core.LoDTensor()
         tensor_x.set(x_data, place)
@@ -50,6 +64,11 @@ for pass_id in range(PASS_NUM):
         outs = exe.run(program,
                        feed={'x': tensor_x,
                              'y': tensor_y},
-                       fetch_list=[avg_cost])
+                       fetch_list=[
+                           avg_cost, program.global_block().var('fc_0.w_1'),
+                           program.global_block().var('fc_0.w_1@GRAD')
+                       ])
         out = np.array(outs[0])
+        w = np.array(outs[1])
+        wg = np.array(outs[2])
         print out