Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into poolmaxpool_with_mask

paddle/gserver/layers/MKLDNNAddtoLayer.cpp

@@ -54,7 +54,6 @@ void MKLDNNAddtoLayer::reshape(
   ow = iw;
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-  printSizeInfo();
 }
 
 void MKLDNNAddtoLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNBatchNormLayer.cpp

@@ -125,7 +125,6 @@ void MKLDNNBatchNormLayer::reshape(
       << "Input channel can not be changed";
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-  printSizeInfo();
 }
 
 void MKLDNNBatchNormLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.cpp

@@ -102,8 +102,6 @@ void MKLDNNConvLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNConvLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNConvLayer.h

@@ -92,7 +92,7 @@ public:
   void printSizeInfo() override {
     MKLDNNLayer::printSizeInfo();
     VLOG(MKLDNN_SIZES) << getName() << ": fh: " << fh_ << ", fw: " << fw_
-                       << ": ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
+                       << ", ph: " << ph_ << ", pw: " << pw_ << ", sh: " << sh_
                        << ", sw: " << sw_ << ", dh: " << dh_ << ", dw: " << dw_;
   }
 

paddle/gserver/layers/MKLDNNFcLayer.cpp

@@ -84,8 +84,6 @@ void MKLDNNFcLayer::reshape(
 
   reshapeOutput(oh, ow);
   resizeOutput(bs, oc);
-
-  printSizeInfo();
 }
 
 void MKLDNNFcLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/gserver/layers/MKLDNNPoolLayer.cpp

@@ -71,8 +71,6 @@ void MKLDNNPoolLayer::reshape(
   reshapeOutput(oh, ow);
 
   resizeOutput(bs, oc * oh * ow);
-
-  printSizeInfo();
 }
 
 void MKLDNNPoolLayer::resetFwd(std::vector<primitive>& pipeline,

paddle/platform/call_once.h

@@ -27,20 +27,22 @@ namespace platform {
 
  This wrap is a hack to avoid this bug.
 */
-template <class Callable, class... Args>
+template <typename Callable, typename... Args>
 inline void call_once(std::once_flag& flag, Callable&& f, Args&&... args) {
   bool good = false;
   std::exception ex;
-  std::call_once(flag, [&]() {
-    try {
-      f(args...);
-      good = true;
-    } catch (const std::exception& e) {
-      ex = e;
-    } catch (...) {
-      ex = std::runtime_error("excption caught in call_once");
-    }
-  });
+  std::call_once(flag,
+                 [&](Args&&... args) {
+                   try {
+                     f(args...);
+                     good = true;
+                   } catch (const std::exception& e) {
+                     ex = e;
+                   } catch (...) {
+                     ex = std::runtime_error("excption caught in call_once");
+                   }
+                 },
+                 args...);
   if (!good) {
     throw std::exception(ex);
   }

python/paddle/v2/framework/layer_helper.py

@@ -4,7 +4,7 @@ import itertools
 from paddle.v2.framework.framework import Variable, g_main_program, \
     g_startup_program, unique_name, Program
 from paddle.v2.framework.initializer import ConstantInitializer, \
-    UniformInitializer
+    UniformInitializer, XavierInitializer
 
 
 class LayerHelper(object):
@@ -61,7 +61,7 @@ class LayerHelper(object):
 
     @property
     def param_attr(self):
-        default = {'name': None, 'initializer': UniformInitializer()}
+        default = {'name': None, 'initializer': XavierInitializer()}
         actual = self.kwargs.get('param_attr', None)
         if actual is None:
             actual = default
@@ -70,10 +70,11 @@ class LayerHelper(object):
                 actual[default_field] = default[default_field]
         return actual
 
+    @property
     def bias_attr(self):
-        default = {'name': None, 'initializer': ConstantInitializer()}
+        default = {'name': None, 'initializer': XavierInitializer()}
         bias_attr = self.kwargs.get('bias_attr', None)
-        if bias_attr is True:
+        if bias_attr is None:
             bias_attr = default
 
         if isinstance(bias_attr, dict):
@@ -166,7 +167,7 @@ class LayerHelper(object):
                 num_flatten_dims = 1
 
         size = list(input_var.shape[num_flatten_dims:])
-        bias_attr = self.bias_attr()
+        bias_attr = self.bias_attr
         if not bias_attr:
             return input_var
 

python/paddle/v2/framework/layers.py

@@ -16,7 +16,7 @@ __all__ = [
 def fc(input,
        size,
        param_attr=None,
-       bias_attr=True,
+       bias_attr=None,
        name=None,
        act=None,
        num_flatten_dims=1,
@@ -125,6 +125,55 @@ def embedding(input,
     return tmp
 
 
+# TODO(qijun): expose H0 and C0
+def dynamic_lstm(input,
+                 size,
+                 data_type='float32',
+                 param_attr=None,
+                 bias_attr=None,
+                 use_peepholes=True,
+                 is_reverse=False,
+                 gate_activation='sigmoid',
+                 cell_activation='tanh',
+                 candidate_activation='tanh',
+                 main_program=None,
+                 startup_program=None):
+    helper = LayerHelper('lstm', **locals())
+    size = size / 4
+    weight = helper.create_parameter(
+        attr=helper.param_attr, shape=[size, 4 * size], dtype=data_type)
+    bias_size = [1, 7 * size]
+    if not use_peepholes:
+        bias_size[1] = 4 * size
+    bias = helper.create_parameter(
+        attr=helper.bias_attr, shape=bias_size, dtype=data_type, suffix='b')
+
+    hidden = helper.create_tmp_variable(data_type)
+    cell = helper.create_tmp_variable(data_type)
+    batch_gate = helper.create_tmp_variable(data_type)
+    batch_cell_pre_act = helper.create_tmp_variable(data_type)
+
+    helper.append_op(
+        type='lstm',
+        inputs={'Input': input,
+                'Weight': weight,
+                'Bias': bias},
+        outputs={
+            'Hidden': hidden,
+            'Cell': cell,
+            'BatchGate': batch_gate,
+            'BatchCellPreAct': batch_cell_pre_act
+        },
+        attrs={
+            'use_peepholes': use_peepholes,
+            'is_reverse': is_reverse,
+            'gate_activation': gate_activation,
+            'cell_activation': cell_activation,
+            'candidate_activation': candidate_activation
+        })
+    return hidden, cell
+
+
 def data(name,
          shape,
          data_type='float32',

python/paddle/v2/framework/optimizer.py

@@ -35,15 +35,21 @@ class Optimizer(object):
         """
         raise NotImplementedError()
 
-    def _initialize_tensors(self, block):
-        """Create all necessary tensors, that will be shared for all parameter updates.
-
-        Tensors like learning rate should be initialized here.
-
-        Args:
-            block: the block in which the loss variable is present
-        """
-        pass
+    def _create_param_lr(self, param_and_grad):
+        # create learning rate variable for every parameter
+        param = param_and_grad[0]
+        param_lr = param.optimize_attr['learning_rate']
+        param_lr_shape = [1]
+        param_lr_var = self.helper.create_global_variable(
+            name=unique_name("learning_rate"),
+            dtype='float32',
+            shape=param_lr_shape,
+            lod_level=1,
+            persistable=True)
+        param_lr = param_lr * self._learning_rate
+        self.helper.set_variable_initializer(
+            var=param_lr_var, initializer=ConstantInitializer(param_lr))
+        return param_lr_var
 
     def _create_accumulators(self, block, parameters):
         """Create all accumulators needed by the parameters
@@ -161,8 +167,6 @@ class Optimizer(object):
             startup_program=startup_program)
         self._create_accumulators(loss.block,
                                   [p[0] for p in parameters_and_grads])
-        # Create any necessary tensors
-        self._initialize_tensors(loss.block)
 
         optimize_ops = []
         for param_and_grad in parameters_and_grads:
@@ -214,27 +218,16 @@ class SGDOptimizer(Optimizer):
         self.type = "sgd"
         self._learning_rate = learning_rate
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _append_optimize_op(self, block, param_and_grad):
         assert isinstance(block, framework.Block)
+
         # create the optimize op
         sgd_op = block.append_op(
             type=self.type,
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={"ParamOut": param_and_grad[0]})
 
@@ -259,19 +252,6 @@ class MomentumOptimizer(Optimizer):
         self._momentum = momentum
         self._use_nesterov = bool(use_nesterov)
 
-    def _initialize_tensors(self, block):
-        assert isinstance(block, framework.Block)
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -290,7 +270,7 @@ class MomentumOptimizer(Optimizer):
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
                 "Velocity": velocity_acc,
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={
                 "ParamOut": param_and_grad[0],
@@ -315,18 +295,6 @@ class AdagradOptimizer(Optimizer):
         self._learning_rate = learning_rate
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -346,7 +314,7 @@ class AdagradOptimizer(Optimizer):
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
                 "Moment": moment_acc,
-                "LearningRate": self._lr
+                "LearningRate": self._create_param_lr(param_and_grad)
             },
             outputs={"ParamOut": param_and_grad[0],
                      "MomentOut": moment_acc},
@@ -378,18 +346,6 @@ class AdamOptimizer(Optimizer):
         self._beta2 = beta2
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         assert isinstance(block, framework.Block)
 
@@ -433,7 +389,7 @@ class AdamOptimizer(Optimizer):
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr,
+                "LearningRate": self._create_param_lr(param_and_grad),
                 "Moment1": moment1,
                 "Moment2": moment2,
                 "Beta1Pow": self._beta1_pow_acc,
@@ -495,18 +451,6 @@ class AdamaxOptimizer(Optimizer):
         self._beta2 = beta2
         self._epsilon = epsilon
 
-    def _initialize_tensors(self, block):
-        lr_shape = [1]
-        # create a variable for learning_rate
-        self._lr = self.helper.create_global_variable(
-            name=unique_name("learning_rate"),
-            dtype='float32',
-            shape=lr_shape,
-            lod_level=1,
-            persistable=True)
-        self.helper.set_variable_initializer(
-            var=self._lr, initializer=ConstantInitializer(self._learning_rate))
-
     def _create_accumulators(self, block, parameters):
         # Create beta1 power accumulator tensor
         beta_shape = [1]
@@ -536,7 +480,7 @@ class AdamaxOptimizer(Optimizer):
             inputs={
                 "Param": param_and_grad[0],
                 "Grad": param_and_grad[1],
-                "LearningRate": self._lr,
+                "LearningRate": self._create_param_lr(param_and_grad),
                 "Moment": moment,
                 "InfNorm": inf_norm,
                 "Beta1Pow": self._beta1_pow_acc

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

+import paddle.v2 as paddle
+import paddle.v2.framework.layers as layers
+import paddle.v2.framework.nets as nets
+import paddle.v2.framework.core as core
+import paddle.v2.framework.optimizer as optimizer
+
+from paddle.v2.framework.framework import Program, g_main_program, g_startup_program
+from paddle.v2.framework.executor import Executor
+
+import numpy as np
+
+
+def stacked_lstm_net(input_dim,
+                     class_dim=2,
+                     emb_dim=128,
+                     hid_dim=512,
+                     stacked_num=3):
+    assert stacked_num % 2 == 1
+    data = layers.data(name="words", shape=[1], data_type="int64")
+    label = layers.data(name="label", shape=[1], data_type="int64")
+
+    emb = layers.embedding(input=data, size=[input_dim, emb_dim])
+    # add bias attr
+
+    # TODO(qijun) linear act
+    fc1 = layers.fc(input=emb, size=hid_dim)
+    lstm1, cell1 = layers.dynamic_lstm(input=fc1, size=hid_dim)
+
+    inputs = [fc1, lstm1]
+
+    for i in range(2, stacked_num + 1):
+        fc = layers.fc(input=inputs, size=hid_dim)
+        lstm, cell = layers.dynamic_lstm(
+            input=fc, size=hid_dim, is_reverse=(i % 2) == 0)
+        inputs = [fc, lstm]
+
+    fc_last = layers.sequence_pool(input=inputs[0], pool_type='max')
+    lstm_last = layers.sequence_pool(input=inputs[1], pool_type='max')
+
+    prediction = layers.fc(input=[fc_last, lstm_last],
+                           size=class_dim,
+                           act='softmax')
+    cost = layers.cross_entropy(input=prediction, label=label)
+    avg_cost = layers.mean(x=cost)
+    adam_optimizer = optimizer.AdamOptimizer(learning_rate=0.002)
+    opts = adam_optimizer.minimize(avg_cost)
+    acc = layers.accuracy(input=prediction, label=label)
+    return avg_cost, acc
+
+
+def to_lodtensor(data, place):
+    seq_lens = [len(seq) for seq in data]
+    cur_len = 0
+    lod = [cur_len]
+    for l in seq_lens:
+        cur_len += l
+        lod.append(cur_len)
+    flattened_data = np.concatenate(data, axis=0).astype("int64")
+    flattened_data = flattened_data.reshape([len(flattened_data), 1])
+    res = core.LoDTensor()
+    res.set(flattened_data, place)
+    res.set_lod([lod])
+    return res
+
+
+def main():
+    BATCH_SIZE = 100
+    PASS_NUM = 5
+
+    word_dict = paddle.dataset.imdb.word_dict()
+    print "load word dict successfully"
+    dict_dim = len(word_dict)
+    class_dim = 2
+
+    cost, acc = stacked_lstm_net(input_dim=dict_dim, class_dim=class_dim)
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.imdb.train(word_dict), buf_size=1000),
+        batch_size=BATCH_SIZE)
+    place = core.CPUPlace()
+    exe = Executor(place)
+
+    exe.run(g_startup_program)
+
+    for pass_id in xrange(PASS_NUM):
+        for data in train_data():
+            tensor_words = to_lodtensor(map(lambda x: x[0], data), place)
+
+            label = np.array(map(lambda x: x[1], data)).astype("int64")
+            label = label.reshape([BATCH_SIZE, 1])
+
+            tensor_label = core.LoDTensor()
+            tensor_label.set(label, place)
+
+            outs = exe.run(g_main_program,
+                           feed={"words": tensor_words,
+                                 "label": tensor_label},
+                           fetch_list=[cost, acc])
+            cost_val = np.array(outs[0])
+            acc_val = np.array(outs[1])
+
+            print("cost=" + str(cost_val) + " acc=" + str(acc_val))
+            if cost_val < 1.0 and acc_val > 0.7:
+                exit(0)
+    exit(1)
+
+
+if __name__ == '__main__':
+    main()