07/Label semantic roles (#5798)

* init label_semantic_roles.py

* add linear_chain_crf and test

* complete test_linear_chain_crf

* correct last layer of db_lstm

* update optimizer and initializer

* update param_initializer of embedding_layer

* support load pre trained embedding

* rm unused parameter

* optimize code

* clean code

* fix test

* add todo

Author: jacquesqiao

paddle/operators/linear_chain_crf_op.h

@@ -271,7 +271,7 @@ class LinearChainCRFOpKernel : public framework::OpKernel<T> {
     ll -= std::log(sum);
     // Now ll is equal to -log(Z).
 
-    const int* lbl = label.data<int>();
+    const int64_t* lbl = label.data<int64_t>();
     PADDLE_ENFORCE_LT(
         static_cast<size_t>(*std::max_element(lbl, lbl + seq_length)), tag_num,
         "An invalid tag label that execesses the largest tag number.");
@@ -449,7 +449,7 @@ class LinearChainCRFGradOpKernel : public framework::OpKernel<T> {
                            Tensor* emission_grad) const {
     const T* w_exps = transition_exps.data<T>();
     const T* x_exps = emission_exps.data<T>();
-    const int* label_value = label.data<int>();
+    const int64_t* label_value = label.data<int64_t>();
     T* beta_value = beta->data<T>();
 
     auto x_dims = emission_exps.dims();

python/paddle/v2/fluid/layer_helper.py

@@ -126,7 +126,10 @@ def create_parameter(self, attr, shape, dtype, suffix='w',
         self.startup_program.global_block().create_parameter(
             dtype=dtype, shape=shape, **attr_copy)
         return self.main_program.global_block().create_parameter(
-            name=attr_copy['name'], dtype=dtype, shape=shape)
+            name=attr_copy['name'],
+            dtype=dtype,
+            shape=shape,
+            trainable=attr_copy.get('trainable', True))
 
     def create_tmp_variable(self, dtype):
         return self.main_program.current_block().create_var(

python/paddle/v2/fluid/layers.py

@@ -112,6 +112,7 @@ def _get_default_bias_initializer():
 def embedding(input,
               size,
               is_sparse=False,
+              param_initializer=None,
               param_attr=None,
               data_type='float32',
               main_program=None,
@@ -136,9 +137,16 @@ def embedding(input,
     to the LayerHelper constructor.
 
     """
+
+    def _get_default_param_initializer():
+        return XavierInitializer()
+
     helper = LayerHelper('embedding', **locals())
     w = helper.create_parameter(
-        attr=helper.param_attr, shape=size, dtype=data_type)
+        attr=helper.param_attr,
+        shape=size,
+        dtype=data_type,
+        initializer=param_initializer or _get_default_param_initializer())
     tmp = helper.create_tmp_variable(data_type)
     helper.append_op(
         type='lookup_table',
@@ -460,6 +468,41 @@ def sums(input, main_program=None, startup_program=None):
     return out
 
 
+def linear_chain_crf(input,
+                     label,
+                     param_attr=None,
+                     param_initializer=None,
+                     main_program=None,
+                     startup_program=None):
+    def _get_default_param_initializer():
+        return XavierInitializer()
+
+    helper = LayerHelper('linear_chain_crf', **locals())
+    size = input.shape[1]
+    transition = helper.create_parameter(
+        attr=helper.param_attr,
+        shape=[size + 2, size],
+        dtype=helper.input_dtype(),
+        initializer=param_initializer or _get_default_param_initializer())
+    alpha = helper.create_tmp_variable(dtype=helper.input_dtype())
+    emission_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
+    transition_exps = helper.create_tmp_variable(dtype=helper.input_dtype())
+    log_likelihood = helper.create_tmp_variable(dtype=helper.input_dtype())
+    helper.append_op(
+        type='linear_chain_crf',
+        inputs={"Emission": [input],
+                "Transition": transition,
+                "Label": label},
+        outputs={
+            "Alpha": [alpha],
+            "EmissionExps": [emission_exps],
+            "TransitionExps": transition_exps,
+            "LogLikelihood": log_likelihood
+        })
+
+    return log_likelihood
+
+
 def assign(input, output, main_program=None, startup_program=None):
     helper = LayerHelper('assign', **locals())
     helper.append_op(

python/paddle/v2/fluid/optimizer.py

@@ -170,7 +170,8 @@ def create_optimization_pass(self,
 
         optimize_ops = []
         for param_and_grad in parameters_and_grads:
-            if param_and_grad[1] is not None:
+            if param_and_grad[0].trainable is True and param_and_grad[
+                    1] is not None:
                 optimize_op = self._append_optimize_op(loss.block,
                                                        param_and_grad)
                 optimize_ops.append(optimize_op)

python/paddle/v2/fluid/tests/book/test_label_semantic_roles.py

@@ -0,0 +1,192 @@
+import numpy as np
+import paddle.v2 as paddle
+import paddle.v2.dataset.conll05 as conll05
+import paddle.v2.fluid.core as core
+import paddle.v2.fluid.framework as framework
+import paddle.v2.fluid.layers as layers
+from paddle.v2.fluid.executor import Executor, g_scope
+from paddle.v2.fluid.optimizer import SGDOptimizer
+
+word_dict, verb_dict, label_dict = conll05.get_dict()
+word_dict_len = len(word_dict)
+label_dict_len = len(label_dict)
+pred_len = len(verb_dict)
+
+mark_dict_len = 2
+word_dim = 32
+mark_dim = 5
+hidden_dim = 512
+depth = 8
+mix_hidden_lr = 1e-3
+
+IS_SPARSE = True
+PASS_NUM = 10
+BATCH_SIZE = 20
+
+embedding_name = 'emb'
+
+
+def load_parameter(file_name, h, w):
+    with open(file_name, 'rb') as f:
+        f.read(16)  # skip header.
+        return np.fromfile(f, dtype=np.float32).reshape(h, w)
+
+
+def db_lstm():
+    # 8 features
+    word = layers.data(name='word_data', shape=[1], data_type='int64')
+    predicate = layers.data(name='verb_data', shape=[1], data_type='int64')
+    ctx_n2 = layers.data(name='ctx_n2_data', shape=[1], data_type='int64')
+    ctx_n1 = layers.data(name='ctx_n1_data', shape=[1], data_type='int64')
+    ctx_0 = layers.data(name='ctx_0_data', shape=[1], data_type='int64')
+    ctx_p1 = layers.data(name='ctx_p1_data', shape=[1], data_type='int64')
+    ctx_p2 = layers.data(name='ctx_p2_data', shape=[1], data_type='int64')
+    mark = layers.data(name='mark_data', shape=[1], data_type='int64')
+
+    predicate_embedding = layers.embedding(
+        input=predicate,
+        size=[pred_len, word_dim],
+        data_type='float32',
+        is_sparse=IS_SPARSE,
+        param_attr={'name': 'vemb'})
+
+    mark_embedding = layers.embedding(
+        input=mark,
+        size=[mark_dict_len, mark_dim],
+        data_type='float32',
+        is_sparse=IS_SPARSE)
+
+    word_input = [word, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2]
+    emb_layers = [
+        layers.embedding(
+            size=[word_dict_len, word_dim],
+            input=x,
+            param_attr={'name': embedding_name,
+                        'trainable': False}) for x in word_input
+    ]
+    emb_layers.append(predicate_embedding)
+    emb_layers.append(mark_embedding)
+
+    hidden_0_layers = [
+        layers.fc(input=emb, size=hidden_dim) for emb in emb_layers
+    ]
+
+    hidden_0 = layers.sums(input=hidden_0_layers)
+
+    lstm_0 = layers.dynamic_lstm(
+        input=hidden_0,
+        size=hidden_dim,
+        candidate_activation='relu',
+        gate_activation='sigmoid',
+        cell_activation='sigmoid')
+
+    # stack L-LSTM and R-LSTM with direct edges
+    input_tmp = [hidden_0, lstm_0]
+
+    for i in range(1, depth):
+        mix_hidden = layers.sums(input=[
+            layers.fc(input=input_tmp[0], size=hidden_dim),
+            layers.fc(input=input_tmp[1], size=hidden_dim)
+        ])
+
+        lstm = layers.dynamic_lstm(
+            input=mix_hidden,
+            size=hidden_dim,
+            candidate_activation='relu',
+            gate_activation='sigmoid',
+            cell_activation='sigmoid',
+            is_reverse=((i % 2) == 1))
+
+        input_tmp = [mix_hidden, lstm]
+
+    feature_out = layers.sums(input=[
+        layers.fc(input=input_tmp[0], size=label_dict_len),
+        layers.fc(input=input_tmp[1], size=label_dict_len)
+    ])
+
+    return feature_out
+
+
+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():
+    # define network topology
+    feature_out = db_lstm()
+    target = layers.data(name='target', shape=[1], data_type='int64')
+    crf_cost = layers.linear_chain_crf(
+        input=feature_out,
+        label=target,
+        param_attr={"name": 'crfw',
+                    "learning_rate": mix_hidden_lr})
+    avg_cost = layers.mean(x=crf_cost)
+    # TODO(qiao)
+    #   1. add crf_decode_layer and evaluator
+    #   2. use other optimizer and check why out will be NAN
+    sgd_optimizer = SGDOptimizer(learning_rate=0.0001)
+    opts = sgd_optimizer.minimize(avg_cost)
+
+    train_data = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.conll05.test(), buf_size=8192),
+        batch_size=BATCH_SIZE)
+    place = core.CPUPlace()
+    exe = Executor(place)
+
+    exe.run(framework.default_startup_program())
+
+    embedding_param = g_scope.find_var(embedding_name).get_tensor()
+    embedding_param.set(
+        load_parameter(conll05.get_embedding(), word_dict_len, word_dim), place)
+
+    batch_id = 0
+    for pass_id in xrange(PASS_NUM):
+        for data in train_data():
+            word_data = to_lodtensor(map(lambda x: x[0], data), place)
+            ctx_n2_data = to_lodtensor(map(lambda x: x[1], data), place)
+            ctx_n1_data = to_lodtensor(map(lambda x: x[2], data), place)
+            ctx_0_data = to_lodtensor(map(lambda x: x[3], data), place)
+            ctx_p1_data = to_lodtensor(map(lambda x: x[4], data), place)
+            ctx_p2_data = to_lodtensor(map(lambda x: x[5], data), place)
+            verb_data = to_lodtensor(map(lambda x: x[6], data), place)
+            mark_data = to_lodtensor(map(lambda x: x[7], data), place)
+            target = to_lodtensor(map(lambda x: x[8], data), place)
+
+            outs = exe.run(framework.default_main_program(),
+                           feed={
+                               'word_data': word_data,
+                               'ctx_n2_data': ctx_n2_data,
+                               'ctx_n1_data': ctx_n1_data,
+                               'ctx_0_data': ctx_0_data,
+                               'ctx_p1_data': ctx_p1_data,
+                               'ctx_p2_data': ctx_p2_data,
+                               'verb_data': verb_data,
+                               'mark_data': mark_data,
+                               'target': target
+                           },
+                           fetch_list=[avg_cost])
+            avg_cost_val = np.array(outs[0])
+
+            if batch_id % 10 == 0:
+                print("avg_cost=" + str(avg_cost_val))
+
+            # exit early for CI
+            exit(0)
+
+            batch_id = batch_id + 1
+
+
+if __name__ == '__main__':
+    main()

python/paddle/v2/fluid/tests/test_layers.py

@@ -1,8 +1,8 @@
+import unittest
+
 import paddle.v2.fluid.layers as layers
 import paddle.v2.fluid.nets as nets
 from paddle.v2.fluid.framework import Program
-import paddle.v2.fluid.core as core
-import unittest
 
 
 class TestBook(unittest.TestCase):
@@ -20,7 +20,8 @@ def test_fit_a_line(self):
         avg_cost = layers.mean(x=cost, main_program=program)
         self.assertIsNotNone(avg_cost)
         program.append_backward(avg_cost)
-        print str(program)
+
+        # print str(program)
 
     def test_recognize_digits_mlp(self):
         program = Program()
@@ -49,7 +50,7 @@ def test_recognize_digits_mlp(self):
             input=predict, label=label, main_program=program)
         avg_cost = layers.mean(x=cost, main_program=program)
         self.assertIsNotNone(avg_cost)
-        print str(program)
+        # print str(program)
 
     def test_simple_conv2d(self):
         program = Program()
@@ -64,7 +65,7 @@ def test_simple_conv2d(self):
             filter_size=[4, 4],
             main_program=program)
 
-        print str(program)
+        # print str(program)
 
     def test_recognize_digits_conv(self):
         program = Program()
@@ -103,7 +104,7 @@ def test_recognize_digits_conv(self):
 
         program.append_backward(avg_cost)
 
-        print str(program)
+        # print str(program)
 
     def test_word_embedding(self):
         program = Program()
@@ -164,7 +165,24 @@ def test_word_embedding(self):
         avg_cost = layers.mean(x=cost, main_program=program)
         self.assertIsNotNone(avg_cost)
 
-        print str(program)
+        # print str(program)
+
+    def test_linear_chain_crf(self):
+        program = Program()
+
+        # Change g_program, so the rest layers use `g_program`
+        images = layers.data(
+            name='pixel',
+            shape=[784],
+            data_type='float32',
+            main_program=program)
+        label = layers.data(
+            name='label', shape=[1], data_type='int32', main_program=program)
+        hidden = layers.fc(input=images, size=128, main_program=program)
+        crf = layers.linear_chain_crf(
+            input=hidden, label=label, main_program=program)
+
+        # print str(program)
 
 
 if __name__ == '__main__':

python/paddle/v2/fluid/tests/test_linear_chain_crf_op.py

@@ -104,7 +104,7 @@ def set_test_data(self):
         transition_exps = np.exp(transition)
 
         labels = np.random.randint(
-            low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int32")
+            low=0, high=TAG_NUM, size=(lod[-1][-1], 1), dtype="int64")
 
         self.inputs = {
             "Emission": (emission, lod),