From ed262b804ce4fa12e82dd0f1f7a3dec1c7ce23f7 Mon Sep 17 00:00:00 2001
From: Aurelius84 <zhangliujie@baidu.com>
Date: Thu, 2 Jul 2020 10:19:36 +0800
Subject: [PATCH] [Dy2stat] Add Sentiment and LAC model for unittest (#25071)

* add sentiment unittest

* add LAC model

* add test=develop

* rename tmp test=develop

* fix timeout test=develop

* undo tmp_var test=develop
---
 .../unittests/dygraph_to_static/test_lac.py   | 573 ++++++++++++++++++
 .../dygraph_to_static/test_sentiment.py       | 355 +++++++++++
 2 files changed, 928 insertions(+)
 create mode 100644 python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
 create mode 100644 python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py

diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
new file mode 100644
index 0000000000..473a919ad9
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_lac.py
@@ -0,0 +1,573 @@
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import print_function
+import math
+import time
+import numpy as np
+import unittest
+
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "2"
+
+import paddle.fluid as fluid
+from paddle.fluid.dygraph import to_variable
+from paddle.fluid.dygraph import Embedding, Linear, GRUUnit
+from paddle.fluid.dygraph import declarative, ProgramTranslator
+
+SEED = 2020
+
+program_translator = ProgramTranslator()
+
+
+class DynamicGRU(fluid.dygraph.Layer):
+    def __init__(self,
+                 size,
+                 h_0=None,
+                 param_attr=None,
+                 bias_attr=None,
+                 is_reverse=False,
+                 gate_activation='sigmoid',
+                 candidate_activation='tanh',
+                 origin_mode=False,
+                 init_size=None):
+        super(DynamicGRU, self).__init__()
+
+        self.gru_unit = GRUUnit(
+            size * 3,
+            param_attr=param_attr,
+            bias_attr=bias_attr,
+            activation=candidate_activation,
+            gate_activation=gate_activation,
+            origin_mode=origin_mode)
+
+        self.size = size
+        self.h_0 = h_0
+        self.is_reverse = is_reverse
+
+    def forward(self, inputs):
+        # Use `to_variable` to create a copy of global h_0 created not in `DynamicGRU`,
+        # to avoid modify it because `h_0` is both used in other `DynamicGRU`.
+        hidden = to_variable(self.h_0)
+        hidden.stop_gradient = True
+
+        res = []
+        for i in range(inputs.shape[1]):
+            if self.is_reverse:
+                j = fluid.layers.shape(inputs)[1] - 1 - i
+            else:
+                # TODO(Aurelius84): In while block, if the var created in parent block
+                # participates in the calculation of gradient, the result of gradient
+                # is incorrect because each step scope always returns the same value
+                # generated by last step. Here we add 0 to create `j` in while block to
+                # avoid this bug, and working on fixing it in next PR.
+                j = i + 0
+            # FIXME(Aurelius84): see above explanation.
+            hidden = fluid.layers.scale(hidden, 1)
+
+            # See above explanation.
+            # input_ = inputs[:, i:i+1, :]  # original code
+            input_ = fluid.layers.slice(
+                inputs, axes=[1], starts=[j], ends=[j + 1])
+            input_ = fluid.layers.reshape(
+                input_, [-1, input_.shape[2]], inplace=False)
+            hidden, reset, gate = self.gru_unit(input_, hidden)
+            hidden_ = fluid.layers.reshape(
+                hidden, [-1, 1, hidden.shape[1]], inplace=False)
+            res.append(hidden_)
+
+        if self.is_reverse:
+            res = res[::-1]
+        res = fluid.layers.concat(res, axis=1)
+        return res
+
+
+class BiGRU(fluid.dygraph.Layer):
+    def __init__(self, input_dim, grnn_hidden_dim, init_bound, h_0=None):
+        super(BiGRU, self).__init__()
+
+        self.pre_gru = Linear(
+            input_dim=input_dim,
+            output_dim=grnn_hidden_dim * 3,
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Uniform(
+                    low=-init_bound, high=init_bound),
+                regularizer=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=1e-4)))
+
+        self.gru = DynamicGRU(
+            size=grnn_hidden_dim,
+            h_0=h_0,
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Uniform(
+                    low=-init_bound, high=init_bound),
+                regularizer=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=1e-4)))
+
+        self.pre_gru_r = Linear(
+            input_dim=input_dim,
+            output_dim=grnn_hidden_dim * 3,
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Uniform(
+                    low=-init_bound, high=init_bound),
+                regularizer=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=1e-4)))
+
+        self.gru_r = DynamicGRU(
+            size=grnn_hidden_dim,
+            is_reverse=True,
+            h_0=h_0,
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Uniform(
+                    low=-init_bound, high=init_bound),
+                regularizer=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=1e-4)))
+
+    def forward(self, input_feature):
+        res_pre_gru = self.pre_gru(input_feature)
+        res_gru = self.gru(res_pre_gru)
+
+        res_pre_gru_r = self.pre_gru_r(input_feature)
+        res_gru_r = self.gru_r(res_pre_gru_r)
+
+        bi_merge = fluid.layers.concat(input=[res_gru, res_gru_r], axis=-1)
+        return bi_merge
+
+
+class LinearChainCRF(fluid.dygraph.Layer):
+    def __init__(self, param_attr, size=None, is_test=False, dtype='float32'):
+        super(LinearChainCRF, self).__init__()
+
+        self._param_attr = param_attr
+        self._dtype = dtype
+        self._size = size
+        self._is_test = is_test
+        self._transition = self.create_parameter(
+            attr=self._param_attr,
+            shape=[self._size + 2, self._size],
+            dtype=self._dtype)
+
+    @property
+    def weight(self):
+        return self._transition
+
+    @weight.setter
+    def weight(self, value):
+        self._transition = value
+
+    def forward(self, input, label, length=None):
+
+        alpha = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype)
+        emission_exps = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype)
+        transition_exps = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype)
+        log_likelihood = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype)
+        this_inputs = {
+            "Emission": [input],
+            "Transition": self._transition,
+            "Label": [label]
+        }
+        if length is not None:
+            this_inputs['Length'] = [length]
+        self._helper.append_op(
+            type='linear_chain_crf',
+            inputs=this_inputs,
+            outputs={
+                "Alpha": [alpha],
+                "EmissionExps": [emission_exps],
+                "TransitionExps": transition_exps,
+                "LogLikelihood": log_likelihood
+            },
+            attrs={"is_test": self._is_test, })
+        return log_likelihood
+
+
+class CRFDecoding(fluid.dygraph.Layer):
+    def __init__(self, param_attr, size=None, is_test=False, dtype='float32'):
+        super(CRFDecoding, self).__init__()
+
+        self._dtype = dtype
+        self._size = size
+        self._is_test = is_test
+        self._param_attr = param_attr
+        self._transition = self.create_parameter(
+            attr=self._param_attr,
+            shape=[self._size + 2, self._size],
+            dtype=self._dtype)
+
+    @property
+    def weight(self):
+        return self._transition
+
+    @weight.setter
+    def weight(self, value):
+        self._transition = value
+
+    def forward(self, input, label=None, length=None):
+
+        viterbi_path = self._helper.create_variable_for_type_inference(
+            dtype=self._dtype)
+        this_inputs = {
+            "Emission": [input],
+            "Transition": self._transition,
+            "Label": label
+        }
+        if length is not None:
+            this_inputs['Length'] = [length]
+        self._helper.append_op(
+            type='crf_decoding',
+            inputs=this_inputs,
+            outputs={"ViterbiPath": [viterbi_path]},
+            attrs={"is_test": self._is_test, })
+        return viterbi_path
+
+
+class ChunkEval(fluid.dygraph.Layer):
+    def __init__(self, num_chunk_types, chunk_scheme,
+                 excluded_chunk_types=None):
+        super(ChunkEval, self).__init__()
+        self.num_chunk_types = num_chunk_types
+        self.chunk_scheme = chunk_scheme
+        self.excluded_chunk_types = excluded_chunk_types
+
+    def forward(self, input, label, seq_length=None):
+
+        precision = self._helper.create_variable_for_type_inference(
+            dtype="float32")
+        recall = self._helper.create_variable_for_type_inference(
+            dtype="float32")
+        f1_score = self._helper.create_variable_for_type_inference(
+            dtype="float32")
+        num_infer_chunks = self._helper.create_variable_for_type_inference(
+            dtype="int64")
+        num_label_chunks = self._helper.create_variable_for_type_inference(
+            dtype="int64")
+        num_correct_chunks = self._helper.create_variable_for_type_inference(
+            dtype="int64")
+
+        this_input = {"Inference": [input], "Label": [label]}
+        if seq_length is not None:
+            this_input["SeqLength"] = [seq_length]
+
+        self._helper.append_op(
+            type='chunk_eval',
+            inputs=this_input,
+            outputs={
+                "Precision": [precision],
+                "Recall": [recall],
+                "F1-Score": [f1_score],
+                "NumInferChunks": [num_infer_chunks],
+                "NumLabelChunks": [num_label_chunks],
+                "NumCorrectChunks": [num_correct_chunks]
+            },
+            attrs={
+                "num_chunk_types": self.num_chunk_types,
+                "chunk_scheme": self.chunk_scheme,
+                "excluded_chunk_types": self.excluded_chunk_types or []
+            })
+        return (precision, recall, f1_score, num_infer_chunks, num_label_chunks,
+                num_correct_chunks)
+
+
+class LexNet(fluid.dygraph.Layer):
+    def __init__(self, args, length=None):
+        super(LexNet, self).__init__()
+        """
+        define the lexical analysis network structure
+        word: stores the input of the model
+        for_infer: a boolean value, indicating if the model to be created is for training or predicting.
+
+        return:
+            for infer: return the prediction
+            otherwise: return the prediction
+        """
+        self.word_emb_dim = args.word_emb_dim
+        self.vocab_size = args.vocab_size
+        self.num_labels = args.num_labels
+        self.grnn_hidden_dim = args.grnn_hidden_dim
+        self.emb_lr = args.emb_learning_rate if 'emb_learning_rate' in dir(
+            args) else 1.0
+        self.crf_lr = args.emb_learning_rate if 'crf_learning_rate' in dir(
+            args) else 1.0
+        self.bigru_num = args.bigru_num
+        self.init_bound = 0.1
+
+        self.word_embedding = Embedding(
+            size=[self.vocab_size, self.word_emb_dim],
+            dtype='float32',
+            param_attr=fluid.ParamAttr(
+                learning_rate=self.emb_lr,
+                name="word_emb",
+                initializer=fluid.initializer.Uniform(
+                    low=-self.init_bound, high=self.init_bound)))
+
+        h_0 = np.zeros((args.batch_size, self.grnn_hidden_dim), dtype="float32")
+        h_0 = to_variable(h_0)
+
+        self.bigru_units = []
+        for i in range(self.bigru_num):
+            if i == 0:
+                self.bigru_units.append(
+                    self.add_sublayer(
+                        "bigru_units%d" % i,
+                        BiGRU(
+                            self.grnn_hidden_dim,
+                            self.grnn_hidden_dim,
+                            self.init_bound,
+                            h_0=h_0)))
+            else:
+                self.bigru_units.append(
+                    self.add_sublayer(
+                        "bigru_units%d" % i,
+                        BiGRU(
+                            self.grnn_hidden_dim * 2,
+                            self.grnn_hidden_dim,
+                            self.init_bound,
+                            h_0=h_0)))
+
+        self.fc = Linear(
+            input_dim=self.grnn_hidden_dim * 2,
+            output_dim=self.num_labels,
+            param_attr=fluid.ParamAttr(
+                initializer=fluid.initializer.Uniform(
+                    low=-self.init_bound, high=self.init_bound),
+                regularizer=fluid.regularizer.L2DecayRegularizer(
+                    regularization_coeff=1e-4)))
+
+        self.linear_chain_crf = LinearChainCRF(
+            param_attr=fluid.ParamAttr(
+                name='linear_chain_crfw', learning_rate=self.crf_lr),
+            size=self.num_labels)
+
+        self.crf_decoding = CRFDecoding(
+            param_attr=fluid.ParamAttr(
+                name='crfw', learning_rate=self.crf_lr),
+            size=self.num_labels)
+        # share weight
+        self.crf_decoding.weight = self.linear_chain_crf.weight
+
+    @declarative
+    def forward(self, word, target, length=None):
+        """
+        Configure the network
+        """
+        word_embed = self.word_embedding(word)
+        input_feature = word_embed
+
+        for i in range(self.bigru_num):
+            bigru_output = self.bigru_units[i](input_feature)
+            input_feature = bigru_output
+
+        emission = self.fc(bigru_output)
+
+        crf_cost = self.linear_chain_crf(
+            input=emission, label=target, length=length)
+        avg_cost = fluid.layers.mean(x=crf_cost)
+        crf_decode = self.crf_decoding(input=emission, length=length)
+        return avg_cost, crf_decode
+
+
+class Args(object):
+    epoch = 1
+    batch_size = 4
+    vocab_size = 100
+    num_labels = 10
+    word_emb_dim = 128
+    grnn_hidden_dim = 128
+    base_learning_rate = 0.01
+    bigru_num = 2
+    print_steps = 1
+    model_save_dir = "./lac_model"
+    dy_param_path = "./lac_dy_param"
+
+
+def get_random_input_data(batch_size, vocab_size, num_labels, max_seq_len=64):
+    local_random = np.random.RandomState(SEED)
+    padding_id = np.int64(0)
+    iter_num = 5
+
+    def __reader__():
+        batch, init_lens = [], []
+        for i in range(iter_num * batch_size):
+            cur_len = local_random.randint(3, max_seq_len)
+            word_ids = local_random.randint(0, vocab_size,
+                                            [cur_len]).astype('int64').tolist()
+            label_ids = local_random.randint(0, num_labels,
+                                             [cur_len]).astype('int64').tolist()
+            batch.append((word_ids, label_ids))
+            init_lens.append(cur_len)
+            if len(batch) == batch_size:
+                batch_max_len = min(max(init_lens), max_seq_len)
+                new_batch = []
+                for words_len, (word_ids, label_ids) in zip(init_lens, batch):
+                    word_ids = word_ids[0:batch_max_len]
+                    words_len = np.int64(len(word_ids))
+                    word_ids += [
+                        padding_id for _ in range(batch_max_len - words_len)
+                    ]
+                    label_ids = label_ids[0:batch_max_len]
+                    label_ids += [
+                        padding_id for _ in range(batch_max_len - words_len)
+                    ]
+                    assert len(word_ids) == len(label_ids)
+                    new_batch.append((word_ids, label_ids, words_len))
+                yield new_batch
+                batch, init_lens = [], []
+
+    return __reader__
+
+
+def create_dataloader(reader, place):
+    data_loader = fluid.io.DataLoader.from_generator(
+        capacity=16, use_double_buffer=True, iterable=True)
+
+    data_loader.set_sample_list_generator(reader, places=place)
+
+    return data_loader
+
+
+def do_train(args, to_static):
+    program_translator.enable(to_static)
+    place = fluid.CUDAPlace(0) if fluid.is_compiled_with_cuda(
+    ) else fluid.CPUPlace()
+    with fluid.dygraph.guard(place):
+        fluid.default_startup_program().random_seed = SEED
+        fluid.default_main_program().random_seed = SEED
+
+        reader = get_random_input_data(args.batch_size, args.vocab_size,
+                                       args.num_labels)
+        train_loader = create_dataloader(reader, place)
+
+        model = LexNet(args)
+        optimizer = fluid.optimizer.AdamOptimizer(
+            learning_rate=args.base_learning_rate,
+            parameter_list=model.parameters())
+        chunk_eval = ChunkEval(
+            int(math.ceil((args.num_labels - 1) / 2.0)), "IOB")
+
+        step = 0
+        chunk_evaluator = fluid.metrics.ChunkEvaluator()
+        chunk_evaluator.reset()
+
+        loss_data = []
+        for epoch_id in range(args.epoch):
+            for batch in train_loader():
+                words, targets, length = batch
+                start_time = time.time()
+                avg_cost, crf_decode = model(words, targets, length)
+                loss_data.append(avg_cost.numpy()[0])
+
+                # backward and optimization
+                avg_cost.backward()
+                optimizer.minimize(avg_cost)
+                model.clear_gradients()
+                end_time = time.time()
+
+                if step % args.print_steps == 0:
+                    (precision, recall, f1_score, num_infer_chunks,
+                     num_label_chunks, num_correct_chunks) = chunk_eval(
+                         input=crf_decode, label=targets, seq_length=length)
+                    outputs = [avg_cost, precision, recall, f1_score]
+                    avg_cost, precision, recall, f1_score = [
+                        np.mean(x.numpy()) for x in outputs
+                    ]
+
+                    print(
+                        "[train] step = %d, loss = %f, P: %f, R: %f, F1: %f, elapsed time %f"
+                        % (step, avg_cost, precision, recall, f1_score,
+                           end_time - start_time))
+
+                step += 1
+        # save inference model
+        if to_static:
+            program_translator.save_inference_model(
+                dirname=args.model_save_dir, feed=[0, 2], fetch=[1])
+        else:
+            fluid.dygraph.save_dygraph(model.state_dict(), args.dy_param_path)
+
+        return np.array(loss_data)
+
+
+class TestLACModel(unittest.TestCase):
+    def setUp(self):
+        self.args = Args()
+        self.place = fluid.CUDAPlace(0) if fluid.is_compiled_with_cuda(
+        ) else fluid.CPUPlace()
+
+    def train(self, to_static):
+        out = do_train(self.args, to_static)
+        return out
+
+    def test_train(self):
+        dy_out = self.train(to_static=False)
+        st_out = self.train(to_static=True)
+        self.assertTrue(
+            np.allclose(dy_out, st_out),
+            msg="dygraph output:\n{},\nstatic output:\n {}.".format(dy_out,
+                                                                    st_out))
+        # Prediction needs trained models, so put `test_predict` at last of `test_train`
+        self.verify_predict()
+
+    def verify_predict(self):
+        reader = get_random_input_data(
+            self.args.batch_size, self.args.vocab_size, self.args.num_labels)
+        for batch in reader():
+            batch = [np.vstack(var) for var in zip(*batch)]
+            dy_pre = self.predict_dygraph(batch)
+            st_pre = self.predict_static(batch)
+            self.assertTrue(
+                np.allclose(dy_pre, st_pre),
+                msg="dy_pre:\n {}\n, st_pre: \n{}.".format(dy_pre, st_pre))
+
+    def predict_dygraph(self, batch):
+        words, targets, length = batch
+        program_translator.enable(False)
+        with fluid.dygraph.guard(self.place):
+            model = LexNet(self.args)
+            # load dygraph trained parameters
+            model_dict, _ = fluid.load_dygraph(self.args.dy_param_path +
+                                               ".pdparams")
+            model.set_dict(model_dict)
+            model.eval()
+
+            _, pred_res = model(
+                to_variable(words), to_variable(targets), to_variable(length))
+
+            return pred_res.numpy()
+
+    def predict_static(self, batch):
+        """
+        LAC model contains h_0 created in `__init__` that is necessary for inferring.
+        Load inference model to test it's ok for prediction.
+        """
+        exe = fluid.Executor(self.place)
+        # load inference model
+        [inference_program, feed_target_names,
+         fetch_targets] = fluid.io.load_inference_model(
+             self.args.model_save_dir, executor=exe)
+
+        words, targets, length = batch
+        pred_res = exe.run(
+            inference_program,
+            feed={feed_target_names[0]: words,
+                  feed_target_names[1]: length},
+            fetch_list=fetch_targets)
+        return pred_res[0]
+
+
+if __name__ == "__main__":
+    unittest.main()
diff --git a/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py
new file mode 100644
index 0000000000..9d73ff1405
--- /dev/null
+++ b/python/paddle/fluid/tests/unittests/dygraph_to_static/test_sentiment.py
@@ -0,0 +1,355 @@
+# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import time
+import unittest
+import numpy as np
+
+import paddle.fluid as fluid
+from paddle.fluid.dygraph.nn import Conv2D, Linear, Embedding
+from paddle.fluid.dygraph import to_variable, ProgramTranslator, declarative
+
+from test_lac import DynamicGRU
+
+SEED = 2020
+program_translator = ProgramTranslator()
+
+# Note: Set True to eliminate randomness.
+#     1. For one operation, cuDNN has several algorithms,
+#        some algorithm results are non-deterministic, like convolution algorithms.
+if fluid.is_compiled_with_cuda():
+    fluid.set_flags({'FLAGS_cudnn_deterministic': True})
+
+
+class SimpleConvPool(fluid.dygraph.Layer):
+    def __init__(self,
+                 num_channels,
+                 num_filters,
+                 filter_size,
+                 use_cudnn=True,
+                 batch_size=None):
+        super(SimpleConvPool, self).__init__()
+        self.batch_size = batch_size
+        self._conv2d = Conv2D(
+            num_channels=num_channels,
+            num_filters=num_filters,
+            filter_size=filter_size,
+            padding=[1, 1],
+            use_cudnn=use_cudnn,
+            act='tanh')
+
+    def forward(self, inputs):
+        x = self._conv2d(inputs)
+        x = fluid.layers.reduce_max(x, dim=-1)
+        x = fluid.layers.reshape(x, shape=[self.batch_size, -1])
+        return x
+
+
+class CNN(fluid.dygraph.Layer):
+    def __init__(self, dict_dim, batch_size, seq_len):
+        super(CNN, self).__init__()
+        self.dict_dim = dict_dim
+        self.emb_dim = 128
+        self.hid_dim = 128
+        self.fc_hid_dim = 96
+        self.class_dim = 2
+        self.channels = 1
+        self.win_size = [3, self.hid_dim]
+        self.batch_size = batch_size
+        self.seq_len = seq_len
+        self.embedding = Embedding(
+            size=[self.dict_dim + 1, self.emb_dim],
+            dtype='float32',
+            is_sparse=False)
+        self._simple_conv_pool_1 = SimpleConvPool(
+            self.channels,
+            self.hid_dim,
+            self.win_size,
+            batch_size=self.batch_size)
+        self._fc1 = Linear(
+            input_dim=self.hid_dim * self.seq_len,
+            output_dim=self.fc_hid_dim,
+            act="softmax")
+        self._fc_prediction = Linear(
+            input_dim=self.fc_hid_dim, output_dim=self.class_dim, act="softmax")
+
+    @declarative
+    def forward(self, inputs, label=None):
+        emb = self.embedding(inputs)
+        o_np_mask = (
+            fluid.layers.reshape(inputs, [-1, 1]) != self.dict_dim).astype(
+                dtype='float32')
+        mask_emb = fluid.layers.expand(o_np_mask, [1, self.hid_dim])
+        emb = emb * mask_emb
+        emb = fluid.layers.reshape(
+            emb, shape=[-1, self.channels, self.seq_len, self.hid_dim])
+        conv_3 = self._simple_conv_pool_1(emb)
+        fc_1 = self._fc1(conv_3)
+        prediction = self._fc_prediction(fc_1)
+
+        cost = fluid.layers.cross_entropy(input=prediction, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=prediction, label=label)
+        return avg_cost, prediction, acc
+
+
+class BOW(fluid.dygraph.Layer):
+    def __init__(self, dict_dim, batch_size, seq_len):
+        super(BOW, self).__init__()
+        self.dict_dim = dict_dim
+        self.emb_dim = 128
+        self.hid_dim = 128
+        self.fc_hid_dim = 96
+        self.class_dim = 2
+        self.batch_size = batch_size
+        self.seq_len = seq_len
+        self.embedding = Embedding(
+            size=[self.dict_dim + 1, self.emb_dim],
+            dtype='float32',
+            is_sparse=False)
+        self._fc1 = Linear(
+            input_dim=self.hid_dim, output_dim=self.hid_dim, act="tanh")
+        self._fc2 = Linear(
+            input_dim=self.hid_dim, output_dim=self.fc_hid_dim, act="tanh")
+        self._fc_prediction = Linear(
+            input_dim=self.fc_hid_dim, output_dim=self.class_dim, act="softmax")
+
+    @declarative
+    def forward(self, inputs, label=None):
+        emb = self.embedding(inputs)
+        o_np_mask = (
+            fluid.layers.reshape(inputs, [-1, 1]) != self.dict_dim).astype(
+                dtype='float32')
+        mask_emb = fluid.layers.expand(o_np_mask, [1, self.hid_dim])
+        emb = emb * mask_emb
+        emb = fluid.layers.reshape(emb, shape=[-1, self.seq_len, self.hid_dim])
+        bow_1 = fluid.layers.reduce_sum(emb, dim=1)
+        bow_1 = fluid.layers.tanh(bow_1)
+        fc_1 = self._fc1(bow_1)
+        fc_2 = self._fc2(fc_1)
+        prediction = self._fc_prediction(fc_2)
+
+        cost = fluid.layers.cross_entropy(input=prediction, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=prediction, label=label)
+        return avg_cost, prediction, acc
+
+
+class GRU(fluid.dygraph.Layer):
+    def __init__(self, dict_dim, batch_size, seq_len):
+        super(GRU, self).__init__()
+        self.dict_dim = dict_dim
+        self.emb_dim = 128
+        self.hid_dim = 128
+        self.fc_hid_dim = 96
+        self.class_dim = 2
+        self.batch_size = batch_size
+        self.seq_len = seq_len
+        self.embedding = Embedding(
+            size=[self.dict_dim + 1, self.emb_dim],
+            dtype='float32',
+            param_attr=fluid.ParamAttr(learning_rate=30),
+            is_sparse=False)
+        h_0 = np.zeros((self.batch_size, self.hid_dim), dtype="float32")
+        h_0 = to_variable(h_0)
+        self._fc1 = Linear(input_dim=self.hid_dim, output_dim=self.hid_dim * 3)
+        self._fc2 = Linear(
+            input_dim=self.hid_dim, output_dim=self.fc_hid_dim, act="tanh")
+        self._fc_prediction = Linear(
+            input_dim=self.fc_hid_dim, output_dim=self.class_dim, act="softmax")
+        self._gru = DynamicGRU(size=self.hid_dim, h_0=h_0)
+
+    @declarative
+    def forward(self, inputs, label=None):
+        emb = self.embedding(inputs)
+        o_np_mask = (fluid.layers.reshape(inputs, [-1, 1]) != self.dict_dim
+                     ).astype('float32')
+        mask_emb = fluid.layers.expand(o_np_mask, [1, self.hid_dim])
+        emb = emb * mask_emb
+        emb = fluid.layers.reshape(
+            emb, shape=[self.batch_size, -1, self.hid_dim])
+        fc_1 = self._fc1(emb)
+        gru_hidden = self._gru(fc_1)
+        gru_hidden = fluid.layers.reduce_max(gru_hidden, dim=1)
+        tanh_1 = fluid.layers.tanh(gru_hidden)
+        fc_2 = self._fc2(tanh_1)
+        prediction = self._fc_prediction(fc_2)
+
+        cost = fluid.layers.cross_entropy(input=prediction, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=prediction, label=label)
+        return avg_cost, prediction, acc
+
+
+class BiGRU(fluid.dygraph.Layer):
+    def __init__(self, dict_dim, batch_size, seq_len):
+        super(BiGRU, self).__init__()
+        self.dict_dim = dict_dim
+        self.emb_dim = 128
+        self.hid_dim = 128
+        self.fc_hid_dim = 96
+        self.class_dim = 2
+        self.batch_size = batch_size
+        self.seq_len = seq_len
+        self.embedding = Embedding(
+            size=[self.dict_dim + 1, self.emb_dim],
+            dtype='float32',
+            param_attr=fluid.ParamAttr(learning_rate=30),
+            is_sparse=False)
+        h_0 = np.zeros((self.batch_size, self.hid_dim), dtype="float32")
+        h_0 = to_variable(h_0)
+        self._fc1 = Linear(input_dim=self.hid_dim, output_dim=self.hid_dim * 3)
+        self._fc2 = Linear(
+            input_dim=self.hid_dim * 2, output_dim=self.fc_hid_dim, act="tanh")
+        self._fc_prediction = Linear(
+            input_dim=self.fc_hid_dim, output_dim=self.class_dim, act="softmax")
+        self._gru_forward = DynamicGRU(
+            size=self.hid_dim, h_0=h_0, is_reverse=False)
+        self._gru_backward = DynamicGRU(
+            size=self.hid_dim, h_0=h_0, is_reverse=True)
+
+    @declarative
+    def forward(self, inputs, label=None):
+        emb = self.embedding(inputs)
+        o_np_mask = (fluid.layers.reshape(inputs, [-1, 1]) != self.dict_dim
+                     ).astype('float32')
+        mask_emb = fluid.layers.expand(o_np_mask, [1, self.hid_dim])
+        emb = emb * mask_emb
+        emb = fluid.layers.reshape(
+            emb, shape=[self.batch_size, -1, self.hid_dim])
+        fc_1 = self._fc1(emb)
+        gru_forward = self._gru_forward(fc_1)
+        gru_backward = self._gru_backward(fc_1)
+        gru_forward_tanh = fluid.layers.tanh(gru_forward)
+        gru_backward_tanh = fluid.layers.tanh(gru_backward)
+        encoded_vector = fluid.layers.concat(
+            input=[gru_forward_tanh, gru_backward_tanh], axis=2)
+        encoded_vector = fluid.layers.reduce_max(encoded_vector, dim=1)
+        fc_2 = self._fc2(encoded_vector)
+        prediction = self._fc_prediction(fc_2)
+        # TODO(Aurelius84): Uncomment the following codes when we support return variable-length vars.
+        # if label is not None:
+        cost = fluid.layers.cross_entropy(input=prediction, label=label)
+        avg_cost = fluid.layers.mean(x=cost)
+        acc = fluid.layers.accuracy(input=prediction, label=label)
+        return avg_cost, prediction, acc
+        # else:
+        #     return prediction
+
+
+def fake_data_reader(class_num, vocab_size, batch_size, padding_size):
+    def reader():
+        batch_data = []
+        while True:
+            label = np.random.randint(0, class_num)
+            seq_len = np.random.randint(padding_size // 2,
+                                        int(padding_size * 1.2))
+            word_ids = np.random.randint(0, vocab_size, [seq_len]).tolist()
+            word_ids = word_ids[:padding_size] + [vocab_size] * (padding_size -
+                                                                 seq_len)
+            batch_data.append((word_ids, [label], seq_len))
+            if len(batch_data) == batch_size:
+                yield batch_data
+                batch_data = []
+
+    return reader
+
+
+class Args(object):
+    epoch = 1
+    batch_size = 4
+    class_num = 2
+    lr = 0.01
+    vocab_size = 1000
+    padding_size = 50
+    log_step = 2
+    train_step = 10
+
+
+def train(args, to_static):
+    program_translator.enable(to_static)
+    place = fluid.CUDAPlace(0) if fluid.is_compiled_with_cuda() \
+        else fluid.CPUPlace()
+
+    with fluid.dygraph.guard(place):
+        np.random.seed(SEED)
+        fluid.default_startup_program().random_seed = SEED
+        fluid.default_main_program().random_seed = SEED
+
+        train_reader = fake_data_reader(args.class_num, args.vocab_size,
+                                        args.batch_size, args.padding_size)
+        train_loader = fluid.io.DataLoader.from_generator(capacity=24)
+        train_loader.set_sample_list_generator(train_reader)
+
+        if args.model_type == 'cnn_net':
+            model = CNN(args.vocab_size, args.batch_size, args.padding_size)
+        elif args.model_type == 'bow_net':
+            model = BOW(args.vocab_size, args.batch_size, args.padding_size)
+        elif args.model_type == 'gru_net':
+            model = GRU(args.vocab_size, args.batch_size, args.padding_size)
+        elif args.model_type == 'bigru_net':
+            model = BiGRU(args.vocab_size, args.batch_size, args.padding_size)
+        sgd_optimizer = fluid.optimizer.Adagrad(
+            learning_rate=args.lr, parameter_list=model.parameters())
+
+        loss_data = []
+        for eop in range(args.epoch):
+            time_begin = time.time()
+            for batch_id, data in enumerate(train_loader()):
+                word_ids, labels, seq_lens = data
+                doc = to_variable(word_ids.numpy().reshape(-1)).astype('int64')
+                label = labels.astype('int64')
+
+                model.train()
+                avg_cost, prediction, acc = model(doc, label)
+                loss_data.append(avg_cost.numpy()[0])
+
+                avg_cost.backward()
+                sgd_optimizer.minimize(avg_cost)
+                model.clear_gradients()
+
+                if batch_id % args.log_step == 0:
+                    time_end = time.time()
+                    used_time = time_end - time_begin
+                    print("step: %d, ave loss: %f, speed: %f steps/s" %
+                          (batch_id, avg_cost.numpy()[0],
+                           args.log_step / used_time))
+                    time_begin = time.time()
+
+                if batch_id == args.train_step:
+                    break
+                batch_id += 1
+    return loss_data
+
+
+class TestSentiment(unittest.TestCase):
+    def setUp(self):
+        self.args = Args()
+
+    def train_model(self, model_type='cnn_net'):
+        self.args.model_type = model_type
+        st_out = train(self.args, True)
+        dy_out = train(self.args, False)
+        self.assertTrue(
+            np.allclose(dy_out, st_out),
+            msg="dy_out:\n {}\n st_out:\n {}".format(dy_out, st_out))
+
+    def test_train(self):
+        model_types = ['cnn_net', 'bow_net', 'gru_net', 'bigru_net']
+        for model_type in model_types:
+            print('training %s ....' % model_type)
+            self.train_model(model_type)
+
+
+if __name__ == '__main__':
+    unittest.main()
-- 
GitLab