release code

ERNIE/batching.py

+#   Copyright (c) 2019 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.
+"""Mask, padding and batching."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+
+def mask(batch_tokens, seg_labels, mask_word_tags, total_token_num, vocab_size, CLS=1, SEP=2, MASK=3):
+    """
+    Add mask for batch_tokens, return out, mask_label, mask_pos;
+    Note: mask_pos responding the batch_tokens after padded;
+    """
+    max_len = max([len(sent) for sent in batch_tokens])
+    mask_label = []
+    mask_pos = []
+    prob_mask = np.random.rand(total_token_num)
+    # Note: the first token is [CLS], so [low=1]
+    replace_ids = np.random.randint(1, high=vocab_size, size=total_token_num)
+    pre_sent_len = 0
+    prob_index = 0
+    for sent_index, sent in enumerate(batch_tokens):
+        mask_flag = False
+        mask_word = mask_word_tags[sent_index]
+        prob_index += pre_sent_len
+        if mask_word:
+            beg = 0
+            for token_index, token in enumerate(sent):
+                seg_label = seg_labels[sent_index][token_index]
+                if seg_label == 1:
+                    continue
+                if beg == 0:
+                    if seg_label != -1:
+                        beg = token_index
+                    continue
+
+                prob = prob_mask[prob_index + beg]
+                if prob > 0.15:
+                    pass
+                else:
+                    for index in xrange(beg, token_index):
+                        prob = prob_mask[prob_index + index]
+                        base_prob = 1.0
+                        if index == beg:
+                            base_prob = 0.15
+                        if base_prob * 0.2 < prob <= base_prob:
+                            mask_label.append(sent[index])
+                            sent[index] = MASK
+                            mask_flag = True
+                            mask_pos.append(sent_index * max_len + index)
+                        elif base_prob * 0.1 < prob <= base_prob * 0.2:
+                            mask_label.append(sent[index])
+                            sent[index] = replace_ids[prob_index + index]
+                            mask_flag = True
+                            mask_pos.append(sent_index * max_len + index)
+                        else:
+                            mask_label.append(sent[index])
+                            mask_pos.append(sent_index * max_len + index)
+
+                if seg_label == -1:
+                    beg = 0
+                else:
+                    beg = token_index
+        else:
+            for token_index, token in enumerate(sent):
+                prob = prob_mask[prob_index + token_index]
+                if prob > 0.15:
+                    continue
+                elif 0.03 < prob <= 0.15:
+                    # mask
+                    if token != SEP and token != CLS:
+                        mask_label.append(sent[token_index])
+                        sent[token_index] = MASK
+                        mask_flag = True
+                        mask_pos.append(sent_index * max_len + token_index)
+                elif 0.015 < prob <= 0.03:
+                    # random replace
+                    if token != SEP and token != CLS:
+                        mask_label.append(sent[token_index])
+                        sent[token_index] = replace_ids[prob_index + token_index]
+                        mask_flag = True
+                        mask_pos.append(sent_index * max_len + token_index)
+                else:
+                    # keep the original token
+                    if token != SEP and token != CLS:
+                        mask_label.append(sent[token_index])
+                        mask_pos.append(sent_index * max_len + token_index)
+
+        pre_sent_len = len(sent)
+
+    mask_label = np.array(mask_label).astype("int64").reshape([-1, 1])
+    mask_pos = np.array(mask_pos).astype("int64").reshape([-1, 1])
+    return batch_tokens, mask_label, mask_pos
+
+
+def prepare_batch_data(insts,
+                       total_token_num,
+                       voc_size=0,
+                       pad_id=None,
+                       cls_id=None,
+                       sep_id=None,
+                       mask_id=None,
+                       return_attn_bias=True,
+                       return_max_len=True,
+                       return_num_token=False):
+
+    batch_src_ids = [inst[0] for inst in insts]
+    batch_sent_ids = [inst[1] for inst in insts]
+    batch_pos_ids = [inst[2] for inst in insts]
+    labels = [inst[3] for inst in insts]
+    labels = np.array(labels).astype("int64").reshape([-1, 1])
+    seg_labels = [inst[4] for inst in insts]
+    mask_word_tags = [inst[5] for inst in insts]
+
+    # First step: do mask without padding
+    assert mask_id >= 0, "[FATAL] mask_id must >= 0"
+    out, mask_label, mask_pos = mask(
+        batch_src_ids,
+        seg_labels,
+        mask_word_tags,
+        total_token_num,
+        vocab_size=voc_size,
+        CLS=cls_id,
+        SEP=sep_id,
+        MASK=mask_id)
+
+    # Second step: padding
+    src_id, next_sent_index, self_attn_bias = pad_batch_data(
+        out, pad_idx=pad_id, return_next_sent_pos=True, return_attn_bias=True)
+    pos_id = pad_batch_data(batch_pos_ids, pad_idx=pad_id)
+    sent_id = pad_batch_data(batch_sent_ids, pad_idx=pad_id)
+
+    return_list = [src_id, pos_id, sent_id, self_attn_bias, mask_label, mask_pos, labels, next_sent_index]
+
+    return return_list
+
+
+def pad_batch_data(insts,
+                   pad_idx=0,
+                   return_pos=False,
+                   return_next_sent_pos=False,
+                   return_attn_bias=False,
+                   return_max_len=False,
+                   return_num_token=False):
+    """
+    Pad the instances to the max sequence length in batch, and generate the
+    corresponding position data and attention bias.
+    """
+    return_list = []
+    max_len = max(len(inst) for inst in insts)
+    # Any token included in dict can be used to pad, since the paddings' loss
+    # will be masked out by weights and make no effect on parameter gradients.
+
+    inst_data = np.array(
+        [inst + list([pad_idx] * (max_len - len(inst))) for inst in insts])
+    return_list += [inst_data.astype("int64").reshape([-1, max_len, 1])]
+
+    # next_sent_pos for extract first token embedding of each sentence
+    if return_next_sent_pos:
+        batch_size = inst_data.shape[0]
+        max_seq_len = inst_data.shape[1]
+        next_sent_index = np.array(
+            range(0, batch_size * max_seq_len, max_seq_len)).astype(
+                "int64").reshape(-1, 1)
+        return_list += [next_sent_index]
+
+    # position data
+    if return_pos:
+        inst_pos = np.array([
+            list(range(0, len(inst))) + [pad_idx] * (max_len - len(inst))
+            for inst in insts
+        ])
+
+        return_list += [inst_pos.astype("int64").reshape([-1, max_len, 1])]
+
+    if return_attn_bias:
+        # This is used to avoid attention on paddings.
+        slf_attn_bias_data = np.array([[0] * len(inst) + [-1e9] *
+                                       (max_len - len(inst)) for inst in insts])
+        slf_attn_bias_data = np.tile(
+            slf_attn_bias_data.reshape([-1, 1, max_len]), [1, max_len, 1])
+        return_list += [slf_attn_bias_data.astype("float32")]
+
+    if return_max_len:
+        return_list += [max_len]
+
+    if return_num_token:
+        num_token = 0
+        for inst in insts:
+            num_token += len(inst)
+        return_list += [num_token]
+
+    return return_list if len(return_list) > 1 else return_list[0]
+
+
+if __name__ == "__main__":
+    pass

ERNIE/finetune/classifier.py

+#   Copyright (c) 2019 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.
+"""Model for classifier."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import time
+import numpy as np
+
+import paddle.fluid as fluid
+
+from model.ernie import ErnieModel
+
+
+def create_model(args,
+                 pyreader_name,
+                 ernie_config,
+                 is_prediction=False):
+    pyreader = fluid.layers.py_reader(
+        capacity=50,
+        shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
+                [-1, args.max_seq_len, 1],
+                [-1, args.max_seq_len, args.max_seq_len], [-1, 1], [-1, 1], [-1, 1]],
+        dtypes=['int64', 'int64', 'int64', 'float', 'int64', 'int64', 'int64'],
+        lod_levels=[0, 0, 0, 0, 0, 0, 0],
+        name=pyreader_name,
+        use_double_buffer=True)
+
+    (src_ids, sent_ids, pos_ids, self_attn_mask, labels,
+     next_sent_index, qids) = fluid.layers.read_file(pyreader)
+
+    ernie = ErnieModel(
+        src_ids=src_ids,
+        position_ids=pos_ids,
+        sentence_ids=sent_ids,
+        self_attn_mask=self_attn_mask,
+        config=ernie_config,
+        use_fp16=args.use_fp16)
+
+    cls_feats = ernie.get_pooled_output(next_sent_index)
+    cls_feats = fluid.layers.dropout(
+        x=cls_feats,
+        dropout_prob=0.1,
+        dropout_implementation="upscale_in_train")
+    logits = fluid.layers.fc(
+        input=cls_feats,
+        size=ernie_config["num_labels"],
+        param_attr=fluid.ParamAttr(
+            name="cls_out_w",
+            initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
+        bias_attr=fluid.ParamAttr(
+            name="cls_out_b", initializer=fluid.initializer.Constant(0.)))
+
+    if is_prediction:
+        probs = fluid.layers.softmax(logits)
+        feed_targets_name = [
+            src_ids.name, pos_ids.name, sent_ids.name, self_attn_mask.name,
+            next_sent_index.name
+        ]
+        return pyreader, probs, feed_targets_name
+
+    ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
+        logits=logits, label=labels, return_softmax=True)
+    loss = fluid.layers.mean(x=ce_loss)
+
+    if args.use_fp16 and args.loss_scaling > 1.0:
+        loss *= args.loss_scaling
+
+    num_seqs = fluid.layers.create_tensor(dtype='int64')
+    accuracy = fluid.layers.accuracy(input=probs, label=labels, total=num_seqs)
+
+    graph_vars = {"loss": loss,
+                  "probs": probs,
+                  "accuracy": accuracy,
+                  "labels": labels,
+                  "num_seqs": num_seqs,
+                  "qids": qids}
+
+    for k, v in graph_vars.items():
+        v.persistable=True
+
+    return pyreader, graph_vars
+
+def evaluate_mrr(preds):
+    last_qid = None
+    total_mrr = 0.0
+    qnum = 0.0
+    rank = 0.0
+    correct = False
+    for qid, score, label in preds:
+        if qid != last_qid:
+            rank = 0.0
+            qnum += 1
+            correct = False
+            last_qid = qid
+
+        rank += 1
+        if not correct and label != 0:
+            total_mrr += 1.0 / rank
+            correct = True
+
+    return total_mrr / qnum
+
+def evaluate_map(preds):
+    def singe_map(st, en):
+        total_p = 0.0
+        correct_num = 0.0
+        for index in xrange(st, en):
+            if int(preds[index][2]) != 0:
+                correct_num += 1
+                total_p += correct_num / (index - st + 1)
+        if int(correct_num) == 0:
+            return 0.0
+        return total_p / correct_num
+
+    last_qid = None
+    total_map = 0.0
+    qnum = 0.0
+    st = 0
+    for i in xrange(len(preds)):
+        qid = preds[i][0]
+        if qid != last_qid:
+            qnum += 1
+            if last_qid != None:
+                total_map += singe_map(st, i)
+            st = i
+            last_qid = qid
+
+    total_map += singe_map(st, len(preds))
+    return total_map / qnum
+
+def evaluate(exe, test_program, test_pyreader, graph_vars, eval_phase):
+    train_fetch_list = [graph_vars["loss"].name,
+                        graph_vars["accuracy"].name,
+                        graph_vars["num_seqs"].name
+                       ] 
+
+    if eval_phase == "train":
+        if "learning_rate" in graph_vars:
+            train_fetch_list.append(graph_vars["learning_rate"].name)
+        outputs = exe.run(fetch_list=train_fetch_list)
+        ret = {"loss":np.mean(outputs[0]), "accuracy":np.mean(outputs[1])}
+        if "learning_rate" in graph_vars:
+            ret["learning_rate"] = float(outputs[4][0])
+        return ret
+
+    test_pyreader.start()
+    total_cost, total_acc, total_num_seqs, total_label_pos_num, total_pred_pos_num, total_correct_num = 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
+    qids, labels, scores = [], [], []
+    time_begin = time.time()
+
+    fetch_list = [graph_vars["loss"].name,
+                  graph_vars["accuracy"].name,
+                  graph_vars["probs"].name,
+                  graph_vars["labels"].name,
+                  graph_vars["num_seqs"].name,
+                  graph_vars["qids"].name]
+    while True:
+        try:
+            np_loss, np_acc, np_probs, np_labels, np_num_seqs, np_qids = exe.run(program=test_program,
+                                                                                 fetch_list=fetch_list)
+            total_cost += np.sum(np_loss * np_num_seqs)
+            total_acc += np.sum(np_acc * np_num_seqs)
+            total_num_seqs += np.sum(np_num_seqs)
+            labels.extend(np_labels.reshape((-1)).tolist())
+            qids.extend(np_qids.reshape(-1).tolist())
+            scores.extend(np_probs[:,1].reshape(-1).tolist())
+            np_preds = np.argmax(np_probs, axis=1).astype(np.float32)
+            total_label_pos_num += np.sum(np_labels)
+            total_pred_pos_num += np.sum(np_preds)
+            total_correct_num += np.sum(np.dot(np_preds, np_labels))
+        except fluid.core.EOFException:
+            test_pyreader.reset()
+            break
+    time_end = time.time()
+
+    if len(qids) == 0:
+        print("[%s evaluation] ave loss: %f, ave acc: %f, data_num: %d, elapsed time: %f s" %
+            (eval_phase, total_cost / total_num_seqs,
+            total_acc / total_num_seqs, total_num_seqs, time_end - time_begin))
+    else:
+        r = total_correct_num / total_label_pos_num
+        p = total_correct_num / total_pred_pos_num
+        f = 2 * p * r / (p + r)
+
+        assert len(qids) == len(labels) == len(scores)
+        preds = sorted(zip(qids, scores, labels), key=lambda elem:(elem[0], -elem[1]))
+        mrr = evaluate_mrr(preds)
+        map = evaluate_map(preds)
+
+        print("[%s evaluation] ave loss: %f, ave_acc: %f, mrr: %f, map: %f, p: %f, r: %f, f1: %f, data_num: %d, elapsed time: %f s" %
+              (eval_phase, total_cost / total_num_seqs,
+              total_acc / total_num_seqs,
+              mrr, map, p, r, f, total_num_seqs, time_end - time_begin))

ERNIE/finetune/sequence_label.py

+#   Copyright (c) 2019 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import argparse
+import numpy as np
+import multiprocessing
+
+import paddle
+import paddle.fluid as fluid
+
+from model.ernie import ErnieModel
+
+def create_model(args,
+                 pyreader_name,
+                 ernie_config,
+                 is_prediction=False):
+    pyreader = fluid.layers.py_reader(
+        capacity=50,
+        shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
+                [-1, args.max_seq_len, 1], [-1, args.max_seq_len, args.max_seq_len],
+                [-1, args.max_seq_len, 1], [-1, 1]],
+        dtypes=['int64', 'int64', 'int64', 'float', 'int64', 'int64'],
+        lod_levels=[0, 0, 0, 0, 0, 0],
+        name=pyreader_name,
+        use_double_buffer=True)
+
+    (src_ids, sent_ids, pos_ids, self_attn_mask, labels,
+     seq_lens) = fluid.layers.read_file(pyreader)
+
+    ernie = ErnieModel(
+        src_ids=src_ids,
+        position_ids=pos_ids,
+        sentence_ids=sent_ids,
+        self_attn_mask=self_attn_mask,
+        config=ernie_config,
+        use_fp16=args.use_fp16)
+
+    enc_out = ernie.get_sequence_output()
+    logits = fluid.layers.fc(
+        input=enc_out,
+        size=args.num_labels,
+        num_flatten_dims=2,
+        param_attr=fluid.ParamAttr(
+            name="cls_seq_label_out_w",
+            initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
+        bias_attr=fluid.ParamAttr(
+            name="cls_seq_label_out_b", initializer=fluid.initializer.Constant(0.)))
+
+    ret_labels = fluid.layers.reshape(x=labels, shape=[-1,1])
+    ret_infers = fluid.layers.reshape(x=fluid.layers.argmax(logits, axis=2), shape=[-1,1])
+
+    labels = fluid.layers.flatten(labels, axis=2)
+    ce_loss, probs = fluid.layers.softmax_with_cross_entropy(
+        logits=fluid.layers.flatten(logits, axis=2),
+        label=labels, return_softmax=True)
+    loss = fluid.layers.mean(x=ce_loss)
+
+    if args.use_fp16 and args.loss_scaling > 1.0:
+        loss *= args.loss_scaling
+
+    graph_vars = {"loss": loss,
+                  "probs": probs,
+                  "labels": ret_labels,
+                  "infers": ret_infers,
+                  "seq_lens": seq_lens}
+
+    for k, v in graph_vars.items():
+        v.persistable=True
+
+    return pyreader, graph_vars
+
+def chunk_eval(np_labels, np_infers, np_lens, tag_num, dev_count=1):
+
+    def extract_bio_chunk(seq):
+        chunks = []
+        cur_chunk = None
+        null_index = tag_num - 1
+        for index in xrange(len(seq)):
+            tag = seq[index]
+            tag_type = tag // 2
+            tag_pos = tag % 2
+
+            if tag == null_index:
+                if cur_chunk is not None:
+                    chunks.append(cur_chunk)
+                    cur_chunk = None
+                continue
+
+            if tag_pos == 0:
+                if cur_chunk is not None:
+                    chunks.append(cur_chunk)
+                    cur_chunk = {}
+                cur_chunk = {"st":index, "en": index + 1, "type": tag_type}
+
+            else:
+                if cur_chunk is None:
+                    cur_chunk = {"st":index, "en": index + 1, "type": tag_type}
+                    continue
+
+                if cur_chunk["type"] == tag_type:
+                    cur_chunk["en"]  = index + 1
+                else:
+                    chunks.append(cur_chunk)
+                    cur_chunk = {"st":index, "en": index + 1, "type": tag_type}
+
+        if cur_chunk is not None:
+            chunks.append(cur_chunk)
+        return chunks
+
+    null_index = tag_num - 1
+    num_label = 0
+    num_infer = 0
+    num_correct = 0
+    labels = np_labels.reshape([-1]).astype(np.int32).tolist()
+    infers = np_infers.reshape([-1]).astype(np.int32).tolist()
+    all_lens = np_lens.reshape([dev_count, -1]).astype(np.int32).tolist()
+
+    base_index = 0
+    for dev_index in xrange(dev_count):
+        lens = all_lens[dev_index]
+        max_len = 0
+        for l in lens:
+            max_len = max(max_len, l)
+
+        for i in xrange(len(lens)):
+            seq_st = base_index + i * max_len + 1
+            seq_en = seq_st + (lens[i] - 2)
+            infer_chunks = extract_bio_chunk(infers[seq_st:seq_en])
+            label_chunks = extract_bio_chunk(labels[seq_st:seq_en])
+            num_infer += len(infer_chunks)
+            num_label += len(label_chunks)
+
+            infer_index = 0
+            label_index = 0
+            while label_index < len(label_chunks) and infer_index < len(infer_chunks):
+                if infer_chunks[infer_index]["st"] < label_chunks[label_index]["st"]:
+                    infer_index += 1
+                elif infer_chunks[infer_index]["st"] > label_chunks[label_index]["st"]:
+                    label_index += 1
+                else:
+                    if infer_chunks[infer_index]["en"] == label_chunks[label_index]["en"] and \
+                            infer_chunks[infer_index]["type"] == label_chunks[label_index]["type"]:
+                        num_correct += 1
+
+                    infer_index += 1
+                    label_index += 1
+
+        base_index += max_len * len(lens)
+
+    return num_label, num_infer, num_correct
+
+def calculate_f1(num_label, num_infer, num_correct):
+    if num_infer == 0:
+        precision = 0.0
+    else:
+        precision = num_correct * 1.0 / num_infer
+
+    if num_label == 0:
+        recall = 0.0
+    else:
+        recall = num_correct * 1.0 / num_label
+
+    if num_correct == 0:
+        f1 = 0.0
+    else:
+        f1 = 2 * precision * recall / (precision + recall)
+    return precision, recall, f1
+
+def evaluate(exe, program, pyreader, graph_vars, tag_num, eval_phase, dev_count=1):
+    fetch_list = [graph_vars["labels"].name,
+                  graph_vars["infers"].name,
+                  graph_vars["seq_lens"].name]
+
+    if eval_phase == "train":
+        fetch_list.append(graph_vars["loss"].name)
+        if "learning_rate" in graph_vars:
+            fetch_list.append(graph_vars["learning_rate"].name)
+        outputs = exe.run(fetch_list=fetch_list)
+        np_labels, np_infers, np_lens, np_loss = outputs[:4]
+        num_label, num_infer, num_correct = chunk_eval(np_labels, np_infers, np_lens, tag_num, dev_count)
+        precision, recall, f1 = calculate_f1(num_label, num_infer, num_correct)
+        outputs = {"precision": precision, "recall": recall, "f1": f1, "loss": np.mean(np_loss)}
+        if "learning_rate" in graph_vars:
+            outputs["lr"] = float(outputs[4][0])
+        return outputs
+
+    else:
+        total_label, total_infer, total_correct = 0.0, 0.0, 0.0
+        time_begin = time.time()
+        pyreader.start()
+        while True:
+            try:
+                np_labels, np_infers, np_lens = exe.run(program=program, fetch_list=fetch_list)
+                label_num, infer_num, correct_num = chunk_eval(np_labels, np_infers, np_lens, tag_num, dev_count)
+                total_infer += infer_num
+                total_label += label_num
+                total_correct += correct_num
+
+            except fluid.core.EOFException:
+                pyreader.reset()
+                break
+
+        precision, recall, f1 = calculate_f1(total_label, total_infer, total_correct)
+        time_end = time.time()
+
+        print("[%s evaluation] f1: %f, precision: %f, recall: %f, elapsed time: %f s" %
+              (eval_phase, f1, precision, recall, time_end - time_begin))

ERNIE/finetune_args.py

+#   Copyright (c) 2019 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import argparse
+
+from utils.args import ArgumentGroup
+
+# yapf: disable
+parser = argparse.ArgumentParser(__doc__)
+model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
+model_g.add_arg("ernie_config_path",         str,  None,           "Path to the json file for ernie model config.")
+model_g.add_arg("init_checkpoint",          str,  None,           "Init checkpoint to resume training from.")
+model_g.add_arg("init_pretraining_params",  str,  None,
+                "Init pre-training params which preforms fine-tuning from. If the "
+                 "arg 'init_checkpoint' has been set, this argument wouldn't be valid.")
+model_g.add_arg("checkpoints",              str,  "checkpoints",  "Path to save checkpoints.")
+
+train_g = ArgumentGroup(parser, "training", "training options.")
+train_g.add_arg("epoch",             int,    3,       "Number of epoches for fine-tuning.")
+train_g.add_arg("learning_rate",     float,  5e-5,    "Learning rate used to train with warmup.")
+train_g.add_arg("lr_scheduler",      str,    "linear_warmup_decay",
+                "scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay'])
+train_g.add_arg("weight_decay",      float,  0.01,    "Weight decay rate for L2 regularizer.")
+train_g.add_arg("warmup_proportion", float,  0.1,
+                "Proportion of training steps to perform linear learning rate warmup for.")
+train_g.add_arg("save_steps",        int,    10000,   "The steps interval to save checkpoints.")
+train_g.add_arg("validation_steps",  int,    1000,    "The steps interval to evaluate model performance.")
+train_g.add_arg("use_fp16",          bool,   False,   "Whether to use fp16 mixed precision training.")
+train_g.add_arg("loss_scaling",      float,  1.0,
+                "Loss scaling factor for mixed precision training, only valid when use_fp16 is enabled.")
+
+log_g = ArgumentGroup(parser,     "logging", "logging related.")
+log_g.add_arg("skip_steps",          int,    10,    "The steps interval to print loss.")
+log_g.add_arg("verbose",             bool,   False, "Whether to output verbose log.")
+
+data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
+data_g.add_arg("train_set",           str,  None,  "Path to training data.")
+data_g.add_arg("test_set",            str,  None,  "Path to test data.")
+data_g.add_arg("dev_set",             str,  None,  "Path to validation data.")
+data_g.add_arg("vocab_path",          str,  None,  "Vocabulary path.")
+data_g.add_arg("max_seq_len",         int,  512,   "Number of words of the longest seqence.")
+data_g.add_arg("batch_size",          int,  32,    "Total examples' number in batch for training. see also --in_tokens.")
+data_g.add_arg("in_tokens",           bool, False,
+              "If set, the batch size will be the maximum number of tokens in one batch. "
+              "Otherwise, it will be the maximum number of examples in one batch.")
+data_g.add_arg("do_lower_case",       bool, True,
+               "Whether to lower case the input text. Should be True for uncased models and False for cased models.")
+data_g.add_arg("random_seed",         int,  0,     "Random seed.")
+data_g.add_arg("label_map_config",    str,  None,  "label_map_path.")
+data_g.add_arg("num_labels",    int,  2,  "label number")
+
+run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
+run_type_g.add_arg("use_cuda",                     bool,   True,  "If set, use GPU for training.")
+run_type_g.add_arg("use_fast_executor",            bool,   False, "If set, use fast parallel executor (in experiment).")
+run_type_g.add_arg("num_iteration_per_drop_scope", int,    10,    "Iteration intervals to drop scope.")
+run_type_g.add_arg("do_train",                     bool,   True,  "Whether to perform training.")
+run_type_g.add_arg("do_val",                       bool,   True,  "Whether to perform evaluation on dev data set.")
+run_type_g.add_arg("do_test",                      bool,   True,  "Whether to perform evaluation on test data set.")
+run_type_g.add_arg("metrics",                      bool,   True,  "Whether to perform evaluation on test data set.")

ERNIE/model/ernie.py

+#   Copyright (c) 2019 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.
+"""Ernie model."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import json
+import numpy as np
+import paddle.fluid as fluid
+from model.transformer_encoder import encoder, pre_process_layer
+
+
+class ErnieConfig(object):
+    def __init__(self, config_path):
+        self._config_dict = self._parse(config_path)
+
+    def _parse(self, config_path):
+        try:
+            with open(config_path) as json_file:
+                config_dict = json.load(json_file)
+        except Exception:
+            raise IOError("Error in parsing Ernie model config file '%s'" %
+                          config_path)
+        else:
+            return config_dict
+
+    def __getitem__(self, key):
+        return self._config_dict[key]
+
+    def print_config(self):
+        for arg, value in sorted(six.iteritems(self._config_dict)):
+            print('%s: %s' % (arg, value))
+        print('------------------------------------------------')
+
+
+class ErnieModel(object):
+    def __init__(self,
+                 src_ids,
+                 position_ids,
+                 sentence_ids,
+                 self_attn_mask,
+                 config,
+                 weight_sharing=True,
+                 use_fp16=False):
+
+        self._emb_size = config['hidden_size']
+        self._n_layer = config['num_hidden_layers']
+        self._n_head = config['num_attention_heads']
+        self._voc_size = config['vocab_size']
+        self._max_position_seq_len = config['max_position_embeddings']
+        self._sent_types = config['type_vocab_size']
+        self._hidden_act = config['hidden_act']
+        self._prepostprocess_dropout = config['hidden_dropout_prob']
+        self._attention_dropout = config['attention_probs_dropout_prob']
+        self._weight_sharing = weight_sharing
+
+        self._word_emb_name = "word_embedding"
+        self._pos_emb_name = "pos_embedding"
+        self._sent_emb_name = "sent_embedding"
+        self._dtype = "float16" if use_fp16 else "float32"
+
+        # Initialize all weigths by truncated normal initializer, and all biases 
+        # will be initialized by constant zero by default.
+        self._param_initializer = fluid.initializer.TruncatedNormal(
+            scale=config['initializer_range'])
+
+        self._build_model(src_ids, position_ids, sentence_ids, self_attn_mask)
+
+    def _build_model(self, src_ids, position_ids, sentence_ids, self_attn_mask):
+        # padding id in vocabulary must be set to 0
+        emb_out = fluid.layers.embedding(
+            input=src_ids,
+            size=[self._voc_size, self._emb_size],
+            dtype=self._dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._word_emb_name, initializer=self._param_initializer),
+            is_sparse=False)
+        position_emb_out = fluid.layers.embedding(
+            input=position_ids,
+            size=[self._max_position_seq_len, self._emb_size],
+            dtype=self._dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._pos_emb_name, initializer=self._param_initializer))
+
+        sent_emb_out = fluid.layers.embedding(
+            sentence_ids,
+            size=[self._sent_types, self._emb_size],
+            dtype=self._dtype,
+            param_attr=fluid.ParamAttr(
+                name=self._sent_emb_name, initializer=self._param_initializer))
+
+        emb_out = emb_out + position_emb_out
+        emb_out = emb_out + sent_emb_out
+
+        emb_out = pre_process_layer(
+            emb_out, 'nd', self._prepostprocess_dropout, name='pre_encoder')
+
+        if self._dtype is "float16":
+            self_attn_mask = fluid.layers.cast(
+                x=self_attn_mask, dtype=self._dtype)
+
+        n_head_self_attn_mask = fluid.layers.stack(
+            x=[self_attn_mask] * self._n_head, axis=1)
+        n_head_self_attn_mask.stop_gradient = True
+
+        self._enc_out = encoder(
+            enc_input=emb_out,
+            attn_bias=n_head_self_attn_mask,
+            n_layer=self._n_layer,
+            n_head=self._n_head,
+            d_key=self._emb_size // self._n_head,
+            d_value=self._emb_size // self._n_head,
+            d_model=self._emb_size,
+            d_inner_hid=self._emb_size * 4,
+            prepostprocess_dropout=self._prepostprocess_dropout,
+            attention_dropout=self._attention_dropout,
+            relu_dropout=0,
+            hidden_act=self._hidden_act,
+            preprocess_cmd="",
+            postprocess_cmd="dan",
+            param_initializer=self._param_initializer,
+            name='encoder')
+
+    def get_sequence_output(self):
+        return self._enc_out
+
+    def get_pooled_output(self, next_sent_index):
+        """Get the first feature of each sequence for classification"""
+        self._reshaped_emb_out = fluid.layers.reshape(
+            x=self._enc_out, shape=[-1, self._emb_size], inplace=True)
+        next_sent_index = fluid.layers.cast(x=next_sent_index, dtype='int32')
+        next_sent_feat = fluid.layers.gather(
+            input=self._reshaped_emb_out, index=next_sent_index)
+        next_sent_feat = fluid.layers.fc(
+            input=next_sent_feat,
+            size=self._emb_size,
+            act="tanh",
+            param_attr=fluid.ParamAttr(
+                name="pooled_fc.w_0", initializer=self._param_initializer),
+            bias_attr="pooled_fc.b_0")
+        return next_sent_feat
+
+    def get_pretraining_output(self, mask_label, mask_pos, labels,
+                               next_sent_index):
+        """Get the loss & accuracy for pretraining"""
+
+        mask_pos = fluid.layers.cast(x=mask_pos, dtype='int32')
+
+        # extract the first token feature in each sentence
+        next_sent_feat = self.get_pooled_output(next_sent_index)
+        # extract masked tokens' feature
+        mask_feat = fluid.layers.gather(
+            input=self._reshaped_emb_out, index=mask_pos)
+
+        # transform: fc
+        mask_trans_feat = fluid.layers.fc(
+            input=mask_feat,
+            size=self._emb_size,
+            act=self._hidden_act,
+            param_attr=fluid.ParamAttr(
+                name='mask_lm_trans_fc.w_0',
+                initializer=self._param_initializer),
+            bias_attr=fluid.ParamAttr(name='mask_lm_trans_fc.b_0'))
+        # transform: layer norm 
+        mask_trans_feat = pre_process_layer(
+            mask_trans_feat, 'n', name='mask_lm_trans')
+
+        mask_lm_out_bias_attr = fluid.ParamAttr(
+            name="mask_lm_out_fc.b_0",
+            initializer=fluid.initializer.Constant(value=0.0))
+        if self._weight_sharing:
+            fc_out = fluid.layers.matmul(
+                x=mask_trans_feat,
+                y=fluid.default_main_program().global_block().var(
+                    self._word_emb_name),
+                transpose_y=True)
+            fc_out += fluid.layers.create_parameter(
+                shape=[self._voc_size],
+                dtype=self._dtype,
+                attr=mask_lm_out_bias_attr,
+                is_bias=True)
+
+        else:
+            fc_out = fluid.layers.fc(input=mask_trans_feat,
+                                     size=self._voc_size,
+                                     param_attr=fluid.ParamAttr(
+                                         name="mask_lm_out_fc.w_0",
+                                         initializer=self._param_initializer),
+                                     bias_attr=mask_lm_out_bias_attr)
+
+        mask_lm_loss = fluid.layers.softmax_with_cross_entropy(
+            logits=fc_out, label=mask_label)
+        mean_mask_lm_loss = fluid.layers.mean(mask_lm_loss)
+
+        next_sent_fc_out = fluid.layers.fc(
+            input=next_sent_feat,
+            size=2,
+            param_attr=fluid.ParamAttr(
+                name="next_sent_fc.w_0", initializer=self._param_initializer),
+            bias_attr="next_sent_fc.b_0")
+
+        next_sent_loss, next_sent_softmax = fluid.layers.softmax_with_cross_entropy(
+            logits=next_sent_fc_out, label=labels, return_softmax=True)
+
+        next_sent_acc = fluid.layers.accuracy(
+            input=next_sent_softmax, label=labels)
+
+        mean_next_sent_loss = fluid.layers.mean(next_sent_loss)
+
+        loss = mean_next_sent_loss + mean_mask_lm_loss
+        return next_sent_acc, mean_mask_lm_loss, loss

ERNIE/model/transformer_encoder.py

+#   Copyright (c) 2019 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.
+"""Transformer encoder."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from functools import partial
+import numpy as np
+
+import paddle.fluid as fluid
+import paddle.fluid.layers as layers
+
+
+def multi_head_attention(queries,
+                         keys,
+                         values,
+                         attn_bias,
+                         d_key,
+                         d_value,
+                         d_model,
+                         n_head=1,
+                         dropout_rate=0.,
+                         cache=None,
+                         param_initializer=None,
+                         name='multi_head_att'):
+    """
+    Multi-Head Attention. Note that attn_bias is added to the logit before
+    computing softmax activiation to mask certain selected positions so that
+    they will not considered in attention weights.
+    """
+    keys = queries if keys is None else keys
+    values = keys if values is None else values
+
+    if not (len(queries.shape) == len(keys.shape) == len(values.shape) == 3):
+        raise ValueError(
+            "Inputs: quries, keys and values should all be 3-D tensors.")
+
+    def __compute_qkv(queries, keys, values, n_head, d_key, d_value):
+        """
+        Add linear projection to queries, keys, and values.
+        """
+        q = layers.fc(input=queries,
+                      size=d_key * n_head,
+                      num_flatten_dims=2,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_query_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_query_fc.b_0')
+        k = layers.fc(input=keys,
+                      size=d_key * n_head,
+                      num_flatten_dims=2,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_key_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_key_fc.b_0')
+        v = layers.fc(input=values,
+                      size=d_value * n_head,
+                      num_flatten_dims=2,
+                      param_attr=fluid.ParamAttr(
+                          name=name + '_value_fc.w_0',
+                          initializer=param_initializer),
+                      bias_attr=name + '_value_fc.b_0')
+        return q, k, v
+
+    def __split_heads(x, n_head):
+        """
+        Reshape the last dimension of inpunt tensor x so that it becomes two
+        dimensions and then transpose. Specifically, input a tensor with shape
+        [bs, max_sequence_length, n_head * hidden_dim] then output a tensor
+        with shape [bs, n_head, max_sequence_length, hidden_dim].
+        """
+        hidden_size = x.shape[-1]
+        # The value 0 in shape attr means copying the corresponding dimension
+        # size of the input as the output dimension size.
+        reshaped = layers.reshape(
+            x=x, shape=[0, 0, n_head, hidden_size // n_head], inplace=True)
+
+        # permuate the dimensions into:
+        # [batch_size, n_head, max_sequence_len, hidden_size_per_head]
+        return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])
+
+    def __combine_heads(x):
+        """
+        Transpose and then reshape the last two dimensions of inpunt tensor x
+        so that it becomes one dimension, which is reverse to __split_heads.
+        """
+        if len(x.shape) == 3: return x
+        if len(x.shape) != 4:
+            raise ValueError("Input(x) should be a 4-D Tensor.")
+
+        trans_x = layers.transpose(x, perm=[0, 2, 1, 3])
+        # The value 0 in shape attr means copying the corresponding dimension
+        # size of the input as the output dimension size.
+        return layers.reshape(
+            x=trans_x,
+            shape=[0, 0, trans_x.shape[2] * trans_x.shape[3]],
+            inplace=True)
+
+    def scaled_dot_product_attention(q, k, v, attn_bias, d_key, dropout_rate):
+        """
+        Scaled Dot-Product Attention
+        """
+        scaled_q = layers.scale(x=q, scale=d_key**-0.5)
+        product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
+        if attn_bias:
+            product += attn_bias
+        weights = layers.softmax(product)
+        if dropout_rate:
+            weights = layers.dropout(
+                weights,
+                dropout_prob=dropout_rate,
+                dropout_implementation="upscale_in_train",
+                is_test=False)
+        out = layers.matmul(weights, v)
+        return out
+
+    q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value)
+
+    if cache is not None:  # use cache and concat time steps
+        # Since the inplace reshape in __split_heads changes the shape of k and
+        # v, which is the cache input for next time step, reshape the cache
+        # input from the previous time step first.
+        k = cache["k"] = layers.concat(
+            [layers.reshape(
+                cache["k"], shape=[0, 0, d_model]), k], axis=1)
+        v = cache["v"] = layers.concat(
+            [layers.reshape(
+                cache["v"], shape=[0, 0, d_model]), v], axis=1)
+
+    q = __split_heads(q, n_head)
+    k = __split_heads(k, n_head)
+    v = __split_heads(v, n_head)
+
+    ctx_multiheads = scaled_dot_product_attention(q, k, v, attn_bias, d_key,
+                                                  dropout_rate)
+
+    out = __combine_heads(ctx_multiheads)
+
+    # Project back to the model size.
+    proj_out = layers.fc(input=out,
+                         size=d_model,
+                         num_flatten_dims=2,
+                         param_attr=fluid.ParamAttr(
+                             name=name + '_output_fc.w_0',
+                             initializer=param_initializer),
+                         bias_attr=name + '_output_fc.b_0')
+    return proj_out
+
+
+def positionwise_feed_forward(x,
+                              d_inner_hid,
+                              d_hid,
+                              dropout_rate,
+                              hidden_act,
+                              param_initializer=None,
+                              name='ffn'):
+    """
+    Position-wise Feed-Forward Networks.
+    This module consists of two linear transformations with a ReLU activation
+    in between, which is applied to each position separately and identically.
+    """
+    hidden = layers.fc(input=x,
+                       size=d_inner_hid,
+                       num_flatten_dims=2,
+                       act=hidden_act,
+                       param_attr=fluid.ParamAttr(
+                           name=name + '_fc_0.w_0',
+                           initializer=param_initializer),
+                       bias_attr=name + '_fc_0.b_0')
+    if dropout_rate:
+        hidden = layers.dropout(
+            hidden,
+            dropout_prob=dropout_rate,
+            dropout_implementation="upscale_in_train",
+            is_test=False)
+    out = layers.fc(input=hidden,
+                    size=d_hid,
+                    num_flatten_dims=2,
+                    param_attr=fluid.ParamAttr(
+                        name=name + '_fc_1.w_0', initializer=param_initializer),
+                    bias_attr=name + '_fc_1.b_0')
+    return out
+
+
+def pre_post_process_layer(prev_out, out, process_cmd, dropout_rate=0.,
+                           name=''):
+    """
+    Add residual connection, layer normalization and droput to the out tensor
+    optionally according to the value of process_cmd.
+    This will be used before or after multi-head attention and position-wise
+    feed-forward networks.
+    """
+    for cmd in process_cmd:
+        if cmd == "a":  # add residual connection
+            out = out + prev_out if prev_out else out
+        elif cmd == "n":  # add layer normalization
+            out_dtype = out.dtype
+            if out_dtype == fluid.core.VarDesc.VarType.FP16:
+                out = layers.cast(x=out, dtype="float32")
+            out = layers.layer_norm(
+                out,
+                begin_norm_axis=len(out.shape) - 1,
+                param_attr=fluid.ParamAttr(
+                    name=name + '_layer_norm_scale',
+                    initializer=fluid.initializer.Constant(1.)),
+                bias_attr=fluid.ParamAttr(
+                    name=name + '_layer_norm_bias',
+                    initializer=fluid.initializer.Constant(0.)))
+            if out_dtype == fluid.core.VarDesc.VarType.FP16:
+                out = layers.cast(x=out, dtype="float16")
+        elif cmd == "d":  # add dropout
+            if dropout_rate:
+                out = layers.dropout(
+                    out,
+                    dropout_prob=dropout_rate,
+                    dropout_implementation="upscale_in_train",
+                    is_test=False)
+    return out
+
+
+pre_process_layer = partial(pre_post_process_layer, None)
+post_process_layer = pre_post_process_layer
+
+
+def encoder_layer(enc_input,
+                  attn_bias,
+                  n_head,
+                  d_key,
+                  d_value,
+                  d_model,
+                  d_inner_hid,
+                  prepostprocess_dropout,
+                  attention_dropout,
+                  relu_dropout,
+                  hidden_act,
+                  preprocess_cmd="n",
+                  postprocess_cmd="da",
+                  param_initializer=None,
+                  name=''):
+    """The encoder layers that can be stacked to form a deep encoder.
+    This module consits of a multi-head (self) attention followed by
+    position-wise feed-forward networks and both the two components companied
+    with the post_process_layer to add residual connection, layer normalization
+    and droput.
+    """
+    attn_output = multi_head_attention(
+        pre_process_layer(
+            enc_input,
+            preprocess_cmd,
+            prepostprocess_dropout,
+            name=name + '_pre_att'),
+        None,
+        None,
+        attn_bias,
+        d_key,
+        d_value,
+        d_model,
+        n_head,
+        attention_dropout,
+        param_initializer=param_initializer,
+        name=name + '_multi_head_att')
+    attn_output = post_process_layer(
+        enc_input,
+        attn_output,
+        postprocess_cmd,
+        prepostprocess_dropout,
+        name=name + '_post_att')
+    ffd_output = positionwise_feed_forward(
+        pre_process_layer(
+            attn_output,
+            preprocess_cmd,
+            prepostprocess_dropout,
+            name=name + '_pre_ffn'),
+        d_inner_hid,
+        d_model,
+        relu_dropout,
+        hidden_act,
+        param_initializer=param_initializer,
+        name=name + '_ffn')
+    return post_process_layer(
+        attn_output,
+        ffd_output,
+        postprocess_cmd,
+        prepostprocess_dropout,
+        name=name + '_post_ffn')
+
+
+def encoder(enc_input,
+            attn_bias,
+            n_layer,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd="n",
+            postprocess_cmd="da",
+            param_initializer=None,
+            name=''):
+    """
+    The encoder is composed of a stack of identical layers returned by calling
+    encoder_layer.
+    """
+    for i in range(n_layer):
+        enc_output = encoder_layer(
+            enc_input,
+            attn_bias,
+            n_head,
+            d_key,
+            d_value,
+            d_model,
+            d_inner_hid,
+            prepostprocess_dropout,
+            attention_dropout,
+            relu_dropout,
+            hidden_act,
+            preprocess_cmd,
+            postprocess_cmd,
+            param_initializer=param_initializer,
+            name=name + '_layer_' + str(i))
+        enc_input = enc_output
+    enc_output = pre_process_layer(
+        enc_output, preprocess_cmd, prepostprocess_dropout, name="post_encoder")
+
+    return enc_output

ERNIE/optimization.py

+#   Copyright (c) 2019 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.
+"""Optimization and learning rate scheduling."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import numpy as np
+import paddle.fluid as fluid
+from utils.fp16 import create_master_params_grads, master_param_to_train_param
+
+
+def linear_warmup_decay(learning_rate, warmup_steps, num_train_steps):
+    """ Applies linear warmup of learning rate from 0 and decay to 0."""
+    with fluid.default_main_program()._lr_schedule_guard():
+        lr = fluid.layers.tensor.create_global_var(
+            shape=[1],
+            value=0.0,
+            dtype='float32',
+            persistable=True,
+            name="scheduled_learning_rate")
+
+        global_step = fluid.layers.learning_rate_scheduler._decay_step_counter()
+
+        with fluid.layers.control_flow.Switch() as switch:
+            with switch.case(global_step < warmup_steps):
+                warmup_lr = learning_rate * (global_step / warmup_steps)
+                fluid.layers.tensor.assign(warmup_lr, lr)
+            with switch.default():
+                decayed_lr = fluid.layers.learning_rate_scheduler.polynomial_decay(
+                    learning_rate=learning_rate,
+                    decay_steps=num_train_steps,
+                    end_learning_rate=0.0,
+                    power=1.0,
+                    cycle=False)
+                fluid.layers.tensor.assign(decayed_lr, lr)
+
+        return lr
+
+
+def optimization(loss,
+                 warmup_steps,
+                 num_train_steps,
+                 learning_rate,
+                 train_program,
+                 startup_prog,
+                 weight_decay,
+                 scheduler='linear_warmup_decay',
+                 use_fp16=False,
+                 loss_scaling=1.0):
+    if warmup_steps > 0:
+        if scheduler == 'noam_decay':
+            scheduled_lr = fluid.layers.learning_rate_scheduler\
+             .noam_decay(1/(warmup_steps *(learning_rate ** 2)),
+                         warmup_steps)
+        elif scheduler == 'linear_warmup_decay':
+            scheduled_lr = linear_warmup_decay(learning_rate, warmup_steps,
+                                               num_train_steps)
+        else:
+            raise ValueError("Unkown learning rate scheduler, should be "
+                             "'noam_decay' or 'linear_warmup_decay'")
+        optimizer = fluid.optimizer.Adam(learning_rate=scheduled_lr)
+    else:
+        optimizer = fluid.optimizer.Adam(learning_rate=learning_rate)
+        scheduled_lr = learning_rate
+
+    clip_norm_thres = 1.0
+    # When using mixed precision training, scale the gradient clip threshold
+    # by loss_scaling
+    if use_fp16 and loss_scaling > 1.0:
+        clip_norm_thres *= loss_scaling
+    fluid.clip.set_gradient_clip(
+        clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=clip_norm_thres))
+
+    def exclude_from_weight_decay(name):
+        if name.find("layer_norm") > -1:
+            return True
+        bias_suffix = ["_bias", "_b", ".b_0"]
+        for suffix in bias_suffix:
+            if name.endswith(suffix):
+                return True
+        return False
+
+    param_list = dict()
+
+    if use_fp16:
+        param_grads = optimizer.backward(loss)
+        master_param_grads = create_master_params_grads(
+            param_grads, train_program, startup_prog, loss_scaling)
+
+        for param, _ in master_param_grads:
+            param_list[param.name] = param * 1.0
+            param_list[param.name].stop_gradient = True
+
+        optimizer.apply_gradients(master_param_grads)
+
+        if weight_decay > 0:
+            for param, grad in master_param_grads:
+                if exclude_from_weight_decay(param.name.rstrip(".master")):
+                    continue
+                with param.block.program._optimized_guard(
+                    [param, grad]), fluid.framework.name_scope("weight_decay"):
+                    updated_param = param - param_list[
+                        param.name] * weight_decay * scheduled_lr
+                    fluid.layers.assign(output=param, input=updated_param)
+
+        master_param_to_train_param(master_param_grads, param_grads,
+                                    train_program)
+
+    else:
+        for param in train_program.global_block().all_parameters():
+            param_list[param.name] = param * 1.0
+            param_list[param.name].stop_gradient = True
+
+        _, param_grads = optimizer.minimize(loss)
+
+        if weight_decay > 0:
+            for param, grad in param_grads:
+                if exclude_from_weight_decay(param.name):
+                    continue
+                with param.block.program._optimized_guard(
+                    [param, grad]), fluid.framework.name_scope("weight_decay"):
+                    updated_param = param - param_list[
+                        param.name] * weight_decay * scheduled_lr
+                    fluid.layers.assign(output=param, input=updated_param)
+
+    return scheduled_lr

ERNIE/pretrain_args.py

+#   Copyright (c) 2019 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 absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import argparse
+
+from utils.args import ArgumentGroup, print_arguments
+
+# yapf: disable
+parser = argparse.ArgumentParser(__doc__)
+parser = argparse.ArgumentParser(__doc__)
+model_g = ArgumentGroup(parser, "model", "model configuration and paths.")
+model_g.add_arg("ernie_config_path",      str,  "./config/ernie_config.json",  "Path to the json file for ernie model config.")
+model_g.add_arg("init_checkpoint",        str,  None,                         "Init checkpoint to resume training from.")
+model_g.add_arg("checkpoints",            str,  "checkpoints",                "Path to save checkpoints.")
+model_g.add_arg("weight_sharing",         bool, True,                         "If set, share weights between word embedding and masked lm.")
+model_g.add_arg("generate_neg_sample",    bool, False,                         "If set, randomly generate negtive samples by positive samples.")
+
+train_g = ArgumentGroup(parser, "training", "training options.")
+train_g.add_arg("epoch",             int,    100,     "Number of epoches for training.")
+train_g.add_arg("learning_rate",     float,  0.0001,  "Learning rate used to train with warmup.")
+train_g.add_arg("lr_scheduler",      str,    "linear_warmup_decay",
+                "scheduler of learning rate.", choices=['linear_warmup_decay', 'noam_decay'])
+train_g.add_arg("weight_decay",      float,  0.01,    "Weight decay rate for L2 regularizer.")
+train_g.add_arg("num_train_steps",   int,    1000000, "Total steps to perform pretraining.")
+train_g.add_arg("warmup_steps",      int,    4000,    "Total steps to perform warmup when pretraining.")
+train_g.add_arg("save_steps",        int,    10000,   "The steps interval to save checkpoints.")
+train_g.add_arg("validation_steps",  int,    1000,    "The steps interval to evaluate model performance.")
+train_g.add_arg("use_fp16",          bool,   False,   "Whether to use fp16 mixed precision training.")
+train_g.add_arg("loss_scaling",      float,  1.0,
+                "Loss scaling factor for mixed precision training, only valid when use_fp16 is enabled.")
+
+log_g = ArgumentGroup(parser,     "logging", "logging related.")
+log_g.add_arg("skip_steps",          int,    10,    "The steps interval to print loss.")
+log_g.add_arg("verbose",             bool,   False, "Whether to output verbose log.")
+
+data_g = ArgumentGroup(parser, "data", "Data paths, vocab paths and data processing options")
+data_g.add_arg("train_filelist",           str,  "",  "Path to training filelist.")
+data_g.add_arg("valid_filelist",           str,  "",  "Path to valid filelist.")
+data_g.add_arg("test_filelist",            str,  "",  "Path to test filelist.")
+data_g.add_arg("vocab_path",          str,  "./config/vocab.txt",  "Vocabulary path.")
+data_g.add_arg("max_seq_len",         int,  512,                   "Number of words of the longest seqence.")
+data_g.add_arg("batch_size",          int,  16,                    "Total examples' number in batch for training. see also --in_tokens.")
+data_g.add_arg("in_tokens",           bool, False,
+              "If set, the batch size will be the maximum number of tokens in one batch. "
+              "Otherwise, it will be the maximum number of examples in one batch.")
+
+run_type_g = ArgumentGroup(parser, "run_type", "running type options.")
+run_type_g.add_arg("is_distributed",    bool,   False,  "If set, then start distributed training.")
+run_type_g.add_arg("use_cuda",          bool,   True,   "If set, use GPU for training.")
+run_type_g.add_arg("use_fast_executor", bool,   False,  "If set, use fast parallel executor (in experiment).")
+run_type_g.add_arg("do_test",           bool,   False,  "Whether to perform evaluation on test data set.")
+# yapf: enable

ERNIE/reader/pretraining.py

+#   Copyright (c) 2019 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
+from __future__ import division
+
+import os
+import numpy as np
+import types
+import gzip
+import logging
+import re
+import six
+import collections
+import tokenization
+
+import paddle
+import paddle.fluid as fluid
+
+from batching import prepare_batch_data
+
+class ErnieDataReader(object):
+    def __init__(self,
+                 filelist,
+                 vocab_path,
+                 batch_size=4096,
+                 max_seq_len=512,
+                 shuffle_files=True,
+                 epoch=100,
+                 voc_size=0,
+                 is_test=False,
+                 generate_neg_sample=False):
+
+        self.vocab = self.load_vocab(vocab_path)
+        self.filelist = filelist
+        self.batch_size = batch_size
+        self.shuffle_files = shuffle_files
+        self.epoch = epoch
+        self.current_epoch = 0
+        self.current_file_index = 0
+        self.total_file = 0
+        self.current_file = None
+        self.voc_size = voc_size
+        self.max_seq_len = max_seq_len
+        self.pad_id = self.vocab["[PAD]"]
+        self.cls_id = self.vocab["[CLS]"]
+        self.sep_id = self.vocab["[SEP]"]
+        self.mask_id = self.vocab["[MASK]"]
+        self.is_test = is_test
+        self.generate_neg_sample = generate_neg_sample
+        assert self.batch_size > 100, "Current batch size means total token's number, \
+                                       it should not be set to too small number."
+
+        if self.is_test:
+            self.epoch = 1
+            self.shuffle_files = False
+
+    def get_progress(self):
+        """return current progress of traning data
+        """
+        return self.current_epoch, self.current_file_index, self.total_file, self.current_file, self.mask_type
+
+    def parse_line(self, line, max_seq_len=512):
+        """ parse one line to token_ids, sentence_ids, pos_ids, label
+        """
+        line = line.strip().split(";")
+        assert len(line) == 5, "One sample must have 5 fields!"
+        (token_ids, sent_ids, pos_ids, seg_labels, label) = line
+        token_ids = [int(token) for token in token_ids.split(" ")]
+        sent_ids = [int(token) for token in sent_ids.split(" ")]
+        pos_ids = [int(token) for token in pos_ids.split(" ")]
+        seg_labels = [int(seg_label) for seg_label in seg_labels.split(" ")]
+        assert len(token_ids) == len(sent_ids) == len(
+            pos_ids) == len(seg_labels
+        ), "[Must be true]len(token_ids) == len(sent_ids) == len(pos_ids) == len(seg_labels)"
+        label = int(label)
+        if len(token_ids) > max_seq_len:
+            return None
+        return [token_ids, sent_ids, pos_ids, label, seg_labels]
+
+    def read_file(self, file):
+        assert file.endswith('.gz'), "[ERROR] %s is not a gzip file" % file
+        with gzip.open(file, "rb") as f:
+            for line in f:
+                parsed_line = self.parse_line(
+                    line, max_seq_len=self.max_seq_len)
+                if parsed_line is None:
+                    continue
+                yield parsed_line
+
+    def convert_to_unicode(self, text):
+        """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+        if six.PY3:
+            if isinstance(text, str):
+                return text
+            elif isinstance(text, bytes):
+                return text.decode("utf-8", "ignore")
+            else:
+                raise ValueError("Unsupported string type: %s" % (type(text)))
+        elif six.PY2:
+            if isinstance(text, str):
+                return text.decode("utf-8", "ignore")
+            elif isinstance(text, unicode):
+                return text
+            else:
+                raise ValueError("Unsupported string type: %s" % (type(text)))
+        else:
+            raise ValueError("Not running on Python2 or Python 3?")
+
+    def load_vocab(self, vocab_file):
+        """Loads a vocabulary file into a dictionary."""
+        vocab = collections.OrderedDict()
+        fin = open(vocab_file)
+        for num, line in enumerate(fin):
+            items = self.convert_to_unicode(line.strip()).split("\t")
+            if len(items) > 2:
+                break
+            token = items[0]
+            index = items[1] if len(items) == 2 else num
+            token = token.strip()
+            vocab[token] = int(index)
+        return vocab
+
+    def random_pair_neg_samples(self, pos_samples):
+        """ randomly generate negtive samples using pos_samples
+
+            Args:
+                pos_samples: list of positive samples
+            
+            Returns:
+                neg_samples: list of negtive samples
+        """
+        np.random.shuffle(pos_samples)
+        num_sample = len(pos_samples)
+        neg_samples = []
+        miss_num = 0
+
+        def split_sent(sample, max_len, sep_id):
+            token_seq, type_seq, pos_seq, label, seg_labels = sample
+            sep_index = token_seq.index(sep_id)
+            left_len = sep_index - 1
+            if left_len <= max_len:
+                return (token_seq[1:sep_index], seg_labels[1:sep_index])
+            else:
+                return [token_seq[sep_index + 1: -1], seg_labels[sep_index + 1 : -1]]
+
+        for i in range(num_sample):
+            pair_index = (i + 1) % num_sample
+            left_tokens, left_seg_labels = split_sent(pos_samples[i],
+                    (self.max_seq_len - 3) // 2, self.sep_id)
+            right_tokens, right_seg_labels = split_sent(pos_samples[pair_index],
+                    self.max_seq_len - 3 - len(left_tokens), self.sep_id)
+
+            token_seq = [self.cls_id] + left_tokens + [self.sep_id] + \
+                    right_tokens + [self.sep_id]
+            if len(token_seq) > self.max_seq_len:
+                miss_num += 1
+                continue
+            type_seq = [0] * (len(left_tokens) + 2) + [1] * (len(right_tokens) + 1)
+            pos_seq = range(len(token_seq))
+            seg_label_seq = [-1] + left_seg_labels + [-1] + right_seg_labels + [-1]
+
+            assert len(token_seq) == len(type_seq) == len(pos_seq) == len(seg_label_seq), \
+                    "[ERROR]len(src_id) == lne(sent_id) == len(pos_id) must be True"
+            neg_samples.append([token_seq, type_seq, pos_seq, 0, seg_label_seq])
+
+        return neg_samples, miss_num
+
+    def mixin_negtive_samples(self, pos_sample_generator, buffer=1000):
+        """ 1. generate negtive samples by randomly group sentence_1 and sentence_2 of positive samples
+            2. combine negtive samples and positive samples
+            
+            Args:
+                pos_sample_generator: a generator producing a parsed positive sample, which is a list: [token_ids, sent_ids, pos_ids, 1]
+
+            Returns:
+                sample: one sample from shuffled positive samples and negtive samples
+        """
+        pos_samples = []
+        num_total_miss = 0
+        pos_sample_num = 0
+        try:
+            while True:
+                while len(pos_samples) < buffer:
+                    pos_sample = next(pos_sample_generator)
+                    label = pos_sample[3]
+                    assert label == 1, "positive sample's label must be 1"
+                    pos_samples.append(pos_sample)
+                    pos_sample_num += 1
+
+                neg_samples, miss_num = self.random_pair_neg_samples(
+                    pos_samples)
+                num_total_miss += miss_num
+                samples = pos_samples + neg_samples
+                pos_samples = []
+                np.random.shuffle(samples)
+                for sample in samples:
+                    yield sample
+        except StopIteration:
+            print("stopiteration: reach end of file")
+            if len(pos_samples) == 1:
+                yield pos_samples[0]
+            elif len(pos_samples) == 0:
+                yield None
+            else:
+                neg_samples, miss_num = self.random_pair_neg_samples(
+                    pos_samples)
+                num_total_miss += miss_num
+                samples = pos_samples + neg_samples
+                pos_samples = []
+                np.random.shuffle(samples)
+                for sample in samples:
+                    yield sample
+            print("miss_num:%d\tideal_total_sample_num:%d\tmiss_rate:%f" %
+                  (num_total_miss, pos_sample_num * 2,
+                   num_total_miss / (pos_sample_num * 2)))
+
+    def data_generator(self):
+        """
+        data_generator
+        """
+        files = open(self.filelist).readlines()
+        self.total_file = len(files)
+        assert self.total_file > 0, "[Error] data_dir is empty"
+
+        def wrapper():
+            def reader():
+                for epoch in range(self.epoch):
+                    self.current_epoch = epoch + 1
+                    if self.shuffle_files:
+                        np.random.shuffle(files)
+                    for index, file in enumerate(files):
+                        file, mask_word_prob = file.strip().split("\t")
+                        mask_word = (np.random.random() < float(mask_word_prob))
+                        self.current_file_index = index + 1
+                        self.current_file = file
+                        if mask_word:
+                            self.mask_type = "mask_word"
+                        else:
+                            self.mask_type = "mask_char"
+
+                        sample_generator = self.read_file(file)
+                        if not self.is_test and self.generate_neg_sample:
+                            sample_generator = self.mixin_negtive_samples(
+                                sample_generator)
+                        for sample in sample_generator:
+                            if sample is None:
+                                continue
+                            sample.append(mask_word)
+                            yield sample
+
+            def batch_reader(reader, batch_size):
+                batch, total_token_num, max_len = [], 0, 0
+                for parsed_line in reader():
+                    token_ids, sent_ids, pos_ids, label, seg_labels, mask_word = parsed_line
+                    max_len = max(max_len, len(token_ids))
+                    if (len(batch) + 1) * max_len <= batch_size:
+                        batch.append(parsed_line)
+                        total_token_num += len(token_ids)
+                    else:
+                        yield batch, total_token_num
+                        batch, total_token_num, max_len = [parsed_line], len(
+                            token_ids), len(token_ids)
+
+                if len(batch) > 0:
+                    yield batch, total_token_num
+
+            for batch_data, total_token_num in batch_reader(reader,
+                                                            self.batch_size):
+                yield prepare_batch_data(
+                    batch_data,
+                    total_token_num,
+                    voc_size=self.voc_size,
+                    pad_id=self.pad_id,
+                    cls_id=self.cls_id,
+                    sep_id=self.sep_id,
+                    mask_id=self.mask_id,
+                    return_attn_bias=True,
+                    return_max_len=False,
+                    return_num_token=False)
+
+        return wrapper
+
+
+if __name__ == "__main__":
+    pass

ERNIE/reader/task_reader.py

+#   Copyright (c) 2019 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 os
+import csv
+import json
+import numpy as np
+from collections import namedtuple
+
+import tokenization
+from batching import pad_batch_data
+
+
+class BaseReader(object):
+    def __init__(self,
+                 vocab_path,
+                 label_map_config=None,
+                 max_seq_len=512,
+                 do_lower_case=True,
+                 in_tokens=False,
+                 random_seed=None):
+        self.max_seq_len = max_seq_len
+        self.tokenizer = tokenization.FullTokenizer(
+            vocab_file=vocab_path, do_lower_case=do_lower_case)
+        self.vocab = self.tokenizer.vocab
+        self.pad_id = self.vocab["[PAD]"]
+        self.cls_id = self.vocab["[CLS]"]
+        self.sep_id = self.vocab["[SEP]"]
+        self.in_tokens = in_tokens
+
+        np.random.seed(random_seed)
+
+        self.current_example = 0
+        self.current_epoch = 0
+        self.num_examples = 0
+
+        if label_map_config:
+            with open(label_map_config) as f:
+                self.label_map = json.load(f)
+        else:
+            self.label_map = None
+        pass
+
+    def get_train_progress(self):
+        """Gets progress for training phase."""
+        return self.current_example, self.current_epoch
+
+    def _read_tsv(self, input_file, quotechar=None):
+        """Reads a tab separated value file."""
+        with open(input_file, "r") as f:
+            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
+            headers = next(reader)
+            Example = namedtuple('Example', headers)
+                
+            examples = []
+            for line in reader:
+                example = Example(*line)
+                examples.append(example)
+            return examples
+
+    def _truncate_seq_pair(self, tokens_a, tokens_b, max_length):
+        """Truncates a sequence pair in place to the maximum length."""
+
+        # This is a simple heuristic which will always truncate the longer sequence
+        # one token at a time. This makes more sense than truncating an equal percent
+        # of tokens from each, since if one sequence is very short then each token
+        # that's truncated likely contains more information than a longer sequence.
+        while True:
+            total_length = len(tokens_a) + len(tokens_b)
+            if total_length <= max_length:
+                break
+            if len(tokens_a) > len(tokens_b):
+                tokens_a.pop()
+            else:
+                tokens_b.pop()
+
+    def _convert_example_to_record(self, example,  max_seq_length, tokenizer):
+        """Converts a single `Example` into a single `Record`."""
+
+        text_a = tokenization.convert_to_unicode(example.text_a)
+        tokens_a = tokenizer.tokenize(text_a)
+        tokens_b = None
+        if "text_b" in example._fields:
+            text_b = tokenization.convert_to_unicode(example.text_b)
+            tokens_b = tokenizer.tokenize(text_b)
+
+        if tokens_b:
+            # Modifies `tokens_a` and `tokens_b` in place so that the total
+            # length is less than the specified length.
+            # Account for [CLS], [SEP], [SEP] with "- 3"
+            self._truncate_seq_pair(tokens_a, tokens_b, max_seq_length - 3)
+        else:
+            # Account for [CLS] and [SEP] with "- 2"
+            if len(tokens_a) > max_seq_length - 2:
+                tokens_a = tokens_a[0:(max_seq_length - 2)]
+
+        # The convention in BERT/ERNIE is:
+        # (a) For sequence pairs:
+        #  tokens:   [CLS] is this jack ##son ##ville ? [SEP] no it is not . [SEP]
+        #  type_ids: 0     0  0    0    0     0       0 0     1  1  1  1   1 1
+        # (b) For single sequences:
+        #  tokens:   [CLS] the dog is hairy . [SEP]
+        #  type_ids: 0     0   0   0  0     0 0
+        #
+        # Where "type_ids" are used to indicate whether this is the first
+        # sequence or the second sequence. The embedding vectors for `type=0` and
+        # `type=1` were learned during pre-training and are added to the wordpiece
+        # embedding vector (and position vector). This is not *strictly* necessary
+        # since the [SEP] token unambiguously separates the sequences, but it makes
+        # it easier for the model to learn the concept of sequences.
+        #
+        # For classification tasks, the first vector (corresponding to [CLS]) is
+        # used as as the "sentence vector". Note that this only makes sense because
+        # the entire model is fine-tuned.
+        tokens = []
+        text_type_ids = []
+        tokens.append("[CLS]")
+        text_type_ids.append(0)
+        for token in tokens_a:
+            tokens.append(token)
+            text_type_ids.append(0)
+        tokens.append("[SEP]")
+        text_type_ids.append(0)
+
+        if tokens_b:
+            for token in tokens_b:
+                tokens.append(token)
+                text_type_ids.append(1)
+            tokens.append("[SEP]")
+            text_type_ids.append(1)
+
+        token_ids = tokenizer.convert_tokens_to_ids(tokens)
+        position_ids = list(range(len(token_ids)))
+
+        if self.label_map:
+            label_id = self.label_map[example.label]
+        else:
+            label_id = example.label
+
+        Record = namedtuple('Record', ['token_ids', 'text_type_ids', 'position_ids', 'label_id', 'qid'])
+
+        qid = None
+        if "qid" in example._fields:
+            qid = example.qid
+
+        record = Record(
+            token_ids=token_ids,
+            text_type_ids=text_type_ids,
+            position_ids=position_ids,
+            label_id=label_id,
+            qid=qid)
+        return record
+
+    def _prepare_batch_data(self, examples, batch_size, phase=None):
+        """generate batch records"""
+        batch_records,  max_len = [], 0
+        for index, example in enumerate(examples):
+            if phase == "train":
+                self.current_example = index
+            record = self._convert_example_to_record(example, self.max_seq_len, self.tokenizer)
+            max_len = max(max_len, len(record.token_ids))
+            if self.in_tokens:
+                to_append = (len(batch_records) + 1) * max_len <= batch_size
+            else:
+                to_append = len(batch_records) < batch_size
+            if to_append:
+                batch_records.append(record)
+            else:
+                yield self._pad_batch_records(batch_records)
+                batch_records, max_len = [record], len(record.token_ids)
+
+        if len(batch_records) > 0:
+            yield self._pad_batch_records(batch_records)
+
+    def get_num_examples(self, input_file):
+        examples = self._read_tsv(input_file)
+        return len(examples)
+
+    def data_generator(self, input_file, batch_size, epoch, shuffle=True, phase=None):
+        examples = self._read_tsv(input_file)
+
+        def wrapper():
+            for epoch_index in range(epoch):
+                if phase == "train":
+                    self.current_example = 0
+                    self.current_epoch = epoch_index
+                if shuffle:
+                    np.random.shuffle(examples)
+
+                for batch_data in self._prepare_batch_data(examples, batch_size, phase=phase):
+                    yield batch_data
+        return wrapper
+
+
+class ClassifyReader(BaseReader):
+    def _read_tsv(self, input_file, quotechar=None):
+        """Reads a tab separated value file."""
+        with open(input_file, "r") as f:
+            reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
+            headers = next(reader)
+            text_indices = [index for index, h in enumerate(headers) if h != "label"]
+            Example = namedtuple('Example', headers)
+            
+            examples = []
+            for line in reader:
+                for index, text in enumerate(line):
+                    if index in text_indices:
+                        line[index] = text.replace(' ', '')
+                example = Example(*line)
+                examples.append(example)
+            return examples 
+    
+    def _pad_batch_records(self, batch_records):
+        batch_token_ids = [record.token_ids for record in batch_records]
+        batch_text_type_ids = [record.text_type_ids for record in batch_records]
+        batch_position_ids = [record.position_ids for record in batch_records]
+        batch_labels = [record.label_id for record in batch_records]
+        batch_labels = np.array(batch_labels).astype("int64").reshape([-1, 1])
+
+        if batch_records[0].qid:
+            batch_qids = [record.qid for record in batch_records]
+            batch_qids = np.array(batch_qids).astype("int64").reshape([-1, 1])
+        else:
+            batch_qids = np.array([]).astype("int64").reshape([-1, 1])
+
+        # padding
+        padded_token_ids, next_sent_index, self_attn_bias = pad_batch_data(
+            batch_token_ids, pad_idx=self.pad_id, return_next_sent_pos=True, return_attn_bias=True)
+        padded_text_type_ids = pad_batch_data(batch_text_type_ids, pad_idx=self.pad_id)
+        padded_position_ids = pad_batch_data(batch_position_ids, pad_idx=self.pad_id)
+
+        return_list = [padded_token_ids, padded_text_type_ids, padded_position_ids, self_attn_bias, batch_labels, next_sent_index, batch_qids]
+
+        return return_list
+
+
+class SequenceLabelReader(BaseReader):
+
+    def _pad_batch_records(self, batch_records):
+        batch_token_ids = [record.token_ids for record in batch_records]
+        batch_text_type_ids = [record.text_type_ids for record in batch_records]
+        batch_position_ids = [record.position_ids for record in batch_records]
+        batch_label_ids = [record.label_ids for record in batch_records]
+        batch_seq_lens =  [len(record.token_ids) for record in batch_records]
+
+        # padding
+        padded_token_ids, self_attn_bias = pad_batch_data(
+            batch_token_ids, pad_idx=self.pad_id, return_next_sent_pos=False, return_attn_bias=True)
+        padded_text_type_ids = pad_batch_data(batch_text_type_ids, pad_idx=self.pad_id)
+        padded_position_ids = pad_batch_data(batch_position_ids, pad_idx=self.pad_id)
+        padded_label_ids = pad_batch_data(batch_label_ids, pad_idx=len(self.label_map)-1)
+        batch_seq_lens = np.array(batch_seq_lens).astype("int64").reshape([-1, 1])
+
+        return_list = [padded_token_ids, padded_text_type_ids, padded_position_ids, self_attn_bias, padded_label_ids, batch_seq_lens]
+        return return_list
+
+    def _reseg_token_label(self, tokens, labels, tokenizer):
+        assert len(tokens) == len(labels)
+        ret_tokens = []
+        ret_labels = []
+        for token, label in zip(tokens, labels):
+            sub_token = tokenizer.tokenize(token)
+            if len(sub_token) == 0:
+                continue
+            ret_tokens.extend(sub_token)
+            ret_labels.append(label)
+            if len(sub_token) < 2:
+                continue
+            sub_label = label
+            if label.startswith("B-"):
+                sub_label = "I-" + label[2:]
+            ret_labels.extend([sub_label] * (len(sub_token) - 1))
+
+        assert len(ret_tokens) == len(ret_labels)
+        return ret_tokens, ret_labels
+
+    def _convert_example_to_record(self, example,  max_seq_length, tokenizer):
+        tokens = tokenization.convert_to_unicode(example.text_a).split(u"")
+        labels = tokenization.convert_to_unicode(example.label).split(u"")
+        tokens, labels = self._reseg_token_label(tokens, labels, tokenizer)
+
+        if len(tokens) > max_seq_length - 2:
+            tokens = tokens[0:(max_seq_length - 2)]
+            labels = labels[0:(max_seq_length - 2)]
+
+        tokens = ["[CLS]"] + tokens + ["[SEP]"]
+        token_ids = tokenizer.convert_tokens_to_ids(tokens)
+        position_ids = list(range(len(token_ids)))
+        text_type_ids = [0] * len(token_ids) 
+        no_entity_id = len(self.label_map) - 1
+        label_ids = [no_entity_id] + [self.label_map[label] for label in labels] + [no_entity_id]
+
+        Record = namedtuple('Record', ['token_ids', 'text_type_ids', 'position_ids', 'label_ids'])
+        record = Record(
+            token_ids=token_ids,
+            text_type_ids=text_type_ids,
+            position_ids=position_ids,
+            label_ids=label_ids)
+        return record
+
+if __name__ == '__main__':
+    pass

ERNIE/run_classifier.py

+#   Copyright (c) 2019 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.
+"""Finetuning on classification tasks."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import argparse
+import numpy as np
+import multiprocessing
+
+import paddle
+import paddle.fluid as fluid
+
+import reader.task_reader as task_reader
+from model.ernie import ErnieConfig
+from finetune.classifier import create_model, evaluate
+from optimization import optimization
+from utils.args import ArgumentGroup, print_arguments
+from utils.init import init_pretraining_params, init_checkpoint
+from finetune_args import parser 
+
+
+args = parser.parse_args()
+
+def main(args):
+    ernie_config = ErnieConfig(args.ernie_config_path)
+    ernie_config.print_config()
+
+    if args.use_cuda:
+        place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
+        dev_count = fluid.core.get_cuda_device_count()
+    else:
+        place = fluid.CPUPlace()
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+    exe = fluid.Executor(place)
+
+    reader = task_reader.ClassifyReader(vocab_path=args.vocab_path,
+                                label_map_config=args.label_map_config,
+                                max_seq_len=args.max_seq_len,
+                                do_lower_case=args.do_lower_case,
+                                in_tokens=args.in_tokens,
+                                random_seed=args.random_seed)
+
+    if not (args.do_train or args.do_val or args.do_test):
+        raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
+                         "least one of them must be True.")
+
+    startup_prog = fluid.Program()
+    if args.random_seed is not None:
+        startup_prog.random_seed = args.random_seed
+
+    if args.do_train:
+        train_data_generator = reader.data_generator(
+            input_file=args.train_set,
+            batch_size=args.batch_size,
+            epoch=args.epoch,
+            shuffle=True,
+            phase="train")
+
+        num_train_examples = reader.get_num_examples(args.train_set)
+
+        if args.in_tokens:
+            max_train_steps = args.epoch * num_train_examples // (
+                args.batch_size // args.max_seq_len) // dev_count
+        else:
+            max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
+
+        warmup_steps = int(max_train_steps * args.warmup_proportion)
+        print("Device count: %d" % dev_count)
+        print("Num train examples: %d" % num_train_examples)
+        print("Max train steps: %d" % max_train_steps)
+        print("Num warmup steps: %d" % warmup_steps)
+
+        train_program = fluid.Program()
+
+        with fluid.program_guard(train_program, startup_prog):
+            with fluid.unique_name.guard():
+                train_pyreader, graph_vars = create_model(
+                    args,
+                    pyreader_name='train_reader',
+                    ernie_config=ernie_config)
+                scheduled_lr = optimization(
+                    loss=graph_vars["loss"],
+                    warmup_steps=warmup_steps,
+                    num_train_steps=max_train_steps,
+                    learning_rate=args.learning_rate,
+                    train_program=train_program,
+                    startup_prog=startup_prog,
+                    weight_decay=args.weight_decay,
+                    scheduler=args.lr_scheduler,
+                    use_fp16=args.use_fp16,
+                    loss_scaling=args.loss_scaling)
+
+                fluid.memory_optimize(
+                    input_program=train_program,
+                    skip_opt_set=[graph_vars["loss"].name,
+                                  graph_vars["probs"].name,
+                                  graph_vars["accuracy"].name,
+                                  graph_vars["num_seqs"].name,
+                    ])
+
+        if args.verbose:
+            if args.in_tokens:
+                lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
+                    program=train_program,
+                    batch_size=args.batch_size // args.max_seq_len)
+            else:
+                lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
+                    program=train_program, batch_size=args.batch_size)
+            print("Theoretical memory usage in training: %.3f - %.3f %s" %
+                  (lower_mem, upper_mem, unit))
+
+    if args.do_val or args.do_test:
+        test_prog = fluid.Program()
+        with fluid.program_guard(test_prog, startup_prog):
+            with fluid.unique_name.guard():
+                test_pyreader, graph_vars = create_model(
+                    args,
+                    pyreader_name='test_reader',
+                    ernie_config=ernie_config)
+
+        test_prog = test_prog.clone(for_test=True)
+
+    exe.run(startup_prog)
+
+    if args.do_train:
+        if args.init_checkpoint and args.init_pretraining_params:
+            print(
+                "WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
+                "both are set! Only arg 'init_checkpoint' is made valid.")
+        if args.init_checkpoint:
+            init_checkpoint(
+                exe,
+                args.init_checkpoint,
+                main_program=startup_prog,
+                use_fp16=args.use_fp16)
+        elif args.init_pretraining_params:
+            init_pretraining_params(
+                exe,
+                args.init_pretraining_params,
+                main_program=startup_prog,
+                use_fp16=args.use_fp16)
+    elif args.do_val or args.do_test:
+        if not args.init_checkpoint:
+            raise ValueError("args 'init_checkpoint' should be set if"
+                             "only doing validation or testing!")
+        init_checkpoint(
+            exe,
+            args.init_checkpoint,
+            main_program=startup_prog,
+            use_fp16=args.use_fp16)
+
+    if args.do_train:
+        exec_strategy = fluid.ExecutionStrategy()
+        if args.use_fast_executor:
+            exec_strategy.use_experimental_executor = True
+        exec_strategy.num_threads = dev_count
+        exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
+
+        train_exe = fluid.ParallelExecutor(
+            use_cuda=args.use_cuda,
+            loss_name=graph_vars["loss"].name,
+            exec_strategy=exec_strategy,
+            main_program=train_program)
+
+        train_pyreader.decorate_tensor_provider(train_data_generator)
+    else:
+        train_exe = None
+
+    if args.do_val or args.do_test:
+        test_exe = fluid.ParallelExecutor(
+            use_cuda=args.use_cuda,
+            main_program=test_prog,
+            share_vars_from=train_exe)
+
+    if args.do_train:
+        train_pyreader.start()
+        steps = 0
+        if warmup_steps > 0:
+            graph_vars["learning_rate"] = scheduled_lr
+
+        time_begin = time.time()
+        while True:
+            try:
+                steps += 1
+                if steps % args.skip_steps != 0:
+                    train_exe.run(fetch_list=[])
+                else:
+                    outputs = evaluate(train_exe, train_program, train_pyreader, graph_vars, "train") 
+
+                    if args.verbose:
+                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
+                        )
+                        verbose += "learning rate: %f" % (
+                            outputs["learning_rate"]
+                            if warmup_steps > 0 else args.learning_rate)
+                        print(verbose)
+
+                    current_example, current_epoch = reader.get_train_progress()
+                    time_end = time.time()
+                    used_time = time_end - time_begin
+                    print("epoch: %d, progress: %d/%d, step: %d, ave loss: %f, "
+                          "ave acc: %f, speed: %f steps/s" %
+                          (current_epoch, current_example, num_train_examples,
+                           steps, outputs["loss"], outputs["accuracy"], args.skip_steps / used_time))
+                    time_begin = time.time()
+
+                if steps % args.save_steps == 0:
+                    save_path = os.path.join(args.checkpoints,
+                                             "step_" + str(steps))
+                    fluid.io.save_persistables(exe, save_path, train_program)
+
+                if steps % args.validation_steps == 0:
+                    # evaluate dev set
+                    if args.do_val:
+                        test_pyreader.decorate_tensor_provider(
+                            reader.data_generator(
+                                args.dev_set,
+                                batch_size=args.batch_size,
+                                epoch=1,
+                                shuffle=False))
+                        evaluate(exe, test_prog, test_pyreader, graph_vars, "dev")
+                    # evaluate test set
+                    if args.do_test:
+                        test_pyreader.decorate_tensor_provider(
+                            reader.data_generator(
+                                args.test_set,
+                                batch_size=args.batch_size,
+                                epoch=1,
+                                shuffle=False))
+                        evaluate(exe, test_prog, test_pyreader, graph_vars, "test")
+            except fluid.core.EOFException:
+                save_path = os.path.join(args.checkpoints, "step_" + str(steps))
+                fluid.io.save_persistables(exe, save_path, train_program)
+                train_pyreader.reset()
+                break
+
+    # final eval on dev set
+    if args.do_val:
+        test_pyreader.decorate_tensor_provider(
+            reader.data_generator(
+                args.dev_set, batch_size=args.batch_size, epoch=1,
+                shuffle=False))
+        print("Final validation result:")
+        evaluate(exe, test_prog, test_pyreader, graph_vars, "dev")
+
+    # final eval on test set
+    if args.do_test:
+        test_pyreader.decorate_tensor_provider(
+            reader.data_generator(
+                args.test_set,
+                batch_size=args.batch_size,
+                epoch=1,
+                shuffle=False))
+        print("Final test result:")
+        evaluate(exe, test_prog, test_pyreader, graph_vars, "test")
+
+
+if __name__ == '__main__':
+    print_arguments(args)
+    main(args)

ERNIE/run_sequence_labeling.py

+#   Copyright (c) 2019 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.
+"""Finetuning on classification tasks."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import numpy as np
+import multiprocessing
+
+import paddle
+import paddle.fluid as fluid
+
+import reader.task_reader as task_reader
+from model.ernie import ErnieConfig
+from optimization import optimization
+from utils.init import init_pretraining_params, init_checkpoint
+from utils.args import print_arguments
+from finetune.sequence_label import create_model, evaluate 
+from finetune_args import parser 
+
+
+args = parser.parse_args()
+
+def main(args):
+    ernie_config = ErnieConfig(args.ernie_config_path)
+    ernie_config.print_config()
+
+    if args.use_cuda:
+        place = fluid.CUDAPlace(int(os.getenv('FLAGS_selected_gpus', '0')))
+        dev_count = fluid.core.get_cuda_device_count()
+    else:
+        place = fluid.CPUPlace()
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+    exe = fluid.Executor(place)
+
+    reader = task_reader.SequenceLabelReader(vocab_path=args.vocab_path,
+                                label_map_config=args.label_map_config,
+                                max_seq_len=args.max_seq_len,
+                                do_lower_case=args.do_lower_case,
+                                in_tokens=args.in_tokens,
+                                random_seed=args.random_seed)
+
+    if not (args.do_train or args.do_val or args.do_test):
+        raise ValueError("For args `do_train`, `do_val` and `do_test`, at "
+                         "least one of them must be True.")
+
+    startup_prog = fluid.Program()
+    if args.random_seed is not None:
+        startup_prog.random_seed = args.random_seed
+
+    if args.do_train:
+        train_data_generator = reader.data_generator(
+            input_file=args.train_set,
+            batch_size=args.batch_size,
+            epoch=args.epoch,
+            shuffle=True,
+            phase="train")
+
+        num_train_examples = reader.get_num_examples(args.train_set)
+
+        if args.in_tokens:
+            max_train_steps = args.epoch * num_train_examples // (
+                args.batch_size // args.max_seq_len) // dev_count
+        else:
+            max_train_steps = args.epoch * num_train_examples // args.batch_size // dev_count
+
+        warmup_steps = int(max_train_steps * args.warmup_proportion)
+        print("Device count: %d" % dev_count)
+        print("Num train examples: %d" % num_train_examples)
+        print("Max train steps: %d" % max_train_steps)
+        print("Num warmup steps: %d" % warmup_steps)
+
+        train_program = fluid.Program()
+
+        with fluid.program_guard(train_program, startup_prog):
+            with fluid.unique_name.guard():
+                train_pyreader, graph_vars = create_model(
+                    args,
+                    pyreader_name='train_reader',
+                    ernie_config=ernie_config)
+                scheduled_lr = optimization(
+                    loss=graph_vars["loss"],
+                    warmup_steps=warmup_steps,
+                    num_train_steps=max_train_steps,
+                    learning_rate=args.learning_rate,
+                    train_program=train_program,
+                    startup_prog=startup_prog,
+                    weight_decay=args.weight_decay,
+                    scheduler=args.lr_scheduler,
+                    use_fp16=args.use_fp16,
+                    loss_scaling=args.loss_scaling)
+
+                fluid.memory_optimize(
+                    input_program=train_program,
+                    skip_opt_set=[graph_vars["loss"].name,
+                                  graph_vars["labels"].name,
+                                  graph_vars["infers"].name,
+                                  graph_vars["seq_lens"].name
+                    ])
+
+        if args.verbose:
+            if args.in_tokens:
+                lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
+                    program=train_program,
+                    batch_size=args.batch_size // args.max_seq_len)
+            else:
+                lower_mem, upper_mem, unit = fluid.contrib.memory_usage(
+                    program=train_program, batch_size=args.batch_size)
+            print("Theoretical memory usage in training: %.3f - %.3f %s" %
+                  (lower_mem, upper_mem, unit))
+
+    if args.do_val or args.do_test:
+        test_prog = fluid.Program()
+        with fluid.program_guard(test_prog, startup_prog):
+            with fluid.unique_name.guard():
+                test_pyreader, graph_vars = create_model(
+                    args,
+                    pyreader_name='test_reader',
+                    ernie_config=ernie_config)
+
+        test_prog = test_prog.clone(for_test=True)
+
+    exe.run(startup_prog)
+
+    if args.do_train:
+        if args.init_checkpoint and args.init_pretraining_params:
+            print(
+                "WARNING: args 'init_checkpoint' and 'init_pretraining_params' "
+                "both are set! Only arg 'init_checkpoint' is made valid.")
+        if args.init_checkpoint:
+            init_checkpoint(
+                exe,
+                args.init_checkpoint,
+                main_program=startup_prog,
+                use_fp16=args.use_fp16)
+        elif args.init_pretraining_params:
+            init_pretraining_params(
+                exe,
+                args.init_pretraining_params,
+                main_program=startup_prog,
+                use_fp16=args.use_fp16)
+    elif args.do_val or args.do_test:
+        if not args.init_checkpoint:
+            raise ValueError("args 'init_checkpoint' should be set if"
+                             "only doing validation or testing!")
+        init_checkpoint(
+            exe,
+            args.init_checkpoint,
+            main_program=startup_prog,
+            use_fp16=args.use_fp16)
+
+    if args.do_train:
+        exec_strategy = fluid.ExecutionStrategy()
+        if args.use_fast_executor:
+            exec_strategy.use_experimental_executor = True
+        exec_strategy.num_threads = dev_count
+        exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
+
+        train_exe = fluid.ParallelExecutor(
+            use_cuda=args.use_cuda,
+            loss_name=graph_vars["loss"].name,
+            exec_strategy=exec_strategy,
+            main_program=train_program)
+
+        train_pyreader.decorate_tensor_provider(train_data_generator)
+    else:
+        train_exe = None
+
+    if args.do_val or args.do_test:
+        test_exe = fluid.ParallelExecutor(
+            use_cuda=args.use_cuda,
+            main_program=test_prog,
+            share_vars_from=train_exe)
+
+    if args.do_train:
+        train_pyreader.start()
+        steps = 0
+        if warmup_steps > 0:
+            graph_vars["learning_rate"] = scheduled_lr
+
+        time_begin = time.time()
+        while True:
+            try:
+                steps += 1
+                if steps % args.skip_steps != 0:
+                    train_exe.run(fetch_list=[])
+                else:
+                    outputs = evaluate(train_exe, train_program, train_pyreader, graph_vars, args.num_labels, "train", dev_count)
+                    if args.verbose:
+                        verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
+                        )
+                        verbose += "learning rate: %f" % (
+                            outputs["lr"] if warmup_steps > 0 else args.learning_rate)
+                        print(verbose)
+
+                    current_example, current_epoch = reader.get_train_progress(
+                    )
+                    time_end = time.time()
+                    used_time = time_end - time_begin
+                    print("epoch: %d, progress: %d/%d, step: %d, loss: %f, "
+                          "f1: %f, precision: %f, recall: %f, speed: %f steps/s" %
+                          (current_epoch, current_example, num_train_examples,
+                           steps, outputs["loss"], outputs["f1"],
+                           outputs["precision"], outputs["recall"],
+                           args.skip_steps / used_time))
+                    time_begin = time.time()
+
+                if steps % args.save_steps == 0:
+                    save_path = os.path.join(args.checkpoints,
+                                             "step_" + str(steps))
+                    fluid.io.save_persistables(exe, save_path, train_program)
+
+                if steps % args.validation_steps == 0:
+                    # evaluate dev set
+                    if args.do_val:
+                        test_pyreader.decorate_tensor_provider(
+                            reader.data_generator(
+                                args.dev_set,
+                                batch_size=args.batch_size,
+                                epoch=1,
+                                shuffle=False))
+                        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "dev")
+                    # evaluate test set
+                    if args.do_test:
+                        test_pyreader.decorate_tensor_provider(
+                            reader.data_generator(
+                                args.test_set,
+                                batch_size=args.batch_size,
+                                epoch=1,
+                                shuffle=False))
+                        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "test")
+
+            except fluid.core.EOFException:
+                save_path = os.path.join(args.checkpoints, "step_" + str(steps))
+                fluid.io.save_persistables(exe, save_path, train_program)
+                train_pyreader.reset()
+                break
+
+    # final eval on dev set
+    if args.do_val:
+        test_pyreader.decorate_tensor_provider(
+            reader.data_generator(
+                args.dev_set, batch_size=args.batch_size, epoch=1,
+                shuffle=False))
+        print("Final validation result:")
+        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "dev")
+
+    # final eval on test set
+    if args.do_test:
+        test_pyreader.decorate_tensor_provider(
+            reader.data_generator(
+                args.test_set,
+                batch_size=args.batch_size,
+                epoch=1,
+                shuffle=False))
+        print("Final test result:")
+        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "test")
+
+
+if __name__ == '__main__':
+    print_arguments(args)
+    main(args)

ERNIE/tokenization.py

+# coding=utf-8
+# Copyright 2018 The Google AI Language Team Authors.
+#
+# 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.
+"""Tokenization classes."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import unicodedata
+import six
+
+
+def convert_to_unicode(text):
+    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text.decode("utf-8", "ignore")
+        elif isinstance(text, unicode):
+            return text
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def printable_text(text):
+    """Returns text encoded in a way suitable for print or `tf.logging`."""
+
+    # These functions want `str` for both Python2 and Python3, but in one case
+    # it's a Unicode string and in the other it's a byte string.
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, unicode):
+            return text.encode("utf-8")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")
+
+
+def load_vocab(vocab_file):
+    """Loads a vocabulary file into a dictionary."""
+    vocab = collections.OrderedDict()
+    fin = open(vocab_file)
+    for num, line in enumerate(fin):
+        items = convert_to_unicode(line.strip()).split("\t")
+        if len(items) > 2:
+            break
+        token = items[0]
+        index = items[1] if len(items) == 2 else num
+        token = token.strip()
+        vocab[token] = int(index)
+    return vocab
+
+
+def convert_by_vocab(vocab, items):
+    """Converts a sequence of [tokens|ids] using the vocab."""
+    output = []
+    for item in items:
+        output.append(vocab[item])
+    return output
+
+
+def convert_tokens_to_ids(vocab, tokens):
+    return convert_by_vocab(vocab, tokens)
+
+
+def convert_ids_to_tokens(inv_vocab, ids):
+    return convert_by_vocab(inv_vocab, ids)
+
+
+def whitespace_tokenize(text):
+    """Runs basic whitespace cleaning and splitting on a peice of text."""
+    text = text.strip()
+    if not text:
+        return []
+    tokens = text.split()
+    return tokens
+
+
+class FullTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.basic_tokenizer = BasicTokenizer(do_lower_case=do_lower_case)
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in self.basic_tokenizer.tokenize(text):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class CharTokenizer(object):
+    """Runs end-to-end tokenziation."""
+
+    def __init__(self, vocab_file, do_lower_case=True):
+        self.vocab = load_vocab(vocab_file)
+        self.inv_vocab = {v: k for k, v in self.vocab.items()}
+        self.wordpiece_tokenizer = WordpieceTokenizer(vocab=self.vocab)
+
+    def tokenize(self, text):
+        split_tokens = []
+        for token in text.lower().split(" "):
+            for sub_token in self.wordpiece_tokenizer.tokenize(token):
+                split_tokens.append(sub_token)
+
+        return split_tokens
+
+    def convert_tokens_to_ids(self, tokens):
+        return convert_by_vocab(self.vocab, tokens)
+
+    def convert_ids_to_tokens(self, ids):
+        return convert_by_vocab(self.inv_vocab, ids)
+
+
+class BasicTokenizer(object):
+    """Runs basic tokenization (punctuation splitting, lower casing, etc.)."""
+
+    def __init__(self, do_lower_case=True):
+        """Constructs a BasicTokenizer.
+
+        Args:
+            do_lower_case: Whether to lower case the input.
+        """
+        self.do_lower_case = do_lower_case
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text."""
+        text = convert_to_unicode(text)
+        text = self._clean_text(text)
+
+        # This was added on November 1st, 2018 for the multilingual and Chinese
+        # models. This is also applied to the English models now, but it doesn't
+        # matter since the English models were not trained on any Chinese data
+        # and generally don't have any Chinese data in them (there are Chinese
+        # characters in the vocabulary because Wikipedia does have some Chinese
+        # words in the English Wikipedia.).
+        text = self._tokenize_chinese_chars(text)
+
+        orig_tokens = whitespace_tokenize(text)
+        split_tokens = []
+        for token in orig_tokens:
+            if self.do_lower_case:
+                token = token.lower()
+                token = self._run_strip_accents(token)
+            split_tokens.extend(self._run_split_on_punc(token))
+
+        output_tokens = whitespace_tokenize(" ".join(split_tokens))
+        return output_tokens
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text):
+        """Splits punctuation on a piece of text."""
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def _tokenize_chinese_chars(self, text):
+        """Adds whitespace around any CJK character."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if self._is_chinese_char(cp):
+                output.append(" ")
+                output.append(char)
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+    def _is_chinese_char(self, cp):
+        """Checks whether CP is the codepoint of a CJK character."""
+        # This defines a "chinese character" as anything in the CJK Unicode block:
+        #     https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+        #
+        # Note that the CJK Unicode block is NOT all Japanese and Korean characters,
+        # despite its name. The modern Korean Hangul alphabet is a different block,
+        # as is Japanese Hiragana and Katakana. Those alphabets are used to write
+        # space-separated words, so they are not treated specially and handled
+        # like the all of the other languages.
+        if ((cp >= 0x4E00 and cp <= 0x9FFF) or  #
+            (cp >= 0x3400 and cp <= 0x4DBF) or  #
+            (cp >= 0x20000 and cp <= 0x2A6DF) or  #
+            (cp >= 0x2A700 and cp <= 0x2B73F) or  #
+            (cp >= 0x2B740 and cp <= 0x2B81F) or  #
+            (cp >= 0x2B820 and cp <= 0x2CEAF) or
+            (cp >= 0xF900 and cp <= 0xFAFF) or  #
+            (cp >= 0x2F800 and cp <= 0x2FA1F)):  #
+            return True
+
+        return False
+
+    def _clean_text(self, text):
+        """Performs invalid character removal and whitespace cleanup on text."""
+        output = []
+        for char in text:
+            cp = ord(char)
+            if cp == 0 or cp == 0xfffd or _is_control(char):
+                continue
+            if _is_whitespace(char):
+                output.append(" ")
+            else:
+                output.append(char)
+        return "".join(output)
+
+
+class WordpieceTokenizer(object):
+    """Runs WordPiece tokenziation."""
+
+    def __init__(self, vocab, unk_token="[UNK]", max_input_chars_per_word=100):
+        self.vocab = vocab
+        self.unk_token = unk_token
+        self.max_input_chars_per_word = max_input_chars_per_word
+
+    def tokenize(self, text):
+        """Tokenizes a piece of text into its word pieces.
+
+        This uses a greedy longest-match-first algorithm to perform tokenization
+        using the given vocabulary.
+
+        For example:
+            input = "unaffable"
+            output = ["un", "##aff", "##able"]
+
+        Args:
+            text: A single token or whitespace separated tokens. This should have
+                already been passed through `BasicTokenizer.
+
+        Returns:
+            A list of wordpiece tokens.
+        """
+
+        text = convert_to_unicode(text)
+
+        output_tokens = []
+        for token in whitespace_tokenize(text):
+            chars = list(token)
+            if len(chars) > self.max_input_chars_per_word:
+                output_tokens.append(self.unk_token)
+                continue
+
+            is_bad = False
+            start = 0
+            sub_tokens = []
+            while start < len(chars):
+                end = len(chars)
+                cur_substr = None
+                while start < end:
+                    substr = "".join(chars[start:end])
+                    if start > 0:
+                        substr = "##" + substr
+                    if substr in self.vocab:
+                        cur_substr = substr
+                        break
+                    end -= 1
+                if cur_substr is None:
+                    is_bad = True
+                    break
+                sub_tokens.append(cur_substr)
+                start = end
+
+            if is_bad:
+                output_tokens.append(self.unk_token)
+            else:
+                output_tokens.extend(sub_tokens)
+        return output_tokens
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if ((cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or
+        (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126)):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False

ERNIE/train.py

+#   Copyright (c) 2019 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.
+"""ERNIE pretraining."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+import time
+import argparse
+import numpy as np
+import multiprocessing
+
+import paddle
+import paddle.fluid as fluid
+
+from reader.pretraining import ErnieDataReader
+from model.ernie import ErnieModel, ErnieConfig
+from optimization import optimization
+from utils.args import ArgumentGroup, print_arguments
+from utils.init import init_checkpoint, init_pretraining_params
+
+from pretrain_args import parser
+
+args = parser.parse_args()
+# yapf: enable.
+
+def create_model(pyreader_name, ernie_config):
+    pyreader = fluid.layers.py_reader(
+        capacity=70,
+        shapes=[[-1, args.max_seq_len, 1], [-1, args.max_seq_len, 1],
+                [-1, args.max_seq_len, 1],
+                [-1, args.max_seq_len, args.max_seq_len], [-1, 1], [-1, 1],
+                [-1, 1], [-1, 1]],
+        dtypes=[
+            'int64', 'int64', 'int64', 'float', 'int64', 'int64', 'int64',
+            'int64'
+        ],
+        lod_levels=[0, 0, 0, 0, 0, 0, 0, 0],
+        name=pyreader_name,
+        use_double_buffer=True)
+
+    (src_ids, pos_ids, sent_ids, self_attn_mask, mask_label, mask_pos, labels,
+     next_sent_index) = fluid.layers.read_file(pyreader)
+
+    ernie = ErnieModel(
+        src_ids=src_ids,
+        position_ids=pos_ids,
+        sentence_ids=sent_ids,
+        self_attn_mask=self_attn_mask,
+        config=ernie_config,
+        weight_sharing=args.weight_sharing,
+        use_fp16=args.use_fp16)
+
+    next_sent_acc, mask_lm_loss, total_loss = ernie.get_pretraining_output(
+        mask_label, mask_pos, labels, next_sent_index)
+
+    if args.use_fp16 and args.loss_scaling > 1.0:
+        total_loss *= args.loss_scaling
+
+    return pyreader, next_sent_acc, mask_lm_loss, total_loss
+
+
+def predict_wrapper(args,
+                    exe,
+                    ernie_config,
+                    test_prog=None,
+                    pyreader=None,
+                    fetch_list=None):
+    # Context to do validation.
+    filelist = args.test_filelist if args.do_test else args.valid_filelist
+    data_reader = ErnieDataReader(
+        filelist,
+        vocab_path=args.vocab_path,
+        batch_size=args.batch_size,
+        voc_size=ernie_config['vocab_size'],
+        shuffle_files=False,
+        epoch=1,
+        max_seq_len=args.max_seq_len,
+        is_test=True)
+
+    if args.do_test:
+        assert args.init_checkpoint is not None, "[FATAL] Please use --init_checkpoint '/path/to/checkpoints' \
+                                                  to specify you pretrained model checkpoints"
+
+        init_pretraining_params(exe, args.init_checkpoint, test_prog)
+
+    def predict(exe=exe, pyreader=pyreader):
+
+        pyreader.decorate_tensor_provider(data_reader.data_generator())
+        pyreader.start()
+
+        cost = 0
+        lm_cost = 0
+        acc = 0
+        steps = 0
+        time_begin = time.time()
+        while True:
+            try:
+                each_next_acc, each_mask_lm_cost, each_total_cost = exe.run(
+                    fetch_list=fetch_list, program=test_prog)
+                acc += each_next_acc
+                lm_cost += each_mask_lm_cost
+                cost += each_total_cost
+                steps += 1
+                if args.do_test and steps % args.skip_steps == 0:
+                    print("[test_set] steps: %d" % steps)
+
+            except fluid.core.EOFException:
+                pyreader.reset()
+                break
+
+        used_time = time.time() - time_begin
+        return cost, lm_cost, acc, steps, (args.skip_steps / used_time)
+
+    return predict
+
+
+def test(args):
+    ernie_config = ErnieConfig(args.ernie_config_path)
+    ernie_config.print_config()
+
+    test_prog = fluid.Program()
+    test_startup = fluid.Program()
+    with fluid.program_guard(test_prog, test_startup):
+        with fluid.unique_name.guard():
+            test_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
+                pyreader_name='test_reader', ernie_config=ernie_config)
+
+    test_prog = test_prog.clone(for_test=True)
+
+    place = fluid.CUDAPlace(0) if args.use_cuda == True else fluid.CPUPlace()
+    exe = fluid.Executor(place)
+    exe.run(test_startup)
+
+    predict = predict_wrapper(
+        args,
+        exe,
+        ernie_config,
+        test_prog=test_prog,
+        pyreader=test_pyreader,
+        fetch_list=[next_sent_acc.name, mask_lm_loss.name, total_loss.name])
+
+    print("test begin")
+    loss, lm_loss, acc, steps, speed = predict()
+    print(
+        "[test_set] loss: %f, global ppl: %f, next_sent_acc: %f, speed: %f steps/s"
+        % (np.mean(np.array(loss) / steps),
+           np.exp(np.mean(np.array(lm_loss) / steps)),
+           np.mean(np.array(acc) / steps), speed))
+
+
+def train(args):
+    print("pretraining start")
+    ernie_config = ErnieConfig(args.ernie_config_path)
+    ernie_config.print_config()
+
+    train_program = fluid.Program()
+    startup_prog = fluid.Program()
+    with fluid.program_guard(train_program, startup_prog):
+        with fluid.unique_name.guard():
+            train_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
+                pyreader_name='train_reader', ernie_config=ernie_config)
+            scheduled_lr = optimization(
+                loss=total_loss,
+                warmup_steps=args.warmup_steps,
+                num_train_steps=args.num_train_steps,
+                learning_rate=args.learning_rate,
+                train_program=train_program,
+                startup_prog=startup_prog,
+                weight_decay=args.weight_decay,
+                scheduler=args.lr_scheduler,
+                use_fp16=args.use_fp16,
+                loss_scaling=args.loss_scaling)
+
+            fluid.memory_optimize(
+                input_program=train_program,
+                skip_opt_set=[
+                    next_sent_acc.name, mask_lm_loss.name, total_loss.name
+                ])
+
+    test_prog = fluid.Program()
+    with fluid.program_guard(test_prog, startup_prog):
+        with fluid.unique_name.guard():
+            test_pyreader, next_sent_acc, mask_lm_loss, total_loss = create_model(
+                pyreader_name='test_reader', ernie_config=ernie_config)
+
+    test_prog = test_prog.clone(for_test=True)
+
+    if args.use_cuda:
+        place = fluid.CUDAPlace(0)
+        dev_count = fluid.core.get_cuda_device_count()
+    else:
+        place = fluid.CPUPlace()
+        dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+
+    print("Device count %d" % dev_count)
+    print("theoretical memory usage: ")
+    print(fluid.contrib.memory_usage(
+        program=train_program, batch_size=args.batch_size // args.max_seq_len))
+
+    nccl2_num_trainers = 1
+    nccl2_trainer_id = 0
+    print("args.is_distributed:", args.is_distributed)
+    if args.is_distributed:
+        worker_endpoints_env = os.getenv("worker_endpoints")
+        worker_endpoints = worker_endpoints_env.split(",")
+        trainers_num = len(worker_endpoints)
+        current_endpoint = os.getenv("current_endpoint")
+        trainer_id = worker_endpoints.index(current_endpoint)
+        if trainer_id == 0:
+            print("train_id == 0, sleep 60s")
+            time.sleep(60)
+        print("worker_endpoints:{} trainers_num:{} current_endpoint:{} \
+              trainer_id:{}"
+                            .format(worker_endpoints, trainers_num,
+                                    current_endpoint, trainer_id))
+
+        # prepare nccl2 env.
+        config = fluid.DistributeTranspilerConfig()
+        config.mode = "nccl2"
+        t = fluid.DistributeTranspiler(config=config)
+        t.transpile(
+            trainer_id,
+            trainers=worker_endpoints_env,
+            current_endpoint=current_endpoint,
+            program=train_program,
+            startup_program=startup_prog)
+        nccl2_num_trainers = trainers_num
+        nccl2_trainer_id = trainer_id
+
+    exe = fluid.Executor(place)
+    exe.run(startup_prog)
+
+    if args.init_checkpoint and args.init_checkpoint != "":
+        init_checkpoint(exe, args.init_checkpoint, train_program, args.use_fp16)
+
+    data_reader = ErnieDataReader(
+        filelist=args.train_filelist,
+        batch_size=args.batch_size,
+        vocab_path=args.vocab_path,
+        voc_size=ernie_config['vocab_size'],
+        epoch=args.epoch,
+        max_seq_len=args.max_seq_len,
+        generate_neg_sample=args.generate_neg_sample)
+
+    exec_strategy = fluid.ExecutionStrategy()
+    if args.use_fast_executor:
+        exec_strategy.use_experimental_executor = True
+    exec_strategy.num_threads = dev_count
+    exec_strategy.num_iteration_per_drop_scope = min(10, args.skip_steps)
+
+    build_strategy = fluid.BuildStrategy()
+    build_strategy.remove_unnecessary_lock = False
+
+    train_exe = fluid.ParallelExecutor(
+        use_cuda=args.use_cuda,
+        loss_name=total_loss.name,
+        build_strategy=build_strategy,
+        exec_strategy=exec_strategy,
+        main_program=train_program,
+        num_trainers=nccl2_num_trainers,
+        trainer_id=nccl2_trainer_id)
+
+    if args.valid_filelist and args.valid_filelist != "":
+        predict = predict_wrapper(
+            args,
+            exe,
+            ernie_config,
+            test_prog=test_prog,
+            pyreader=test_pyreader,
+            fetch_list=[
+                next_sent_acc.name, mask_lm_loss.name, total_loss.name
+            ])
+
+    train_pyreader.decorate_tensor_provider(data_reader.data_generator())
+    train_pyreader.start()
+    steps = 0
+    cost = []
+    lm_cost = []
+    acc = []
+    time_begin = time.time()
+    while steps < args.num_train_steps:
+        try:
+            steps += nccl2_num_trainers
+            skip_steps = args.skip_steps * nccl2_num_trainers
+
+            if nccl2_trainer_id != 0:
+                train_exe.run(fetch_list=[])
+                continue
+
+            if steps % skip_steps != 0:
+                train_exe.run(fetch_list=[])
+            else:
+                each_next_acc, each_mask_lm_cost, each_total_cost, np_lr = train_exe.run(
+                    fetch_list=[
+                        next_sent_acc.name, mask_lm_loss.name, total_loss.name,
+                        scheduled_lr.name
+                    ])
+                acc.extend(each_next_acc)
+                lm_cost.extend(each_mask_lm_cost)
+                cost.extend(each_total_cost)
+
+                print("feed_queue size", train_pyreader.queue.size())
+                time_end = time.time()
+                used_time = time_end - time_begin
+                epoch, current_file_index, total_file, current_file, mask_type = data_reader.get_progress(
+                )
+                print("current learning_rate:%f" % np_lr[0])
+                print("epoch: %d, progress: %d/%d, step: %d, loss: %f, "
+                      "ppl: %f, next_sent_acc: %f, speed: %f steps/s, file: %s, mask_type: %s"
+                      % (epoch, current_file_index, total_file, steps,
+                         np.mean(np.array(cost)),
+                         np.mean(np.exp(np.array(lm_cost))),
+                         np.mean(np.array(acc)), skip_steps / used_time,
+                         current_file, mask_type))
+                cost = []
+                lm_cost = []
+                acc = []
+                time_begin = time.time()
+
+            if steps % args.save_steps == 0:
+                save_path = os.path.join(args.checkpoints, "step_" + str(steps))
+                fluid.io.save_persistables(exe, save_path, train_program)
+
+            if args.valid_filelist and steps % args.validation_steps == 0:
+                vali_cost, vali_lm_cost, vali_acc, vali_steps, vali_speed = predict(
+                )
+                print("[validation_set] epoch: %d, step: %d, "
+                      "loss: %f, global ppl: %f, batch-averged ppl: %f, "
+                      "next_sent_acc: %f, speed: %f steps/s" %
+                      (epoch, steps,
+                       np.mean(np.array(vali_cost) / vali_steps),
+                       np.exp(np.mean(np.array(vali_lm_cost) / vali_steps)),
+                       np.mean(np.exp(np.array(vali_lm_cost) / vali_steps)),
+                       np.mean(np.array(vali_acc) / vali_steps), vali_speed))
+
+        except fluid.core.EOFException:
+            train_pyreader.reset()
+            break
+
+
+if __name__ == '__main__':
+    print_arguments(args)
+    if args.do_test:
+        test(args)
+    else:
+        train(args)

ERNIE/utils/args.py

+#   Copyright (c) 2019 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.
+"""Arguments for configuration."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import six
+import argparse
+
+
+def str2bool(v):
+    # because argparse does not support to parse "true, False" as python
+    # boolean directly
+    return v.lower() in ("true", "t", "1")
+
+
+class ArgumentGroup(object):
+    def __init__(self, parser, title, des):
+        self._group = parser.add_argument_group(title=title, description=des)
+
+    def add_arg(self, name, type, default, help, **kwargs):
+        type = str2bool if type == bool else type
+        self._group.add_argument(
+            "--" + name,
+            default=default,
+            type=type,
+            help=help + ' Default: %(default)s.',
+            **kwargs)
+
+
+def print_arguments(args):
+    print('-----------  Configuration Arguments -----------')
+    for arg, value in sorted(six.iteritems(vars(args))):
+        print('%s: %s' % (arg, value))
+    print('------------------------------------------------')

ERNIE/utils/fp16.py

+#   Copyright (c) 2019 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 paddle
+import paddle.fluid as fluid
+
+
+def cast_fp16_to_fp32(i, o, prog):
+    prog.global_block().append_op(
+        type="cast",
+        inputs={"X": i},
+        outputs={"Out": o},
+        attrs={
+            "in_dtype": fluid.core.VarDesc.VarType.FP16,
+            "out_dtype": fluid.core.VarDesc.VarType.FP32
+        })
+
+
+def cast_fp32_to_fp16(i, o, prog):
+    prog.global_block().append_op(
+        type="cast",
+        inputs={"X": i},
+        outputs={"Out": o},
+        attrs={
+            "in_dtype": fluid.core.VarDesc.VarType.FP32,
+            "out_dtype": fluid.core.VarDesc.VarType.FP16
+        })
+
+
+def copy_to_master_param(p, block):
+    v = block.vars.get(p.name, None)
+    if v is None:
+        raise ValueError("no param name %s found!" % p.name)
+    new_p = fluid.framework.Parameter(
+        block=block,
+        shape=v.shape,
+        dtype=fluid.core.VarDesc.VarType.FP32,
+        type=v.type,
+        lod_level=v.lod_level,
+        stop_gradient=p.stop_gradient,
+        trainable=p.trainable,
+        optimize_attr=p.optimize_attr,
+        regularizer=p.regularizer,
+        gradient_clip_attr=p.gradient_clip_attr,
+        error_clip=p.error_clip,
+        name=v.name + ".master")
+    return new_p
+
+
+def create_master_params_grads(params_grads, main_prog, startup_prog,
+                               loss_scaling):
+    master_params_grads = []
+    tmp_role = main_prog._current_role
+    OpRole = fluid.core.op_proto_and_checker_maker.OpRole
+    main_prog._current_role = OpRole.Backward
+    for p, g in params_grads:
+        # create master parameters
+        master_param = copy_to_master_param(p, main_prog.global_block())
+        startup_master_param = startup_prog.global_block()._clone_variable(
+            master_param)
+        startup_p = startup_prog.global_block().var(p.name)
+        cast_fp16_to_fp32(startup_p, startup_master_param, startup_prog)
+        # cast fp16 gradients to fp32 before apply gradients
+        if g.name.find("layer_norm") > -1:
+            if loss_scaling > 1:
+                scaled_g = g / float(loss_scaling)
+            else:
+                scaled_g = g
+            master_params_grads.append([p, scaled_g])
+            continue
+        master_grad = fluid.layers.cast(g, "float32")
+        if loss_scaling > 1:
+            master_grad = master_grad / float(loss_scaling)
+        master_params_grads.append([master_param, master_grad])
+    main_prog._current_role = tmp_role
+    return master_params_grads
+
+
+def master_param_to_train_param(master_params_grads, params_grads, main_prog):
+    for idx, m_p_g in enumerate(master_params_grads):
+        train_p, _ = params_grads[idx]
+        if train_p.name.find("layer_norm") > -1:
+            continue
+        with main_prog._optimized_guard([m_p_g[0], m_p_g[1]]):
+            cast_fp32_to_fp16(m_p_g[0], train_p, main_prog)

ERNIE/utils/init.py

+#   Copyright (c) 2019 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 os
+import six
+import ast
+import copy
+
+import numpy as np
+import paddle.fluid as fluid
+
+
+def cast_fp32_to_fp16(exe, main_program):
+    print("Cast parameters to float16 data format.")
+    for param in main_program.global_block().all_parameters():
+        if not param.name.endswith(".master"):
+            param_t = fluid.global_scope().find_var(param.name).get_tensor()
+            data = np.array(param_t)
+            if param.name.find("layer_norm") == -1:
+                param_t.set(np.float16(data).view(np.uint16), exe.place)
+            master_param_var = fluid.global_scope().find_var(param.name +
+                                                             ".master")
+            if master_param_var is not None:
+                master_param_var.get_tensor().set(data, exe.place)
+
+
+def init_checkpoint(exe, init_checkpoint_path, main_program, use_fp16=False):
+    assert os.path.exists(
+        init_checkpoint_path), "[%s] cann't be found." % init_checkpoint_path
+
+    def existed_persitables(var):
+        if not fluid.io.is_persistable(var):
+            return False
+        return os.path.exists(os.path.join(init_checkpoint_path, var.name))
+
+    fluid.io.load_vars(
+        exe,
+        init_checkpoint_path,
+        main_program=main_program,
+        predicate=existed_persitables)
+    print("Load model from {}".format(init_checkpoint_path))
+
+    if use_fp16:
+        cast_fp32_to_fp16(exe, main_program)
+
+
+def init_pretraining_params(exe,
+                            pretraining_params_path,
+                            main_program,
+                            use_fp16=False):
+    assert os.path.exists(pretraining_params_path
+                          ), "[%s] cann't be found." % pretraining_params_path
+
+    def existed_params(var):
+        if not isinstance(var, fluid.framework.Parameter):
+            return False
+        return os.path.exists(os.path.join(pretraining_params_path, var.name))
+
+    fluid.io.load_vars(
+        exe,
+        pretraining_params_path,
+        main_program=main_program,
+        predicate=existed_params)
+    print("Load pretraining parameters from {}.".format(
+        pretraining_params_path))
+
+    if use_fp16:
+        cast_fp32_to_fp16(exe, main_program)