Merge pull request #12 from tianxin1860/develop

code format

ERNIE/batching.py

@@ -19,7 +19,15 @@ 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):
+
+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;
@@ -90,7 +98,8 @@ def mask(batch_tokens, seg_labels, mask_word_tags, total_token_num, vocab_size,
                     # random replace
                     if token != SEP and token != CLS:
                         mask_label.append(sent[token_index])
-                        sent[token_index] = replace_ids[prob_index + token_index]
+                        sent[token_index] = replace_ids[prob_index +
+                                                        token_index]
                         mask_flag = True
                         mask_pos.append(sent_index * max_len + token_index)
                 else:
@@ -143,7 +152,10 @@ def prepare_batch_data(insts,
     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_list = [
+        src_id, pos_id, sent_id, self_attn_bias, mask_label, mask_pos, labels,
+        next_sent_index
+    ]
 
     return return_list
 
@@ -207,4 +219,5 @@ def pad_batch_data(insts,
 
 
 if __name__ == "__main__":
+
     pass

ERNIE/finetune/classifier.py

@@ -25,22 +25,20 @@ import paddle.fluid as fluid
 from model.ernie import ErnieModel
 
 
-def create_model(args,
-                 pyreader_name,
-                 ernie_config,
-                 is_prediction=False):
+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]],
+                [-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)
+    (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,
@@ -57,7 +55,7 @@ def create_model(args,
         dropout_implementation="upscale_in_train")
     logits = fluid.layers.fc(
         input=cls_feats,
-        size=ernie_config["num_labels"],
+        size=args.num_labels,
         param_attr=fluid.ParamAttr(
             name="cls_out_w",
             initializer=fluid.initializer.TruncatedNormal(scale=0.02)),
@@ -82,18 +80,21 @@ def create_model(args,
     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}
+    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
+        v.persistable = True
 
     return pyreader, graph_vars
 
+
 def evaluate_mrr(preds):
     last_qid = None
     total_mrr = 0.0
@@ -114,6 +115,7 @@ def evaluate_mrr(preds):
 
     return total_mrr / qnum
 
+
 def evaluate_map(preds):
     def singe_map(st, en):
         total_p = 0.0
@@ -142,17 +144,18 @@ def evaluate_map(preds):
     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
-                       ] 
+    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])}
+        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
@@ -162,22 +165,21 @@ def evaluate(exe, test_program, test_pyreader, graph_vars, eval_phase):
     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]
+    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)
+            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())
+            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)
@@ -188,20 +190,23 @@ def evaluate(exe, test_program, test_pyreader, graph_vars, eval_phase):
     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))
+        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]))
+        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))
+        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_args.py

@@ -64,7 +64,7 @@ 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")
+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.")
@@ -74,3 +74,4 @@ run_type_g.add_arg("do_train",                     bool,   True,  "Whether to pe
 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.")
+# yapf: enable

ERNIE/pretrain_args.py

@@ -24,7 +24,6 @@ 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.")

ERNIE/reader/pretraining.py

@@ -30,6 +30,7 @@ import paddle.fluid as fluid
 
 from batching import prepare_batch_data
 
+
 class ErnieDataReader(object):
     def __init__(self,
                  filelist,
@@ -81,8 +82,8 @@ class ErnieDataReader(object):
         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
+        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:
@@ -153,14 +154,17 @@ class ErnieDataReader(object):
             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]]
+                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)
+            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]

ERNIE/reader/task_reader.py

@@ -62,7 +62,7 @@ class BaseReader(object):
             reader = csv.reader(f, delimiter="\t", quotechar=quotechar)
             headers = next(reader)
             Example = namedtuple('Example', headers)
-                
+
             examples = []
             for line in reader:
                 example = Example(*line)
@@ -85,7 +85,7 @@ class BaseReader(object):
             else:
                 tokens_b.pop()
 
-    def _convert_example_to_record(self, example,  max_seq_length, tokenizer):
+    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)
@@ -148,7 +148,9 @@ class BaseReader(object):
         else:
             label_id = example.label
 
-        Record = namedtuple('Record', ['token_ids', 'text_type_ids', 'position_ids', 'label_id', 'qid'])
+        Record = namedtuple(
+            'Record',
+            ['token_ids', 'text_type_ids', 'position_ids', 'label_id', 'qid'])
 
         qid = None
         if "qid" in example._fields:
@@ -164,11 +166,12 @@ class BaseReader(object):
 
     def _prepare_batch_data(self, examples, batch_size, phase=None):
         """generate batch records"""
-        batch_records,  max_len = [], 0
+        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)
+            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
@@ -187,7 +190,12 @@ class BaseReader(object):
         examples = self._read_tsv(input_file)
         return len(examples)
 
-    def data_generator(self, input_file, batch_size, epoch, shuffle=True, phase=None):
+    def data_generator(self,
+                       input_file,
+                       batch_size,
+                       epoch,
+                       shuffle=True,
+                       phase=None):
         examples = self._read_tsv(input_file)
 
         def wrapper():
@@ -198,8 +206,10 @@ class BaseReader(object):
                 if shuffle:
                     np.random.shuffle(examples)
 
-                for batch_data in self._prepare_batch_data(examples, batch_size, phase=phase):
+                for batch_data in self._prepare_batch_data(
+                        examples, batch_size, phase=phase):
                     yield batch_data
+
         return wrapper
 
 
@@ -209,9 +219,11 @@ class ClassifyReader(BaseReader):
         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"]
+            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):
@@ -219,8 +231,8 @@ class ClassifyReader(BaseReader):
                         line[index] = text.replace(' ', '')
                 example = Example(*line)
                 examples.append(example)
-            return examples 
-    
+            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]
@@ -236,33 +248,50 @@ class ClassifyReader(BaseReader):
 
         # 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]
+            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]
+        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]
+            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):
@@ -285,7 +314,7 @@ class SequenceLabelReader(BaseReader):
         assert len(ret_tokens) == len(ret_labels)
         return ret_tokens, ret_labels
 
-    def _convert_example_to_record(self, example,  max_seq_length, tokenizer):
+    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)
@@ -297,11 +326,15 @@ class SequenceLabelReader(BaseReader):
         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) 
+        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]
+        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 = namedtuple(
+            'Record',
+            ['token_ids', 'text_type_ids', 'position_ids', 'label_ids'])
         record = Record(
             token_ids=token_ids,
             text_type_ids=text_type_ids,
@@ -309,5 +342,6 @@ class SequenceLabelReader(BaseReader):
             label_ids=label_ids)
         return record
 
+
 if __name__ == '__main__':
     pass

ERNIE/run_classifier.py

@@ -32,11 +32,11 @@ 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 
-
+from finetune_args import parser
 
 args = parser.parse_args()
 
+
 def main(args):
     ernie_config = ErnieConfig(args.ernie_config_path)
     ernie_config.print_config()
@@ -49,12 +49,13 @@ def main(args):
         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)
+    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 "
@@ -108,10 +109,11 @@ def main(args):
 
                 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,
+                    skip_opt_set=[
+                        graph_vars["loss"].name,
+                        graph_vars["probs"].name,
+                        graph_vars["accuracy"].name,
+                        graph_vars["num_seqs"].name,
                     ])
 
         if args.verbose:
@@ -201,7 +203,8 @@ def main(args):
                 if steps % args.skip_steps != 0:
                     train_exe.run(fetch_list=[])
                 else:
-                    outputs = evaluate(train_exe, train_program, train_pyreader, graph_vars, "train") 
+                    outputs = evaluate(train_exe, train_program, train_pyreader,
+                                       graph_vars, "train")
 
                     if args.verbose:
                         verbose = "train pyreader queue size: %d, " % train_pyreader.queue.size(
@@ -217,7 +220,8 @@ def main(args):
                     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))
+                           steps, outputs["loss"], outputs["accuracy"],
+                           args.skip_steps / used_time))
                     time_begin = time.time()
 
                 if steps % args.save_steps == 0:
@@ -254,7 +258,9 @@ def main(args):
     if args.do_val:
         test_pyreader.decorate_tensor_provider(
             reader.data_generator(
-                args.dev_set, batch_size=args.batch_size, epoch=1,
+                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")
@@ -273,4 +279,5 @@ def main(args):
 
 if __name__ == '__main__':
     print_arguments(args)
+
     main(args)

ERNIE/run_sequence_labeling.py

@@ -30,12 +30,12 @@ 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 
-
+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()
@@ -48,12 +48,13 @@ def main(args):
         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)
+    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 "
@@ -107,10 +108,9 @@ def main(args):
 
                 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
+                    skip_opt_set=[
+                        graph_vars["loss"].name, graph_vars["labels"].name,
+                        graph_vars["infers"].name, graph_vars["seq_lens"].name
                     ])
 
         if args.verbose:
@@ -200,24 +200,26 @@ def main(args):
                 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)
+                    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)
+                            outputs["lr"]
+                            if warmup_steps > 0 else args.learning_rate)
                         print(verbose)
 
-                    current_example, current_epoch = reader.get_train_progress(
-                    )
+                    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))
+                          "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:
@@ -234,7 +236,8 @@ def main(args):
                                 batch_size=args.batch_size,
                                 epoch=1,
                                 shuffle=False))
-                        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "dev")
+                        evaluate(exe, test_prog, test_pyreader, graph_vars,
+                                 args.num_labels, "dev")
                     # evaluate test set
                     if args.do_test:
                         test_pyreader.decorate_tensor_provider(
@@ -243,7 +246,8 @@ def main(args):
                                 batch_size=args.batch_size,
                                 epoch=1,
                                 shuffle=False))
-                        evaluate(exe, test_prog, test_pyreader, graph_vars, args.num_labels, "test")
+                        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))
@@ -255,7 +259,9 @@ def main(args):
     if args.do_val:
         test_pyreader.decorate_tensor_provider(
             reader.data_generator(
-                args.dev_set, batch_size=args.batch_size, epoch=1,
+                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")

ERNIE/train.py

@@ -35,8 +35,10 @@ 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,
@@ -224,8 +226,7 @@ def train(args):
             print("train_id == 0, sleep 60s")
             time.sleep(60)
         print("worker_endpoints:{} trainers_num:{} current_endpoint:{} \
-              trainer_id:{}"
-                            .format(worker_endpoints, trainers_num,
+              trainer_id:{}".format(worker_endpoints, trainers_num,
                                     current_endpoint, trainer_id))
 
         # prepare nccl2 env.
@@ -319,13 +320,14 @@ def train(args):
                 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))
+                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 = []
@@ -341,8 +343,7 @@ def train(args):
                 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),
+                      (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))