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

BERT/train.py

@@ -313,6 +313,9 @@ def train(args):
     exec_strategy.num_threads = dev_count
     exec_strategy.num_iteration_per_drop_scope = args.num_iteration_per_drop_scope
 
+    # use_ngraph is for CPU only, please refer to README_ngraph.md for details
+    use_ngraph = os.getenv('FLAGS_use_ngraph')
+    if not use_ngraph:
         train_exe = fluid.ParallelExecutor(
             use_cuda=args.use_cuda,
             loss_name=total_loss.name,
@@ -320,6 +323,8 @@ def train(args):
             main_program=train_program,
             num_trainers=nccl2_num_trainers,
             trainer_id=nccl2_trainer_id)
+    else:
+        train_exe = exe
 
     if args.validation_set_dir and args.validation_set_dir != "":
         predict = predict_wrapper(
@@ -345,17 +350,30 @@ def train(args):
             skip_steps = args.skip_steps * nccl2_num_trainers
 
             if nccl2_trainer_id != 0:
+                if use_ngraph:
+                    train_exe.run(fetch_list=[], program=train_program)
+                else:
                     train_exe.run(fetch_list=[])
                 continue
 
             if steps % skip_steps != 0:
+                if use_ngraph:
+                    train_exe.run(fetch_list=[], program=train_program)
+                else:
                     train_exe.run(fetch_list=[])
+
             else:
+                if use_ngraph:
                     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
-                    ])
+                            scheduled_lr.name], program=train_program)
+                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)
@@ -398,7 +416,6 @@ def train(args):
             train_pyreader.reset()
             break
 
-
 if __name__ == '__main__':
     print_arguments(args)
     if args.do_test:

ELMo/README.md

@@ -20,7 +20,7 @@ ELMo(Embeddings from Language Models) 是重要的通用语义表示模型之一
 | Task | 评估指标 | Baseline | +ELMo |
 | :------| :------: | :------: |:------: |
 | [LAC](https://github.com/baidu/lac) | F1 | 87.3% | **88.4%** |
-| [阅读理解](github.com/PaddlePaddle/models/tree/develop/PaddleNLP/machine_reading_comprehension) | Rouge-L | 39.4% | **40.4%** |
+| [阅读理解](https://github.com/PaddlePaddle/models/tree/develop/PaddleNLP/unarchived/machine_reading_comprehension) | Rouge-L | 39.4% | **40.4%** |
 
 **Note**:
 

ELMo/train.py

@@ -555,6 +555,7 @@ def train_loop(args,
             valid_ppl = eval(vocab, infer_progs, dev_count, logger, args)
             logger.info("valid ppl {}".format(valid_ppl))
         if batch_id > 0 and batch_id % args.save_interval == 0:
+            epoch_id = int(batch_id / n_batches_per_epoch)
             model_path = os.path.join(args.para_save_dir,
                                       str(batch_id + epoch_id))
             if not os.path.isdir(model_path):

ERNIE/finetune/classifier.py

@@ -156,7 +156,7 @@ def evaluate(exe, test_program, test_pyreader, graph_vars, eval_phase):
         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])
+            ret["learning_rate"] = float(outputs[3][0])
         return ret
 
     test_pyreader.start()