add sequence label README.md

demo/sequence-labeling/README.md

+# PaddleHub 序列标注

demo/sequence-labeling/run_sequence_labeling.sh → demo/sequence-labeling/run_sequence_label.sh

 export CUDA_VISIBLE_DEVICES=0
 
-CKPT_DIR="./ckpt_sequence_labeling"
+CKPT_DIR="./ckpt_sequence_label"
 python -u sequence_label.py \
                    --batch_size 16 \
                    --weight_decay  0.01 \
-                   --checkpoint_dir $CKPT_DIR \
                    --num_epoch 3 \
+                   --checkpoint_dir $CKPT_DIR \
                    --max_seq_len 256 \
                    --learning_rate 5e-5

demo/sequence-labeling/sequence_label.py

@@ -14,6 +14,7 @@
 """Finetuning on sequence labeling task."""
 
 import argparse
+import ast
 
 import paddle.fluid as fluid
 import paddlehub as hub

demo/text-classification/README.md

@@ -133,7 +133,7 @@ hub.finetune_and_eval(task=cls_task, data_reader=reader, feed_list=feed_list, co
 * `eval_interval`: 模型评估的间隔，默认每100个step评估一次验证集
 * `save_ckpt_interval`: 模型保存间隔，请根据任务大小配置，默认只保存验证集效果最好的模型和训练结束的模型
 * `use_cuda`: 是否使用GPU训练，默认为False
-* `checkpoint_dir`: 模型checkpoint保存路径
+* `checkpoint_dir`: 模型checkpoint保存路径, 若用户没有指定，程序会自动生成
 * `num_epoch`: finetune的轮数
 * `batch_size`: 训练的批大小，如果使用GPU，请根据实际情况调整batch_size
 * `enable_memory_optim`: 是否使用内存优化， 默认为True

demo/text-classification/run_predict.sh

 export CUDA_VISIBLE_DEVICES=5
 
-CKPT_DIR="./ckpt_chnsenticorp/best_model"
+CKPT_DIR="./ckpt_20190414203357/best_model"
 python -u predict.py --checkpoint_dir $CKPT_DIR --max_seq_len 128

demo/text-classification/simple_demo.py

 import paddle.fluid as fluid
 import paddlehub as hub
 
-# Step1
+# Step1: Select pre-trained model
 module = hub.Module(name="ernie")
 inputs, outputs, program = module.context(trainable=True, max_seq_len=128)
 
-# Step2
+# Step2: Prepare Dataset and DataReader
 dataset = hub.dataset.ChnSentiCorp()
 reader = hub.reader.ClassifyReader(
     dataset=dataset, vocab_path=module.get_vocab_path(), max_seq_len=128)
 
-# Step3
+# Step3: Construct transfer learning task
 with fluid.program_guard(program):
     label = fluid.layers.data(name="label", shape=[1], dtype='int64')
-
     pooled_output = outputs["pooled_output"]
-
-    cls_task = hub.create_text_classification_task(
+    cls_task = hub.create_text_cls_task(
         feature=pooled_output, label=label, num_classes=dataset.num_labels)
 
-# Step4
+# Step4: Setup config then start finetune
 strategy = hub.AdamWeightDecayStrategy(learning_rate=5e-5, weight_decay=0.01)
 
 config = hub.RunConfig(
-    use_cuda=True, num_epoch=3, batch_size=32, strategy=strategy)
+    use_cuda=True,
+    checkpoint_dir="./ckpt",
+    num_epoch=3,
+    batch_size=32,
+    strategy=strategy)
 
 feed_list = [
     inputs["input_ids"].name, inputs["position_ids"].name,

paddlehub/finetune/evaluate.py

@@ -57,6 +57,7 @@ def evaluate_cls_task(task, data_reader, feed_list, phase="test", config=None):
         avg_loss = loss_sum / num_eval_examples
         avg_acc = acc_sum / num_eval_examples
         eval_speed = eval_step / eval_time_used
+
     logger.info(
         "[%s dataset evaluation result] loss=%.5f acc=%.5f [step/sec: %.2f]" %
         (phase, avg_loss, avg_acc, eval_speed))
@@ -64,11 +65,11 @@ def evaluate_cls_task(task, data_reader, feed_list, phase="test", config=None):
     return avg_loss, avg_acc, eval_speed
 
 
-def evaluate_seq_labeling_task(task,
-                               data_reader,
-                               feed_list,
-                               phase="test",
-                               config=None):
+def evaluate_seq_label_task(task,
+                            data_reader,
+                            feed_list,
+                            phase="test",
+                            config=None):
     fetch_list = [
         task.variable("labels").name,
         task.variable("infers").name,

paddlehub/finetune/finetune.py

@@ -27,7 +27,7 @@ from visualdl import LogWriter
 from paddlehub.common.logger import logger
 from paddlehub.finetune.strategy import AdamWeightDecayStrategy, DefaultStrategy
 from paddlehub.finetune.checkpoint import load_checkpoint, save_checkpoint
-from paddlehub.finetune.evaluate import evaluate_cls_task, evaluate_seq_labeling_task
+from paddlehub.finetune.evaluate import evaluate_cls_task, evaluate_seq_label_task
 import paddlehub as hub
 
 
@@ -126,13 +126,13 @@ def _finetune_seq_label_task(task,
                         exe=exe)
 
                 if do_eval and global_step % config.eval_interval == 0:
-                    evaluate_seq_labeling_task(
+                    evaluate_seq_label_task(
                         task,
                         data_reader,
                         feed_list,
                         phase="test",
                         config=config)
-                    evaluate_seq_labeling_task(
+                    evaluate_seq_label_task(
                         task,
                         data_reader,
                         feed_list,
@@ -148,9 +148,9 @@ def _finetune_seq_label_task(task,
             exe=exe)
 
         if do_eval:
-            evaluate_seq_labeling_task(
+            evaluate_seq_label_task(
                 task, data_reader, feed_list, phase="dev", config=config)
-            evaluate_seq_labeling_task(
+            evaluate_seq_label_task(
                 task, data_reader, feed_list, phase="test", config=config)
         logger.info("PaddleHub finetune finished.")
 
@@ -164,8 +164,8 @@ def _finetune_cls_task(task, data_reader, feed_list, config=None,
 
     num_epoch = config.num_epoch
     batch_size = config.batch_size
-    log_writter = LogWriter(
-        os.path.join(config.checkpoint_dir, "vdllog"), sync_cycle=10)
+    log_writer = LogWriter(
+        os.path.join(config.checkpoint_dir, "vdllog"), sync_cycle=1)
 
     place, dev_count = hub.common.get_running_device_info(config)
     with fluid.program_guard(main_program, startup_program):
@@ -190,10 +190,10 @@ def _finetune_cls_task(task, data_reader, feed_list, config=None,
         logger.info("PaddleHub finetune start")
 
         # add visualdl scalar
-        with log_writter.mode("train") as logw:
+        with log_writer.mode("train") as logw:
             train_loss_scalar = logw.scalar(tag="loss[train]")
             train_acc_scalar = logw.scalar(tag="accuracy[train]")
-        with log_writter.mode("evaluate") as logw:
+        with log_writer.mode("evaluate") as logw:
             eval_loss_scalar = logw.scalar(tag="loss[evaluate]")
             eval_acc_scalar = logw.scalar(tag="accuracy[evaluate]")
 
@@ -276,8 +276,7 @@ def finetune_and_eval(task, data_reader, feed_list, config=None):
     if task.task_type == "sequence_labeling":
         _finetune_seq_label_task(
             task, data_reader, feed_list, config, do_eval=True)
-    # if it's image_classification and text classificaiton
-    else:
+    elif task.task_type == "image_classification" or task.task_type == "text_classification":
         _finetune_cls_task(task, data_reader, feed_list, config, do_eval=True)