From 6c1b67bc5e6f6a0e9eb7e6b0562f0ae5ee1fc795 Mon Sep 17 00:00:00 2001 From: Zeyu Chen Date: Sun, 14 Apr 2019 23:30:25 +0800 Subject: [PATCH] add sequence label README.md --- demo/sequence-labeling/README.md | 1 + ...ence_labeling.sh => run_sequence_label.sh} | 4 ++-- demo/sequence-labeling/sequence_label.py | 1 + demo/text-classification/README.md | 2 +- demo/text-classification/run_predict.sh | 2 +- demo/text-classification/simple_demo.py | 18 +++++++++------- paddlehub/finetune/evaluate.py | 11 +++++----- paddlehub/finetune/finetune.py | 21 +++++++++---------- 8 files changed, 32 insertions(+), 28 deletions(-) create mode 100644 demo/sequence-labeling/README.md rename demo/sequence-labeling/{run_sequence_labeling.sh => run_sequence_label.sh} (89%) diff --git a/demo/sequence-labeling/README.md b/demo/sequence-labeling/README.md new file mode 100644 index 00000000..d13437c1 --- /dev/null +++ b/demo/sequence-labeling/README.md @@ -0,0 +1 @@ +# PaddleHub 序列标注 diff --git a/demo/sequence-labeling/run_sequence_labeling.sh b/demo/sequence-labeling/run_sequence_label.sh similarity index 89% rename from demo/sequence-labeling/run_sequence_labeling.sh rename to demo/sequence-labeling/run_sequence_label.sh index bb34ce72..55ca0bd3 100644 --- a/demo/sequence-labeling/run_sequence_labeling.sh +++ b/demo/sequence-labeling/run_sequence_label.sh @@ -1,10 +1,10 @@ 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 diff --git a/demo/sequence-labeling/sequence_label.py b/demo/sequence-labeling/sequence_label.py index 03bfb967..9a028fe0 100644 --- a/demo/sequence-labeling/sequence_label.py +++ b/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 diff --git a/demo/text-classification/README.md b/demo/text-classification/README.md index 53f649e3..e4233f95 100644 --- a/demo/text-classification/README.md +++ b/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 diff --git a/demo/text-classification/run_predict.sh b/demo/text-classification/run_predict.sh index 57b522ab..bc664cb8 100644 --- a/demo/text-classification/run_predict.sh +++ b/demo/text-classification/run_predict.sh @@ -1,4 +1,4 @@ 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 diff --git a/demo/text-classification/simple_demo.py b/demo/text-classification/simple_demo.py index ecb0ce98..c7d77ec9 100644 --- a/demo/text-classification/simple_demo.py +++ b/demo/text-classification/simple_demo.py @@ -1,29 +1,31 @@ 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, diff --git a/paddlehub/finetune/evaluate.py b/paddlehub/finetune/evaluate.py index f106c7f9..9af5bf0f 100644 --- a/paddlehub/finetune/evaluate.py +++ b/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, diff --git a/paddlehub/finetune/finetune.py b/paddlehub/finetune/finetune.py index 5a5bda64..37d5b138 100644 --- a/paddlehub/finetune/finetune.py +++ b/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) -- GitLab