fix ci

9f64ef8e · shenyuhan · 7ea3c245 · 7151f6bb · 9f64ef8e · 9f64ef8e
11 changed file
--- a/README.md
+++ b/README.md
@@ -37,8 +37,9 @@ PaddleHub是基于PaddlePaddle生态下的预训练模型管理和迁移学习
 * Python==2.7 or Python>=3.5
 * PaddlePaddle>=1.4.0
-除上述依赖外，PaddleHub的预训练模型和预置数据集需要连接服务端进行下载，请确保机器可以正常访问网络
+除上述依赖外，PaddleHub的预训练模型和预置数据集需要连接服务端进行下载，请确保机器可以正常访问网络。若本地已存在相关的数据集和预训练模型，则可以离线运行PaddleHub。
+**NOTE:** 若是出现离线运行PaddleHub错误，请更新PaddleHub 1.1.1版本之上。
 pip安装方式如下：
 ```shell
@@ -102,9 +103,11 @@ PaddleHub如何完成迁移学习，详情参考[wiki教程](https://github.com/
 PaddleHub如何自定义迁移任务，详情参考[wiki教程](https://github.com/PaddlePaddle/PaddleHub/wiki/PaddleHub:-%E8%87%AA%E5%AE%9A%E4%B9%89Task)
-如何使用PaddleHub超参优化功能[autofinetune使用教程](https://github.com/PaddlePaddle/PaddleHub/blob/develop/tutorial/autofinetune.md)
+如何使用PaddleHub超参优化功能，详情参考[autofinetune使用教程](https://github.com/PaddlePaddle/PaddleHub/blob/develop/tutorial/autofinetune.md)
-如何使用PaddleHub“端到端地”完成文本相似度计算[word2vce使用教程](https://github.com/PaddlePaddle/PaddleHub/blob/develop/tutorial/sentence_sim.ipynb)
+如何使用PaddleHub“端到端地”完成文本相似度计算，详情参考[word2vce使用教程](https://github.com/PaddlePaddle/PaddleHub/blob/develop/tutorial/sentence_sim.ipynb)
+如何使用ULMFiT策略微调PaddleHub预训练模型，详情参考[PaddleHub 迁移学习与ULMFiT微调策略](https://github.com/PaddlePaddle/PaddleHub/blob/develop/tutorial/strategy_exp.md)
 ## FAQ

--- a/RELEASE.md
+++ b/RELEASE.md
@@ -14,6 +14,7 @@
  * 新增**阅读理解Fine-tune任务**和**回归Fine-tune任务**
  * 新增多指标评测
  * 优化predict接口
+  * 可视化工具支持使用tb_paddle
 # PaddleHub v1.1.2
@@ -69,7 +70,5 @@
 正式发布PaddleHub预训练模型管理工具，旨在帮助用户更高效的管理模型并开展迁移学习的工作。
 **预训练模型管理**: 通过hub命令行可完成PaddlePaddle生态的预训练模型下载、搜索、版本管理等功能。
 **命令行一键使用**: 无需代码，通过命令行即可直接使用预训练模型进行预测，快速调研训练模型效果。目前版本支持以下模型：词法分析LAC；情感分析Senta；目标检测SSD；图像分类ResNet, MobileNet, NASNet等。
 **迁移学习**: 提供了基于预训练模型的Finetune API，用户通过少量代码即可完成迁移学习，包括BERT/ERNIE文本分类、序列标注、图像分类迁移等。
--- a/paddlehub/autofinetune/autoft.py
+++ b/paddlehub/autofinetune/autoft.py
@@ -164,15 +164,15 @@ class BaseTuningStrategy(object):
        params_cudas_dirs = []
        solution_results = []
        cnt = 0
-        solutions_ckptdirs = {}
+        solutions_modeldirs = {}
        mkdir(output_dir)
        for idx, solution in enumerate(solutions):
            cuda = self.is_cuda_free["free"][0]
-            ckptdir = output_dir + "/ckpt-" + str(idx)
+            modeldir = output_dir + "/model-" + str(idx) + "/"
            log_file = output_dir + "/log-" + str(idx) + ".info"
-            params_cudas_dirs.append([solution, cuda, ckptdir, log_file])
+            params_cudas_dirs.append([solution, cuda, modeldir, log_file])
-            solutions_ckptdirs[tuple(solution)] = ckptdir
+            solutions_modeldirs[tuple(solution)] = modeldir
            self.is_cuda_free["free"].remove(cuda)
            self.is_cuda_free["busy"].append(cuda)
            if len(params_cudas_dirs
@@ -190,7 +190,7 @@ class BaseTuningStrategy(object):
        self.feedback(solutions, solution_results)
-        return solutions_ckptdirs
+        return solutions_modeldirs
 class HAZero(BaseTuningStrategy):

--- a/paddlehub/autofinetune/evaluator.py
+++ b/paddlehub/autofinetune/evaluator.py
@@ -36,11 +36,12 @@ else:
 class BaseEvaluator(object):
-    def __init__(self, params_file, finetunee_script):
+    def __init__(self, params_file, finetunee_script, options_str=""):
        with io.open(params_file, 'r', encoding='utf8') as f:
            self.params = yaml.safe_load(f)
        self.finetunee_script = finetunee_script
        self.model_rewards = {}
+        self.options_str = options_str
    def get_init_params(self):
        init_params = []
@@ -108,13 +109,14 @@ class BaseEvaluator(object):
 class FullTrailEvaluator(BaseEvaluator):
-    def __init__(self, params_file, finetunee_script):
+    def __init__(self, params_file, finetunee_script, options_str=""):
-        super(FullTrailEvaluator, self).__init__(params_file, finetunee_script)
+        super(FullTrailEvaluator, self).__init__(
+            params_file, finetunee_script, options_str=options_str)
    def run(self, *args):
        params = args[0][0]
        num_cuda = args[0][1]
-        ckpt_dir = args[0][2]
+        saved_params_dir = args[0][2]
        log_file = args[0][3]
        params = self.convert_params(params)
        if not self.is_valid_params(params):
@@ -125,12 +127,11 @@ class FullTrailEvaluator(BaseEvaluator):
        f.close()
        if is_windows():
-            run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --checkpoint_dir=%s %s >%s 2>&1" % \
+            run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                    (num_cuda, self.finetunee_script, ckpt_dir, param_str, log_file)
+                    (num_cuda, self.finetunee_script, saved_params_dir, param_str, self.options_str, log_file)
        else:
-            run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --checkpoint_dir=%s %s >%s 2>&1" % \
+            run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                    (num_cuda, self.finetunee_script, ckpt_dir, param_str, log_file)
+                    (num_cuda, self.finetunee_script, saved_params_dir, param_str, self.options_str, log_file)
        try:
            os.system(run_cmd)
            with open(log_file, "r") as f:
@@ -142,20 +143,21 @@ class FullTrailEvaluator(BaseEvaluator):
                % param_str.replace("--", ""))
            eval_result = 0.0
        reward = self.get_reward(eval_result)
-        self.model_rewards[ckpt_dir] = reward
+        self.model_rewards[saved_params_dir] = reward
        return reward
 class ModelBasedEvaluator(BaseEvaluator):
-    def __init__(self, params_file, finetunee_script):
+    def __init__(self, params_file, finetunee_script, options_str=""):
-        super(ModelBasedEvaluator, self).__init__(params_file, finetunee_script)
+        super(ModelBasedEvaluator, self).__init__(
-        self.half_best_model_ckpt = []
+            params_file, finetunee_script, options_str=options_str)
+        self.half_best_model_path = []
        self.run_count = 0
    def run(self, *args):
        params = args[0][0]
        num_cuda = args[0][1]
-        ckpt_dir = args[0][2]
+        saved_params_dir = args[0][2]
        log_file = args[0][3]
        params = self.convert_params(params)
        if not self.is_valid_params(params):
@@ -165,22 +167,23 @@ class ModelBasedEvaluator(BaseEvaluator):
        f = open(log_file, "w")
        f.close()
-        if len(self.half_best_model_ckpt) > 0:
+        if len(self.half_best_model_path) > 0:
-            model_path = self.half_best_model_ckpt[self.run_count % len(
+            model_path = self.half_best_model_path[self.run_count % len(
-                self.half_best_model_ckpt)] + "/best_model"
+                self.half_best_model_path)]
            if is_windows():
-                run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --epochs=1 --model_path %s --checkpoint_dir=%s %s >%s 2>&1" % \
+                run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --epochs=1 --model_path %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                        (num_cuda, self.finetunee_script, model_path, ckpt_dir, param_str, log_file)
+                        (num_cuda, self.finetunee_script, model_path, saved_params_dir, param_str, self.options_str, log_file)
            else:
-                run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --epochs=1 --model_path %s --checkpoint_dir=%s %s >%s 2>&1" % \
+                run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --epochs=1 --model_path %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                        (num_cuda, self.finetunee_script, model_path, ckpt_dir, param_str, log_file)
+                        (num_cuda, self.finetunee_script, model_path, saved_params_dir, param_str, self.options_str, log_file)
        else:
            if is_windows():
-                run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --checkpoint_dir=%s %s >%s 2>&1" % \
+                run_cmd = "set FLAGS_eager_delete_tensor_gb=0.0&set CUDA_VISIBLE_DEVICES=%s&python -u %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                        (num_cuda, self.finetunee_script, ckpt_dir, param_str, log_file)
+                        (num_cuda, self.finetunee_script, saved_params_dir, param_str, self.options_str, log_file)
            else:
-                run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --checkpoint_dir=%s %s >%s 2>&1" % \
+                run_cmd = "export FLAGS_eager_delete_tensor_gb=0.0; export CUDA_VISIBLE_DEVICES=%s; python -u %s --saved_params_dir=%s %s %s >%s 2>&1" % \
-                        (num_cuda, self.finetunee_script, ckpt_dir, param_str, log_file)
+                        (num_cuda, self.finetunee_script, saved_params_dir, param_str, self.options_str, log_file)
        self.run_count += 1
        try:
@@ -194,7 +197,7 @@ class ModelBasedEvaluator(BaseEvaluator):
                % param_str.replace("--", ""))
            eval_result = 0.0
        reward = self.get_reward(eval_result)
-        self.model_rewards[ckpt_dir] = reward
+        self.model_rewards[saved_params_dir] = reward
        return reward
    def new_round(self):
@@ -202,7 +205,7 @@ class ModelBasedEvaluator(BaseEvaluator):
        half_size = int(len(self.model_rewards) / 2)
        if half_size < 1:
            half_size = 1
-        self.half_best_model_ckpt = list({
+        self.half_best_model_path = list({
            key
            for key in sorted(
                self.model_rewards, key=self.model_rewards.get, reverse=False)

--- a/paddlehub/commands/autofinetune.py
+++ b/paddlehub/commands/autofinetune.py
@@ -91,6 +91,21 @@ class AutoFineTuneCommand(BaseCommand):
            type=str,
            default="HAZero",
            help="Choices: HAZero or PSHE2.")
+        self.arg_config_group.add_argument(
+            'opts',
+            help='See utils/config.py for all options',
+            default=None,
+            nargs=argparse.REMAINDER)
+    def convert_to_other_options(self, config_list):
+        if len(config_list) % 2 != 0:
+            raise ValueError(
+                "Command for finetuned task options config format error! Please check it: {}"
+                .format(config_list))
+        options_str = ""
+        for key, value in zip(config_list[0::2], config_list[1::2]):
+            options_str += "--" + key + "=" + value + " "
+        return options_str
    def execute(self, argv):
        if not argv:
@@ -109,6 +124,11 @@ class AutoFineTuneCommand(BaseCommand):
            description=
            "Autofintune configuration for controlling autofinetune behavior, not required"
        )
+        self.arg_finetuned_task_group = self.parser.add_argument_group(
+            title="Finetuned task config options",
+            description=
+            "Finetuned task configuration for controlling finetuned task behavior, not required"
+        )
        self.add_params_file_arg()
        self.add_autoft_config_arg()
@@ -118,12 +138,20 @@ class AutoFineTuneCommand(BaseCommand):
            return False
        self.args = self.parser.parse_args(argv[1:])
+        options_str = ""
+        if self.args.opts is not None:
+            options_str = self.convert_to_other_options(self.args.opts)
        if self.args.evaluate_choice.lower() == "fulltrail":
-            evaluator = FullTrailEvaluator(self.args.param_file,
+            evaluator = FullTrailEvaluator(
-                                           self.fintunee_script)
+                self.args.param_file,
+                self.fintunee_script,
+                options_str=options_str)
        elif self.args.evaluate_choice.lower() == "modelbased":
-            evaluator = ModelBasedEvaluator(self.args.param_file,
+            evaluator = ModelBasedEvaluator(
-                                            self.fintunee_script)
+                self.args.param_file,
+                self.fintunee_script,
+                options_str=options_str)
        else:
            raise ValueError(
                "The evaluate %s is not defined!" % self.args.evaluate_choice)
@@ -145,13 +173,13 @@ class AutoFineTuneCommand(BaseCommand):
                             self.args.tuning_strategy)
        run_round_cnt = 0
-        solutions_ckptdirs = {}
+        solutions_modeldirs = {}
        print("PaddleHub Autofinetune starts.")
        while (not autoft.is_stop()) and run_round_cnt < self.args.round:
            print("PaddleHub Autofinetune starts round at %s." % run_round_cnt)
            output_dir = autoft._output_dir + "/round" + str(run_round_cnt)
            res = autoft.step(output_dir)
-            solutions_ckptdirs.update(res)
+            solutions_modeldirs.update(res)
            evaluator.new_round()
            run_round_cnt = run_round_cnt + 1
        print("PaddleHub Autofinetune ends.")
@@ -164,17 +192,15 @@ class AutoFineTuneCommand(BaseCommand):
                print("%s=%s" % (hparam_name, best_hparams[index]))
                f.write(hparam_name + "\t:\t" + str(best_hparams[index]) + "\n")
            f.write("\n\n\n")
-            f.write("\t".join(autoft.hparams_name_list) + "\toutput_dir\n\n")
+            f.write("\t".join(autoft.hparams_name_list) +
-            logger.info(
+                    "\tsaved_params_dir\n\n")
-                "The checkpont directory of programs ran with hyperparamemters searched are saved as log_file.txt ."
-            )
            print(
                "The checkpont directory of programs ran with hyperparamemters searched are saved as log_file.txt ."
            )
-            for solution, ckptdir in solutions_ckptdirs.items():
+            for solution, modeldir in solutions_modeldirs.items():
                param = evaluator.convert_params(solution)
                param = [str(p) for p in param]
-                f.write("\t".join(param) + "\t" + ckptdir + "\n\n")
+                f.write("\t".join(param) + "\t" + modeldir + "\n\n")
        return True

--- a/paddlehub/finetune/task/basic_task.py
+++ b/paddlehub/finetune/task/basic_task.py
@@ -588,13 +588,13 @@ class BasicTask(object):
                    run_states = self._run(do_eval=do_eval)
                    self.env.current_epoch += 1
-                # Save checkpoint after finetune
-                self.save_checkpoint()
                # Final evaluation
                if self._base_data_reader.get_dev_examples() != []:
                    self.eval(phase="dev")
                if self._base_data_reader.get_test_examples() != []:
                    self.eval(phase="test", load_best_model=True)
+                # Save checkpoint after finetune
+                self.save_checkpoint()
            self._finetune_end_event(run_states)
            return run_states

--- a/requirements.txt
+++ b/requirements.txt
+#for python2, you should use requirements_py2.txt
 pre-commit
 protobuf >= 3.1.0
 yapf == 0.26.0
 pyyaml
 numpy
-#[py2]numpy == 1.16.0
 Pillow
 six >= 1.10.0
 chardet == 3.0.4
 requests
 pandas
-#[py2]pandas == 0.24.0
 flake8
 tb-paddle
 tb-nightly

--- a/requirements_py2.txt
+++ b/requirements_py2.txt
+pre-commit
+protobuf >= 3.1.0
+yapf == 0.26.0
+pyyaml
+numpy < 1.17.0
+Pillow
+six >= 1.10.0
+chardet == 3.0.4
+requests
+pandas < 0.25.0
+flake8
+tb-paddle
+tb-nightly
+cma == 2.7.0
--- a/setup.py
+++ b/setup.py
@@ -36,7 +36,7 @@ REQUIRED_PACKAGES = [
 ]
 if max_version < 3:
-    REQUIRED_PACKAGES += ["numpy == 1.16.0", "pandas == 0.24.0"]
+    REQUIRED_PACKAGES += ["numpy < 1.17.0", "pandas < 0.25.0"]
 else:
    REQUIRED_PACKAGES += ["numpy", "pandas"]

--- a/tutorial/autofinetune.md
+++ b/tutorial/autofinetune.md
@@ -11,6 +11,10 @@ PaddleHub Auto Fine-tune提供两种超参优化策略：
 * PSHE2: 采用粒子群算法，最优超参数组合就是所求问题的解。现在想求得最优解就是要找到更新超参数组合，即如何更新超参数，才能让算法更快更好的收敛到最优解。PSE2算法根据超参数本身历史的最优，在一定随机扰动的情况下决定下一步的更新方向。
 PaddleHub Auto Fine-tune提供两种超参评估策略：
 * FullTrail: 给定一组超参，利用这组超参从头开始Finetune一个新模型，之后在数据集dev部分评估这个模型
@@ -75,11 +79,21 @@ parser.add_argument("--warmup_prop", type=float, default=0.1, help="warmup_prop.
 parser.add_argument("--weight_decay", type=float, default=0.01, help="weight_decay.")
 parser.add_argument("--max_seq_len", type=int, default=128, help="Number of words of the longest seqence.")
 parser.add_argument("--checkpoint_dir", type=str, default=None, help="Directory to model checkpoint")
+parser.add_argument("--saved_params_dir", type=str, default="", help="Directory for saving model during ")
 parser.add_argument("--model_path", type=str, default="", help="load model path")
 args = parser.parse_args()
 # yapf: enable.
+def is_path_valid(path):
+    if path == "":
+        return False
+    path = os.path.abspath(path)
+    dirname = os.path.dirname(path)
+    if not os.path.exists(dirname):
+        os.mkdir(dirname)
+    return True
 if __name__ == '__main__':
    # Load Paddlehub ERNIE pretrained model
    module = hub.Module(name="ernie")
@@ -133,7 +147,7 @@ if __name__ == '__main__':
        config=config,
        metrics_choices=metrics_choices)
-    # Finetune and evaluate by PaddleHub's API
+    # Load model from the defined model path or not
    if args.model_path != "":
        with cls_task.phase_guard(phase="train"):
            cls_task.init_if_necessary()
@@ -144,12 +158,17 @@ if __name__ == '__main__':
    run_states = cls_task.eval()
    eval_avg_score, eval_avg_loss, eval_run_speed = cls_task._calculate_metrics(run_states)
-print(eval_avg_score["acc"], end="")
+    # Move ckpt/best_model to the defined saved parameters directory
+    if is_path_valid(args.saved_params_dir) and os.path.exists(config.checkpoint_dir+"/best_model/"):
+        shutil.copytree(config.checkpoint_dir+"/best_model/", args.saved_params_dir)
+        shutil.rmtree(config.checkpoint_dir)
+    print(eval_avg_score["acc"], end="")
 ```
 **Note**:以上是finetunee.py的写法。
 > finetunee.py必须可以接收待优化超参数选项参数, 并且待搜素超参数选项名字和yaml文件中的超参数名字保持一致.
-> finetunee.py必须有checkpoint_dir这个选项。
+> finetunee.py必须有saved_params_dir这个选项。
 > PaddleHub Auto Fine-tune超参评估策略选择为ModelBased，finetunee.py必须有model_path选项。
@@ -167,7 +186,7 @@ print(eval_avg_score["acc"], end="")
 ```shell
 $ OUTPUT=result/
 $ hub autofinetune finetunee.py --param_file=hparam.yaml --cuda=['1','2'] --popsize=5 --round=10
-$ --output_dir=${OUTPUT} --evaluate_choice=fulltrail --tuning_strategy=pshe2
+ --output_dir=${OUTPUT} --evaluate_choice=fulltrail --tuning_strategy=pshe2
 ```
 其中，选项
@@ -202,8 +221,16 @@ $ tensorboard --logdir $OUTPUT/tb_paddle --host ${HOST_IP} --port ${PORT_NUM}
 ## 五、其他
-如在使用Auto Fine-tune功能时，输出信息中包含如下字样：
+1. 如在使用Auto Fine-tune功能时，输出信息中包含如下字样：
 **WARNING：Program which was ran with hyperparameters as ... was crashed!**
 首先根据终端上的输出信息，确定这个输出信息是在第几个round（如round 3），之后查看${OUTPUT}/round3/下的日志文件信息log.info, 查看具体出错原因。
+2. PaddleHub AutoFinetune 命令行支持从启动命令hub autofinetune传入finetunee.py中不需要搜索的选项参数，如上述示例中的max_seq_len选项，可以参照以下方式传入。
+```shell
+$ OUTPUT=result/
+$ hub autofinetune finetunee.py --param_file=hparam.yaml --cuda=['1','2'] --popsize=5 --round=10
+ --output_dir=${OUTPUT} --evaluate_choice=fulltrail --tuning_strategy=pshe2 max_seq_len 128
+```
--- a/tutorial/strategy_exp.md
+++ b/tutorial/strategy_exp.md