TaaS 模型训练的运行参数通过环境变量的传递。

TaaS 平台模型训练的 Base 框架调整后，用户的代码文件作为主入口文件。如果用户使用了 argparse 来
解析命令参数的话，将会出现无法处理 TaaS 运行参数，于是将 TaaS 运行参数通过环境变量的方式提供。

.gitignore

@@ -1,3 +1,4 @@
 *~
 *.pyc
 .DS_Store
+

caicloud.tensorflow/VERSION

@@ -1 +1 @@
-2.0.0
+2.0.1

caicloud.tensorflow/caicloud/clever/examples/mnist/train_model.sh

@@ -18,4 +18,8 @@
 rm -rf /tmp/caicloud-dist-tf
 rm -rf /tmp/saved_model/mnist
 
-python mnist_export.py --max_steps=2000 --data_dir="/tmp/mnist-data" --export_dir="/tmp/saved_model/mnist"
+export TF_MAX_STEPS=3000
+export TF_LOGDIR=/tmp/mnist-log
+export TF_SAVE_CHECKPOINTS_SECS=60
+export TF_SAVE_SUMMARIES_STEPS=10
+python mnist_export.py --data_dir="/tmp/mnist-data" --export_dir="/tmp/saved_model/mnist"

caicloud.tensorflow/caicloud/clever/examples/ptb/train-model.sh

@@ -18,10 +18,10 @@
 rm -rf /tmp/ptb_saved_model
 rm -rf /tmp/ptb
 
+export TF_MAX_STEPS=10000
+export TF_LOGDIR=/tmp/ptb
+export TF_SAVE_CHECKPOINTS_SECS=60
+export TF_SAVE_SUMMARIES_STEPS=10
 python ptb_caicloud_taas.py \
-       --max_steps=10000 \
-       --save_checkpoints_secs=3 \
-       --save_summaries_secs=3 \
-       --logdir=/tmp/ptb \
        --data_path=./simple-examples/data \
        --save_path=/tmp/ptb_saved_model

caicloud.tensorflow/caicloud/clever/tensorflow/dist_base.py

@@ -20,19 +20,19 @@
 from datetime import datetime
 import tensorflow as tf
 from caicloud.clever.tensorflow import model_exporter
+import os
 
-tf.app.flags.DEFINE_integer("max_steps",
-                            1,
-                            "maximum train steps.")
-tf.app.flags.DEFINE_string("logdir",
-                           "/tmp/caicloud-dist-tf",
-                           "saves checkpoints and summaries directory path.")
-tf.app.flags.DEFINE_integer("save_checkpoints_secs", 60,
-                            "save checkpoints after every special seconds")
-tf.app.flags.DEFINE_integer("save_summaries_secs", 120,
-                            "save summaries after every special seconds")
+class RunConfig(object):
+    def __init__(self):
+        self.is_chief = True
+        self.use_gpu = False
+        self.sync = False
+        self.max_steps = int(os.getenv("TF_MAX_STEPS", "1"))
+        self.logdir = os.getenv("TF_LOGDIR", "/tmp/caicloud-dist-tf")
+        self.save_checkpoints_secs = int(os.getenv("TF_SAVE_CHECKPOINTS_SECS", "600"))
+        self.save_summaries_steps = int(os.getenv("TF_SAVE_SUMMARIES_STEPS", "100"))
 
-FLAGS = tf.app.flags.FLAGS
+cfg = RunConfig()
 
 class ModelFnHandler(object):
     """model_fn 函数返回的一些模型信息。"""
@@ -196,16 +196,16 @@ def run(self, train_fn):
             if self._gen_init_fn is not None:
                 init_fn = self._gen_init_fn()
 
-        logdir = FLAGS.logdir
+        logdir = cfg.logdir
         sv = tf.train.Supervisor(
             logdir=logdir,
             graph=g,
             init_op=init_op,
             summary_op=summary_op,
             saver=saver,
             global_step=self._global_step,
-            save_model_secs=FLAGS.save_checkpoints_secs,
-            save_summaries_secs=FLAGS.save_summaries_secs,
+            save_model_secs=cfg.save_checkpoints_secs,
+            save_summaries_secs=cfg.save_summaries_steps,
             init_fn=init_fn)
         # Get a TensorFlow session managed by the supervisor.
         sess = sv.prepare_or_wait_for_session('')
@@ -218,7 +218,7 @@ def run(self, train_fn):
         step = 0
         while not sv.should_stop():
             step = sess.run(self._global_step)
-            if step > FLAGS.max_steps:
+            if step > cfg.max_steps:
                 break
             should_stop = train_fn(sess, step)
             if should_stop: