Tensorpack初学

官方文档：Tensorpack tutorial

• logger.auto_set_dir()，logger.set_logger_dir(dirname, action=None)

Use logger.set_logger_dir() to set log directory to “./train_log/{scriptname}:{name}”. “scriptname”就是当前python文件的名字
action 有三种类型：[“k”(keep),”d”(delete),”q”(quit)]

• BatchData()

假如：BatchData(dataset.Mnist(‘train’), 128)，len(Mnist(‘train’) )=60000，batch_size=128，最后的len(BatchData) = 60000//128 = 468；BatchData(dataset.Mnist(‘train’), 128, remainder=True)，那么len(BatchData) = 469。

• PrintData(train)，默认num=1

打印前几个Datapoints，方便debug。

• Mymodel(ModelDesc)

class MyModel(ModelDesc):
  def inputs(self):
  		"""
        Define all the inputs (with type, shape, name) that the graph will need.
        """
    return [tf.TensorSpec(shape, dtype, name), tf.TensorSpec(shape, dtype, name), ... ]

  def build_graph(self, tensorA, tensorB, ...):  # inputs
    	"""This function should build the model which takes the input variables (defined above)
        and return cost at the end."""
    return cost   # returns the cost tensor

  def optimizer(self):
    return tf.train.GradientDescentOptimizer(0.1)

• TrainConfig

config = TrainConfig(
   model=MyModel()
   dataflow=my_dataflow,
   # data=my_inputsource, # alternatively, use an InputSource
   callbacks=[...],    # some default callbacks are automatically applied
   # some default monitors are automatically applied
   steps_per_epoch=300,   # default to the size of your InputSource/DataFlow
)

• 模型保存

 ModelSaver(),   # save the model after every epoch
 MaxSaver('validation_accuracy'),  # save the model with highest accuracy (prefix 'validation_')

• Argscope

with argscope(Conv2D, filters=32, kernel_size=3, activation=tf.nn.relu):
  l = (LinearWrap(image)  # the starting brace is only for line-breaking
       .Conv2D('conv0')
       .MaxPooling('pool0', 2)
       .Conv2D('conv1', padding='SAME')
       .Conv2D('conv2', kernel_size=5)
       .FullyConnected('fc0', 512, activation=tf.nn.relu)
       .Dropout('dropout', rate=0.5)
       .tf.multiply(0.5)
       .apply(func, *args, **kwargs)
       .FullyConnected('fc1', units=10, activation=tf.identity)())

  with argscope([Conv2D, MaxPooling], data_format='NCHW'), argscope(Conv2D, kernel_size=1):
    l2 = LinearWrap(image).Conv2D('conv3', strides=2).MaxPooling('pool3', 2)

等价于下面：

l = Conv2D('conv0', image, 32, 3, activation=tf.nn.relu)
l = MaxPooling('pool0', l, 2)
l = Conv2D('conv1', l, 32, 3, padding='SAME', activation=tf.nn.relu)
l = Conv2D('conv2', l, 32, 5, activation=tf.nn.relu)
l = FullyConnected('fc0', l, 512, activation=tf.nn.relu)
l = Dropout('dropout', l, rate=0.5)
l = tf.multiply(l, 0.5)
l = func(l, *args, **kwargs)
l = FullyConnected('fc1', l, 10, activation=tf.identity)

l2 = Conv2D('conv3', image, 32, 1, strides=2, data_format='NCHW', activation=tf.nn.relu)
l2 = MaxPooling('pool3', l2, 2, data_format='NCHW')

- launch_train_with_config(config, trainer)

等价下面两个函数

trainer.setup_graph(
    my_model.get_input_signature(),
    my_input_source,  # or QueueInput(my_dataflow)
    my_model.build_graph,
    my_model.get_optimizer)
trainer.train_with_defaults(
    callbacks=config.callbacks,
    monitors=config.monitors,
    session_creator=config.session_creator,
    session_init=config.session_init,
    steps_per_epoch=config.steps_per_epoch,
    starting_epoch=config.starting_epoch,
    max_epoch=config.max_epoch,
    extra_callbacks=config.extra_callbacks)