Training Interface

Tensorpack trainers have a verbose interface for maximum flexibility. Then, there are interfaces built on top of trainers to simplify the use, when you don’t want to customize too much.

With ModelDesc and TrainConfig

This interface is enough for most types of single-cost tasks. A lot of examples are written in this interface.

SingleCost trainers expects 4 arguments to setup the graph: input signatures, InputSource, get_cost function, and an optimizer. ModelDesc describes a model by packing 3 of them together into one object:

class MyModel(ModelDesc):
  def inputs(self):
    return [tf.TensorSpec(shape, dtype, name), tf.TensorSpec(shape, dtype, name), ... ]

  def build_graph(self, tensorA, tensorB, ...):  # inputs
    # build the graph
    return cost   # returns the cost tensor

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

inputs() should define the metainfo of all the inputs your graph will take to build.

build_graph() takes inputs tensors that matches what you’ve defined in inputs().

You can use any symbolic functions in build_graph, including TensorFlow core library functions and other symbolic libraries. build_graph will be the tower function, so you need to follow some rules. Because this interface is specialized for single-cost training, you need to return the cost tensor.

After defining such a model, use it with TrainConfig and launch_train_with_config:

config = TrainConfig(
   # 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

trainer = SomeTrainer()
# trainer = SyncMultiGPUTrainerParameterServer(8)
launch_train_with_config(config, trainer)

See the docs of TrainConfig and launch_train_with_config for detailed functionalities.

The function launch_train_with_config(config, trainer) uses the raw trainer interface under the hood, and is almost equivalent to the following two lines of code:

    my_input_source,  # or QueueInput(my_dataflow)

If you need more control (e.g., if you want to construct the callbacks after setting up the graph), you can write the above two lines by yourself instead. You don’t need to construct a TrainConfig any more in that case. The function launch_train_with_config exists mainly for historical reasons.

Keras Interface

Some wrappers were made on top of tensorpack trainers, to create a Keras-like interface. See the experimental Tensorpack+Keras examples for details.

Raw Trainer Interface

To get a lower-level control, you can also access trainer methods directly:

Build the graph: For single-cost trainers, build the graph by calling SingleCostTrainer.setup_graph. For other types of tasks, you can build the graph by yourself.

Start training: Call Trainer.train() to start training, or call Trainer.train_with_defaults() which applies some defaults options for common use cases.

Read their API documentation and the advanced trainer tutorial for more details.