In research we do training of various kind. The only assumption tensorpack Trainer class makes about your training, is that your training follows this pattern:

for epoch_num in range(starting_epoch, max_epoch):
    for local_step in range(steps_per_epoch):
  1. Training is running some iterations. Tensorpack base trainer implements the logic of running the iteration. Users or derived trainers should implement what the iteration is.

  2. Trainer assumes the existence of "epoch", i.e. that the iterations run in double for-loops. But an epoch doesn't need to be a full pass of your dataset, the size of an epoch can be any number you set and it only affects the schedule of callbacks. In other words, an "epoch" in tensorpack is the default period to run callbacks (validation, summary, checkpoint, etc.).

Common Trainers

Most neural network training tasks are single-cost optimization. Tensorpack provides some trainer implementations for such tasks. These trainers will build the graph based on the given ModelDesc, and minimizes ModelDesc.cost.

To use trainers, pass a TrainConfig to configure them:

config = TrainConfig(
           # data=my_inputsource, # alternatively, use a customized InputSource

# start training:
SomeTrainer(config, other_arguments).train()

# start multi-GPU training with synchronous update:
# SyncMultiGPUTrainerParameterServer(config).train()

When you set the DataFlow (rather than the InputSource) in the config, tensorpack trainers automatically adopt certain prefetch mechanism, as mentioned in the Input Pipeline tutorial. You can set the InputSource instead, to customize this behavior.

Existing multi-GPU trainers include the logic of data-parallel training. You can enable them by just one line, and all the necessary logic to achieve the best performance was baked into the trainers already. The trainers can reach the same performance as the official tensorflow benchmark.

Please note that in data-parallel training, in each iteration all towers (all replicates of the model) will take tensors from the InputSource (instead of taking one for all and split). So the total batch size would be (batch size of InputSource/DataFlow) * #GPU.

Custom Trainers

You can easily write a trainer for other types of training. See Write a Trainer.