Model

Model(
   run_outputs, run_inputs = None, train_outputs = None, train_inputs = None,
   train_loss = None, eval_inputs = None, eval_outputs = None, eval_score = None,
   name = 'Model'
)

Base Model

Args

train_loss : either a callable, layer, or dictionary mapping a callable or
train_inputs : defaults to run inputs, if loss is provided you can either layer to the target outputs.

supply the inputs to the train graph that include the loss, or let the Model create inputs for you.

Methods:

.draw

source

.draw(
   path = 'graph.pdf'
)

.set_optimizer

source

.set_optimizer(
   optimizer, **config
)

Set the optimizer for this model

Optimizer Hyper-Parameters

The arguments passed to the optimizer constructor can be either regular Python values, tensors, or a callable. If they are callable, they will called during apply_gradients() to get the value for the hyper parameter.

Args

optimizer (Optimizer) : optimizer class or instance
config : dictionary with parameters for the optimizer, if you want to modify these parameters during training pass an tx.Param as the value for the given parameter instead of constant value.

Returns

optimizer (Optimizer) : the configured optimizer instance.

.run

source

.run(
   input_feed, compiled_graph = False
)

.train_step

source

.train_step(
   input_feed
)

.eval_step

source

.eval_step(
   input_feed
)

Args

input_feed:

Returns

*eval_output, eval_score ((eval outputs,eval score)):

.train

source

.train(
   train_data, validation_data = None, test_data = None, epochs = 1,
   steps_per_epoch = None, callbacks = []
)

Main training loop

Args

train_data : an iterable of dictionaries from Input Layers to values {Input:data}.
validation_data : an iterable of dictionaries from Input Layers to values {Input:data}.
test_data : an iterable of dictionaries from Input Layers to values {Input:data}.
epochs (int) : number of training epochs.
steps_per_epoch : number of steps in an epoch, if not None, epochs are incremented each time
callbacks : Callback functions scheduled during the training. (calling iter on this object should yield an iterator for an epoch.)

this number of steps pass even if the entire train_data has not been transversed.