Trainer#

class Trainer(args, model=None, optimizer=None)#

Bases: object

This class represents a training object, that has all the needed components for a training, like dataLoaders, optimizer, model etc.

__init__(args, model=None, optimizer=None)#

Initializes the model, optimizer and data generators, based on user defined arguments. In case model and optimizer are passed - they are used, In case any of those arguments is None, the values are initialized from scratch, based on user arguments

Parameters:

  • args (Arguments) – user defined arguments

  • model (nn.Model) – Either a partually trained model (loaded from check point) or None. In case no model is passed, the model is loaded based on user arguments, by name

  • optimizer (Optimizer) – The optimizer to use during training (loaded from check point) or None. In case no model is passed, the optimizer is loaded based on user arguments

Methods

__init__(args[, model, optimizer])

Initializes the model, optimizer and data generators, based on user defined arguments.

approximate_input_range_from_data(args)

Runs the model on the given data.

get_all_ranges(args)

lists all activations ranges (min and max) in a single list

get_model()

Returns the model to be trained

get_optimizer()

Returns the optimizer of the training

get_range_aware_act(args)

test(args, epoch)

tests the model on the given data

train_step(args, epoch, t)

performs a single train step (a single forward and backward pass)

validation(args, epoch)

tests the model on the given data
approximate_input_range_from_data(args)#

Runs the model on the given data. As the result, each range aware activation in the model (if any) will hold the correct actual data range - minimum and maximum values of all inputs it recieved during the approximation run on the approximation data set. This ranges are used when approximation the range aware relu by range aware polynomial, as polynomial approximation requires a range

Parameters:

args (Arguments) – user arguments

get_all_ranges(args)#

lists all activations ranges (min and max) in a single list

Parameters:

args (Arguments) – user arguments

Returns:

ranges (two numbers for each activation, that represent minimum and maximum of the activation range)

Return type:

list<float>

get_model()#

Returns the model to be trained

Returns:

the model to be trained

Return type:

nn.model

get_optimizer()#

Returns the optimizer of the training

Returns:

the optimizer of the training

Return type:

optimizer

test(args, epoch: int)#

tests the model on the given data

Parameters:

  • args (Arguments) – user arguments

  • epoch (int) – epoch number

Raises:

Exception – Exit because of NaNs, if Nan values are present in the output, and args.continue_with_nans=False

Returns:

metrics np.array: confusion matrix

Return type:

MetricsTracker

train_step(args, epoch: int, t: tqdm)#

performs a single train step (a single forward and backward pass)

Parameters:

  • args (Arguments) – user arguments

  • epoch (int) – the current epoch number (for epoch summary printing)

  • t (tqdm) – training progress (the progress is updated after each epoch)

Raises:

Exception – Exit because of NaNs, if Nan values are present in the output, and args.continue_with_nans=False

Returns:

metrics np.array: confusion matrix

Return type:

MetricsTracker

validation(args, epoch: int)#

tests the model on the given data

Parameters:

  • args (Arguments) – user arguments

  • epoch (int) – epoch number

Raises:

Exception – Exit because of NaNs, if Nan values are present in the output, and args.continue_with_nans=False

Returns:

metrics np.array: confusion matrix

Return type:

MetricsTracker