olympus.tasks.gan module

class olympus.tasks.gan.GAN(generator: torch.nn.modules.module.Module, discriminator: torch.nn.modules.module.Module, generator_optimizer: torch.optim.optimizer.Optimizer, discriminator_optimizer: torch.optim.optimizer.Optimizer, latent_vector_size: int = 10, criterion: torch.nn.modules.module.Module = CrossEntropyLoss())

Bases: olympus.tasks.task.Task

Attributes:

device

metrics

Methods

criterion(\*input, \*\*kwargs)
eval_loss(self, batch)	This is used to compute validation and test loss
fit(self, step, input, context)	Execute a single batch
get_space(self, \*\*fidelities)	Return missing hyper parameters that need to be set using init
init(self, \*\*kwargs)	Used to initialize the hyperparameters is any
load_state_dict(self, state[, strict])	Try to load a previous unfinished state to resume
state_dict(self[, destination, prefix, …])	Save a state the task can go back to if an error occur

accuracy
discriminate
discriminate_probabilities
finish
generate
report
resumed
set_device
summary

accuracy(self, batch, target)

criterion(*input, **kwargs) = CrossEntropyLoss()

discriminate(self, batch)

discriminate_probabilities(self, images)

finish(self)

fit(self, step, input, context)

Execute a single batch

Parameters:

epoch: int

current step in the training process

context: dict

Optional Context

Notes

You should wrap whatever code you have here inside a BadResumeGuard to prevent users from resuming a failed task that can have a bad states

To resume a task, you need to create a clean one with the same hyper parameters. It will pickup automatically where at its last checkpoint

generate(self, latent_vector)

latent_vector_size = 10