deepdisc.training.trainers

Classes

`LazyAstroTrainer`
`LazyAstroEvaluator`

Functions

`return_lazy_trainer`(model, loader, optimizer, cfg, ...)	Return a trainer for models built on LazyConfigs
`return_savehook`(output_name, save_period)	Returns a hook for saving the model
`return_schedulerhook`(optimizer)	Returns a hook for the learning rate
`return_evallosshook`(val_per, model, test_loader)	Returns a hook for evaulating the loss
`return_optimizer`(cfg)	Returns an optimizer for training

Module Contents

class LazyAstroTrainer(model, data_loader, optimizer, cfg)[source]

Bases: detectron2.engine.SimpleTrainer

checkpointer[source]

lossList = [][source]

lossdict_epochs[source]

vallossList = [][source]

vallossdict_epochs[source]

period = 20[source]

iterCount = 0[source]

scheduler[source]

valloss = 0[source]

vallossdict[source]

set_period(p)[source]

run_step()[source]

classmethod build_lr_scheduler(cfg, optimizer)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

add_val_loss(val_loss)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

add_val_loss_dict(val_loss_dict)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

class LazyAstroEvaluator(model, data_loader, optimizer, cfg)[source]

Bases: detectron2.engine.SimpleTrainer

checkpointer[source]

lossList = [][source]

lossdict_epochs[source]

vallossList = [][source]

vallossdict_epochs[source]

period = 20[source]

iterCount = 0[source]

scheduler[source]

valloss = 0[source]

vallossdict[source]

set_period(p)[source]

run_step()[source]

classmethod build_lr_scheduler(cfg, optimizer)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

add_val_loss(val_loss)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

add_val_loss_dict(val_loss_dict)[source]: It now calls detectron2.solver.build_lr_scheduler(). Overwrite it if you’d like a different scheduler.

return_lazy_trainer(model, loader, optimizer, cfg, hooklist)[source]

Return a trainer for models built on LazyConfigs

Parameters:

model (torch model) – pointer to file
loader (detectron2 data loader)
optimizer (detectron2 optimizer)
cfg (.py file) – The LazyConfig used to build the model, and also stores config vals for data loaders
hooklist (list) – The list of hooks to use for the trainer

Return type:

trainer

return_savehook(output_name, save_period)[source]

Returns a hook for saving the model

Parameters:: output_name (str) – name of output file to save
Return type:: a SaveHook

return_schedulerhook(optimizer)[source]

Returns a hook for the learning rate

Parameters:: optimizer (detectron2 optimizer) – the optimizer that controls the learning rate
Return type:: a CustomLRScheduler hook

return_evallosshook(val_per, model, test_loader)[source]

Returns a hook for evaulating the loss

Parameters:

val_per (int) – the frequency with which to calculate validation loss
model (torch.nn.module) – the model
test_loader (data loader) – the loader to read in the eval data

Return type:

a LossEvalHook

return_optimizer(cfg)[source]

Returns an optimizer for training

Parameters:: cfg (.py file) – The LazyConfig used to build the model
Return type:: a pytorch optimizer