def backward(self, loss: dy.Expression,
dynet_profiling: numbers.Integral) -> None:
"""
Perform backward pass to accumulate gradients.
Args:
loss: Result of self.training_step(...)
dynet_profiling: if > 0, print the computation graph
"""
if dynet_profiling and dynet_profiling > 0:
dy.print_text_graphviz()
loss.backward()
def update_weights(self, loss, trainer, dynet_profiling):
"""
Standardized way to perform backward pass and parameter updates.
:param loss: Result of self.training_step(...)
:param trainer: DyNet trainer / xnmt.optimizer object
:param dynet_profiling: if > 0, print the computation graph
"""
if dynet_profiling and dynet_profiling > 0:
dy.print_text_graphviz()
loss.backward()
trainer.update()
def update_weights(self, loss, trainer, dynet_profiling):
"""
Standardized way to perform backward pass and parameter updates.
Args:
loss: Result of self.training_step(...)
trainer (XnmtOptimizer): DyNet trainer
dynet_profiling (int): if > 0, print the computation graph
"""
if dynet_profiling and dynet_profiling > 0:
dy.print_text_graphviz()
loss.backward()
trainer.update()
def print_cg_conditional() -> None:
if settings.PRINT_CG_ON_ERROR:
dy.print_text_graphviz()