forked from mrq/DL-Art-School
Make loss accumulator for logs accumulate better
This commit is contained in:
parent
4f8c4d130c
commit
e7957e4897
|
@ -22,7 +22,7 @@ class ConfigurableStep(Module):
|
|||
self.env = env
|
||||
self.opt = env['opt']
|
||||
self.gen_outputs = opt_step['generator_outputs']
|
||||
self.loss_accumulator = LossAccumulator()
|
||||
self.loss_accumulator = LossAccumulator(buffer_sz=opt_get(opt_step, ['loss_log_buffer'], 50))
|
||||
self.optimizers = None
|
||||
self.scaler = GradScaler(enabled=self.opt['fp16'])
|
||||
self.grads_generated = False
|
||||
|
|
Loading…
Reference in New Issue
Block a user