MCPcopy
hub / github.com/zai-org/CogView / step

Method step

fp16/fp16.py:399–453  ·  view source on GitHub ↗

If no closure is supplied, :attr:`step` should be called after ``fp16_optimizer_obj.backward(loss)``. :attr:`step` updates the fp32 master copy of parameters using the optimizer supplied to :class:`FP16_Optimizer`'s constructor, then copies the updated fp32 params i

(self, closure=None)

Source from the content-addressed store, hash-verified

397 current.data.copy_(saved.data)
398
399 def step(self, closure=None): # could add clip option.
400 """
401 If no closure is supplied, :attr:`step` should be called after
402 ``fp16_optimizer_obj.backward(loss)``.
403 :attr:`step` updates the fp32 master copy of parameters using the optimizer supplied to
404 :class:`FP16_Optimizer`'s constructor, then copies the updated fp32 params into the fp16 params
405 originally referenced by :class:`FP16_Optimizer`'s constructor, so the user may immediately run
406 another forward pass using their model.
407
408 If a closure is supplied, :attr:`step` may be called without a prior call to
409 :attr:`backward(loss)`.
410 This control flow is identical to `ordinary Pytorch optimizer use`_ with closures.
411 However, the user should take care that any ``loss.backward()`` call within the closure
412 has been replaced by ``fp16_optimizer_obj.backward(loss)``.
413
414 Args:
415 closure (optional): Closure that will be supplied to the underlying optimizer originally passed to :class:`FP16_Optimizer`'s constructor. closure should call :attr:`zero_grad()` on the :class:`FP16_Optimizer` object, compute the loss, call :attr:`backward(loss)`, and return the loss.
416
417 Example with closure::
418
419 # optimizer is assumed to be an FP16_Optimizer object, previously constructed from an
420 # existing pytorch optimizer.
421 for input, target in dataset:
422 def closure():
423 optimizer.zero_grad()
424 output = model(input)
425 loss = loss_fn(output, target)
426 # loss.backward() becomes:
427 optimizer.backward(loss)
428 return loss
429 optimizer.step(closure)
430
431 .. warning::
432 Currently, calling :attr:`step` with a closure is not compatible with dynamic loss scaling.
433
434 .. _`ordinary Pytorch optimizer use`:
435 http://pytorch.org/docs/master/optim.html#optimizer-step-closure
436 """
437
438 scale = self.loss_scaler.loss_scale
439 self._update_scale(self.overflow)
440
441 if self.overflow:
442 self.maybe_print("OVERFLOW! Skipping step. Attempted loss scale: {}, reducing to {}"
443 .format(scale, self.loss_scale))
444 return
445
446 if closure is not None:
447 retval = self._step_with_closure(closure)
448 else:
449 retval = self.optimizer.step()
450
451 self._master_params_to_model_params()
452
453 return retval
454
455 def _step_with_closure(self, closure):
456 def wrapped_closure():

Callers 2

train_stepFunction · 0.45
_step_with_closureMethod · 0.45

Calls 4

_update_scaleMethod · 0.95
maybe_printMethod · 0.95
_step_with_closureMethod · 0.95

Tested by

no test coverage detected