def __next__(self):
batch_input, batch_target = next(self.data_loader_iter)
tup = (batch_input, batch_target)
def is_tensor(obj):
return isinstance(obj, torch.Tensor)
to_device_fn = partial(torch.Tensor.to, device=self.device)
batch_input, batch_target = objwalk(tup, is_tensor, to_device_fn)
return batch_input, batch_target, self.kwargs
def test_objwalk(objwalk_objects):
start_obj = objwalk_objects[0]
ref_obj = objwalk_objects[1]
def is_target_class(obj):
return isinstance(obj, ObjwalkTestClass)
fn_to_apply = partial(ObjwalkTestClass.member_fn, val=OBJWALK_REF_VAL)
test_obj = objwalk(start_obj, is_target_class, fn_to_apply)
assert test_obj == ref_obj
def __next__(self):
if self.num_iter >= self._num_data_iter:
raise StopIteration
self.num_iter += 1
dataloader_output = next(self.data_loader_iter)
device = next(self._model.parameters()).device
to_device_fn = partial(torch.Tensor.to, device=device)
dataloader_output = objwalk(dataloader_output, is_tensor, to_device_fn)
args, kwargs = self._data_loader.get_inputs(dataloader_output)
self._model.zero_grad()
outputs = self._model(*args, **kwargs)
loss = self._criterion(outputs,
self._data_loader.get_target(dataloader_output))
loss.backward(create_graph=True)
grads = self._parameter_handler.get_gradients()
self._model.zero_grad()
return grads
def wrap_nncf_model_inputs_with_objwalk(model_args, model_kwargs):
model_args = objwalk(model_args, is_tensor, nncf_model_input)
model_kwargs = objwalk(model_kwargs, is_tensor, nncf_model_input)
return model_args, model_kwargs
def _infer_batch(self, args_kwargs_tuple, device):
to_device_fn = partial(torch.Tensor.to, device=device)
args, kwargs = objwalk(args_kwargs_tuple, is_tensor, to_device_fn)
self.model(*args, **kwargs)