def _sync_from_torch(self):
for child in self.children():
child._sync_from_torch()
for name, param in self.module.named_parameters(recurse=False):
getattr(self, name).array = tensor.asarray(param)
for name, buffer in self.module.named_buffers(recurse=False):
setattr(self, name, tensor.asarray(buffer))
def __init__(self, module):
super().__init__()
self._module = module
with self.init_scope():
for name, child in module.named_children():
setattr(self, name, TorchModule(child))
for name, param in module.named_parameters(recurse=False):
ch_param = chainer.Parameter(tensor.asarray(param))
setattr(self, name, ch_param)
# Gradient computed at PyTorch side is automatically
# synchronized to Chainer side with this hook.
param.register_hook(_get_grad_setter(ch_param))
for name, buffer in module.named_buffers(recurse=False):
self.add_persistent(name, tensor.asarray(buffer))
def test_asarray_multi_gpu():
if torch.cuda.device_count() < 2:
pytest.skip('Not enough GPUs')
t = torch.arange(5, dtype=torch.float32, device='cuda:1')
a = tensor.asarray(t)
assert isinstance(a, cupy.ndarray)
with cupy.cuda.Device(1):
a += 1
numpy.testing.assert_array_equal(a.get(), t.cpu().numpy())
def __init__(self, module):
super().__init__()
self._module = module
with self.init_scope():
for name, child in module.named_children():
if name == 'module':
# DataParallel objects have the model stored as `module`
# causing a conflict.
name = 'wrapped_module'
setattr(self, name, TorchModule(child))
for name, param in module.named_parameters(recurse=False):
ch_param = chainer.Parameter(tensor.asarray(param))
setattr(self, name, ch_param)
# Gradient computed at PyTorch side is automatically
# synchronized to Chainer side with this hook.
param.register_hook(_get_grad_setter(ch_param))
for name, buffer in module.named_buffers(recurse=False):
self.add_persistent(name, tensor.asarray(buffer))
def set_model(self):
chainer_model = MLP()
chainer_model.to_device(self.device)
self.device.use()
dummy_input = self.train_dataset[0][0]
dummy_input = chainer.Variable(tensor.asarray(dummy_input))
dummy_input.to_device(self.device)
chainer_model(dummy_input)
# dummy_input = iter(self.train_loader).next()
# dummy_input = chainer_prepare_batch(dummy_input, self.device)
# chainer_model(dummy_input[0][0])
self.model = cpm.LinkAsTorchModel(chainer_model)
def collate_to_array(batch):
data = torch.utils.data._utils.collate.default_collate(batch)
return [tensor.asarray(x) for x in data]
def hook(grad):
param.grad = tensor.asarray(grad)
def test_asarray_cpu():
t = torch.arange(5, dtype=torch.float32)
a = tensor.asarray(t)
assert isinstance(a, numpy.ndarray)
a += 1
numpy.testing.assert_array_equal(a, t.numpy())
def test_asarray_empty_gpu():
t = torch.tensor([], dtype=torch.float32, device='cuda')
a = tensor.asarray(t)
def test_asarray_empty_cpu():
t = torch.tensor([], dtype=torch.float32)
a = tensor.asarray(t)
def test_asarray_gpu():
t = torch.arange(5, dtype=torch.float32, device='cuda')
a = tensor.asarray(t)
assert isinstance(a, cupy.ndarray)
a += 1
numpy.testing.assert_array_equal(a.get(), t.cpu().numpy())
def test_asarray_empty_gpu():
t = torch.tensor([], dtype=torch.float32, device='cuda')
a = tensor.asarray(t)
assert isinstance(a, cupy.ndarray)
def test_asarray_empty_cpu():
t = torch.tensor([], dtype=torch.float32)
a = tensor.asarray(t)
assert isinstance(a, numpy.ndarray)