def test_decorate(self):
param_names = [param.name for param in self.layer.parameters()]
self.optimizer = paddle.incubate.asp.decorate(self.optimizer)
program = paddle.static.default_main_program()
for name in param_names:
mask_var = ASPHelper._get_program_asp_info(program).mask_vars.get(
name, None)
if ASPHelper._is_supported_layer(program, name):
self.assertTrue(mask_var is not None)
else:
self.assertTrue(mask_var is None)
def test_save_and_load(self):
path = "/tmp/paddle_asp_save_dy/"
net_path = path + "asp_net.pdparams"
opt_path = path + "asp_opt.pdopt"
paddle.save(self.layer.state_dict(), net_path)
paddle.save(self.optimizer.state_dict(), opt_path)
asp_info = ASPHelper._get_program_asp_info(
paddle.static.default_main_program())
for param_name in asp_info.mask_vars:
mask = asp_info.mask_vars[param_name]
asp_info.update_mask_vars(
param_name, paddle.ones(shape=mask.shape, dtype=mask.dtype))
asp_info.update_masks(param_name, np.ones(shape=mask.shape))
net_state_dict = paddle.load(net_path)
opt_state_dict = paddle.load(opt_path)
self.layer.set_state_dict(net_state_dict)
self.optimizer.set_state_dict(opt_state_dict)
imgs = paddle.to_tensor(np.random.randn(64, 3, 32, 32),
dtype='float32',
place=self.place,
stop_gradient=False)
labels = paddle.to_tensor(np.random.randint(10, size=(64, 1)),
dtype='float32',
place=self.place,
stop_gradient=False)
loss_fn = paddle.nn.MSELoss(reduction='mean')
output = self.layer(imgs)
loss = loss_fn(output, labels)
loss.backward()
self.optimizer.step()
self.optimizer.clear_grad()
for param in self.layer.parameters():
if ASPHelper._is_supported_layer(
paddle.static.default_main_program(), param.name):
mat = param.numpy()
self.assertTrue(
paddle.fluid.contrib.sparsity.check_sparsity(mat.T,
n=2,
m=4))