def __call__(self, param, grad, block):
"""Add L2 weight decay ops to network
Adds L2 weight decay ops.
L2WeightDecay = reg_coeff * parameter
Args:
param: parameter variable for which regularization is applied
block: block in which variable is to be created
Returns:
new variable for weight decay
"""
assert isinstance(param, framework.Variable)
assert isinstance(block, framework.Block)
if framework.in_dygraph_mode():
return _C_ops.scale(param, "scale", self._regularization_coeff)
else:
decay = block.create_var(dtype=param.dtype,
shape=param.shape,
lod_level=param.lod_level)
# Append Op to calculate decay
block.append_op(type='scale',
inputs={"X": param},
outputs={"Out": decay},
attrs={"scale": self._regularization_coeff})
return decay
def test(self):
input_data = np.ones([1, 1])
w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
out = _C_ops.scale(w, 'scale', 0.1)
out.backward()
w.grad.scale_(scale=0.5)
w._reset_grad_inplace_version(False)
assert w.grad._inplace_version() == 1
def test(self):
paddle.set_device('cpu')
input_data = np.ones([2, 2]).astype('float32')
w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
_clear_grad = clear_grad(w, a="1")
w._register_backward_hook(_clear_grad)
for i in range(10):
out = _C_ops.scale(w, 'scale', 0.1)
out.backward()
def test(self):
input_data = np.ones([1, 1])
w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
_clear_grad = clear_grad_test_0(w, a="1")
w._register_backward_hook(_clear_grad)
for i in range(2):
print(" Step: ", i)
out0 = _C_ops.scale(w, 'scale', 0.1)
out = _C_ops.matmul_v2(out0, w, 'trans_x', False, 'trans_y', False)
out.backward()
assert w.grad[0] == 0.15
def test(self):
input_data = np.ones([1, 1])
w = paddle.to_tensor(input_data, 'float32', stop_gradient=False)
c = Counter()
_clear_grad = clear_grad_test_1(w, c)
w._register_backward_hook(_clear_grad)
for c.step in range(5):
out0 = _C_ops.scale(w, 'scale', 0.1)
out = _C_ops.matmul_v2(out0, w, 'trans_x', False, 'trans_y', False)
out.backward()
if c.step == 1:
w.clear_gradient(False)
assert c.num_calls == 1
c.num_calls = 0
def _scalar_elementwise_op_(var, scale, bias):
return _C_ops.scale(var, 'scale', scale, 'bias', bias)
def _scalar_elementwise_op_(var, scale, bias):
if framework.in_dygraph_mode():
return _C_ops.final_state_scale(var, float(scale), bias, True)
return _C_ops.scale(var, 'scale', scale, 'bias', bias)
