def test_sgd_momentum_static():
_, data_shape, _, label_shape = get_input()
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01, momentum=0.9)
@trace
def f(data, label):
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
slots = TensorDict()
for param in mlp.parameters():
slots[param] = np.zeros(param.shape).astype(np.float32)
for _ in range(3):
f(
np.random.random(data_shape).astype(np.float32),
np.random.randint(0, 10, label_shape).astype(np.int32),
)
orig_params = TensorDict()
grads = TensorDict()
for param in mlp.parameters():
orig_params[param] = np.copy(param.numpy())
grads[param] = np.copy(param.grad.numpy())
opt.step()
for param in mlp.parameters():
slot = slots[param]
orig_param = orig_params[param]
slot *= 0.9
slot -= param.grad.numpy() * 0.01
assertTensorClose(param.numpy(), orig_param + slot)
def test_update_lr():
data, data_shape, label, label_shape = get_input()
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01)
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
opt.step()
for group in opt.param_groups:
group["lr"] += 0.02
for _ in range(3):
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
for param in mlp.parameters():
grad = F.grad(loss, param, use_virtual_grad=False)
assertTensorClose(grad.numpy(), param.grad.numpy())
orig_params = []
for param in mlp.parameters():
orig_params.append(np.copy(param.numpy()))
opt.step()
for param, orig_param in zip(mlp.parameters(), orig_params):
assertTensorClose(param.numpy(),
orig_param - param.grad.numpy() * 0.03)
def test_sgd_simple():
data, data_shape, label, label_shape = get_input()
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01, weight_decay=0.1)
for idx in range(3):
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
if idx % 2:
opt.zero_grad()
else:
mlp.zero_grad()
opt.backward(loss)
grads = TensorDict()
orig_params = TensorDict()
for param in mlp.parameters():
grad = F.grad(loss, param, use_virtual_grad=False)
assertTensorClose(grad.numpy(), param.grad.numpy())
grads[param] = np.copy(grad.numpy())
orig_params[param] = np.copy(param.numpy())
opt.step()
for param in mlp.parameters():
assertTensorClose(param.numpy(),
orig_params[param] * 0.999 - grads[param] * 0.01)
def _test_optimizer(opt_str, test_case, check_class, update_lr=False):
iter_num = 3
data, data_shape, label, label_shape = get_input()
net = MLP()
opt = getattr(optimizer, opt_str)(net.parameters(), **test_case)
check_func = check_class(net, **test_case)
step = 0
# eager graph
for i in range(iter_num):
if update_lr and i == 1: # change learning rate
for group in opt.param_groups:
group["lr"] += 0.01
check_func.lr += 0.01
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = net(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
ori_params = TensorDict()
for param in net.parameters():
ori_params[param] = np.copy(param.numpy())
opt.step()
step += 1
check_func(ori_params, net.parameters(), step)
# static graph
@trace
def train_func(data, label):
pred = net(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.backward(loss)
for i in range(iter_num):
if update_lr and i == 1: # change learning rate
for group in opt.param_groups:
group["lr"] += 0.01
check_func.lr += 0.01
opt.zero_grad()
ori_params = TensorDict()
for param in net.parameters():
ori_params[param] = np.copy(param.numpy())
train_func(
np.random.random(data_shape).astype(np.float32),
np.random.randint(0, 10, label_shape).astype(np.int32),
)
opt.step()
step += 1
check_func(ori_params, net.parameters(), step)
def test_compile_multi_times_static():
return # XXX: rewrite or remove this test
with Graph() as cg:
cg.set_option("eager_evaluation", False)
data = Input("data", shape=(2, 28))
label = Input("label", shape=(2, ), dtype=np.int32)
mlp = MLP()
opt = SGD(mlp.parameters(requires_grad=True), lr=0.01)
pred0 = mlp(data)
pred = F.softmax(pred0)
loss = F.square_loss(pred, label.reshape(2, 1))
opt.zero_grad()
grads = opt.backward(loss)
opt.step()
f0 = compile(pred, None)
f1 = compile([pred, loss], grads, copy=True)
data = np.random.random((2, 28)).astype(np.float32)
label = np.random.randint(0, 10, (2, )).astype(np.float32)
out0 = f0(data=data)
out1 = f1(data=data, label=label)
assertTensorClose(out0[0], out1[0])
_ = compile([pred, loss], grads, copy=False)
with pytest.raises(mgb.MegBrainError):
f0(data=data)
def test_multi_step_lr():
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01, momentum=0.9)
scheduler = MultiStepLR(opt, [3, 6, 8])
lr = np.array(0.01, dtype=np.float32)
for i in range(10):
for group in opt.param_groups:
assertTensorClose(
np.array(group["lr"], dtype=np.float32),
(lr * 0.1**bisect_right([3, 6, 8], i)).astype(np.float32),
max_err=5e-6,
)
scheduler.step()
def test_compile_multi_times_eager():
return # XXX: rewrite or remove this test
data = Input("data", shape=(2, 28))
label = Input("label", shape=(2, ), dtype=np.int32)
mlp = MLP()
opt = SGD(mlp.parameters(requires_grad=True), lr=0.01)
pred0 = mlp(data)
pred = F.softmax(pred0)
loss = F.square_loss(pred, label.reshape(2, 1))
opt.zero_grad()
grads = opt.backward(loss)
opt.step()
f0 = compile(pred, None)
f1 = compile([pred, loss], grads, copy=False)
for _ in range(3):
data = np.random.random((2, 28)).astype(np.float32)
label = np.random.randint(0, 10, (2, )).astype(np.float32)
out0 = f0(data=data)
out1 = f1(data=data, label=label)
assertTensorClose(out0[0], out1[0])
def test_optimizer_serialization():
data, data_shape, label, label_shape = get_input()
mlp = MLP()
opt = SGD(mlp.parameters(), lr=0.01, momentum=0.9)
slots = TensorDict()
for param in mlp.parameters():
slots[param] = np.zeros(param.shape).astype(np.float32)
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
opt.backward(loss)
opt.step()
for param in mlp.parameters():
slot = slots[param]
slot *= 0.9
slot -= param.grad.numpy() * 0.01
with BytesIO() as fout:
save(opt.state_dict(), fout)
fout.seek(0)
state_dict = load(fout)
opt1 = SGD(mlp.parameters(), lr=0.02, momentum=0.8)
opt1.load_state_dict(state_dict)
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt1.zero_grad()
opt1.backward(loss)
orig_params = TensorDict()
for param in mlp.parameters():
orig_params[param] = np.copy(param.numpy())
opt1.step()
for param in mlp.parameters():
orig_param = orig_params[param]
slot = slots[param]
slot *= 0.9
slot -= param.grad.numpy() * 0.01
assertTensorClose(param.numpy(), orig_param + slot)
def test_adam():
data, data_shape, label, label_shape = get_input()
mlp = MLP()
beta0 = 0.8
beta1 = 0.9
eps = 1e-4
opt = Adam(mlp.parameters(), lr=0.01, betas=(beta0, beta1), eps=eps)
m_slots = TensorDict()
v_slots = TensorDict()
for param in mlp.parameters():
m_slots[param] = np.zeros(param.shape).astype(np.float32)
v_slots[param] = np.zeros(param.shape).astype(np.float32)
step_size = 0
def check_value():
for param in mlp.parameters():
grad = param.grad.numpy()
orig_param = orig_params[param]
m = m_slots[param]
v = v_slots[param]
m *= beta0
m += (1 - beta0) * grad
v *= beta1
v += (1 - beta1) * grad * grad
update = (m / (1 - beta0**step_size)) / (
np.sqrt(v / (1 - beta1**step_size)) + eps)
assertTensorClose(param.numpy(), orig_param - 0.01 * update)
# eager
for _ in range(3):
data.set_value(np.random.random(data_shape).astype(np.float32))
label.set_value(np.random.randint(0, 10, label_shape))
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.zero_grad()
grads = opt.backward(loss)
orig_params = TensorDict()
for param in mlp.parameters():
orig_params[param] = np.copy(param.numpy())
opt.step()
step_size += 1
check_value()
# static
@trace
def f(data, label):
pred = mlp(data)
loss = F.square_loss(pred, label.reshape(-1, 1))
opt.backward(loss)
for _ in range(3):
opt.zero_grad()
orig_params = TensorDict()
for param in mlp.parameters():
orig_params[param] = np.copy(param.numpy())
f(
np.random.random(data_shape).astype(np.float32),
np.random.randint(0, 10, label_shape).astype(np.int32),
)
opt.step()
step_size += 1
check_value()