def test_pynative_resnet50():
context.set_context(mode=context.PYNATIVE_MODE, device_target="Ascend")
batch_size = 32
num_classes = 10
net = resnet50(batch_size, num_classes)
criterion = CrossEntropyLoss()
optimizer = Momentum(learning_rate=0.01, momentum=0.9,
params=filter(lambda x: x.requires_grad, net.get_parameters()))
net_with_criterion = WithLossCell(net, criterion)
net_with_criterion.set_grad()
train_network = GradWrap(net_with_criterion)
train_network.set_train()
step = 0
max_step = 21
exceed_num = 0
data_set = create_dataset(repeat_num=1, training=True, batch_size=batch_size)
for element in data_set.create_dict_iterator(num_epochs=1):
step = step + 1
if step > max_step:
break
start_time = time.time()
input_data = element["image"]
input_label = element["label"]
loss_output = net_with_criterion(input_data, input_label)
grads = train_network(input_data, input_label)
optimizer(grads)
end_time = time.time()
cost_time = end_time - start_time
print("======step: ", step, " loss: ", loss_output.asnumpy(), " cost time: ", cost_time)
if step > 1 and cost_time > 0.25:
exceed_num = exceed_num + 1
assert exceed_num < 20
def test_loss_scale_fp16_overflow():
inputs = Tensor(np.ones([16, 16]).astype(np.float32))
label = Tensor(np.zeros([16, 16]).astype(np.float32))
net = NetFP16(16, 16)
net.set_train()
loss = MSELoss()
optimizer = Lamb(net.trainable_params(), learning_rate=0.01)
net_with_loss = WithLossCell(net, loss)
net_with_loss.set_grad()
train_network = TrainOneStepWithLossScaleCell(net_with_loss, optimizer,
scale_sense=Tensor(np.full((1), np.finfo(np.float32).max),
dtype=mstype.float32))
output_1 = train_network(inputs, label)
output_2 = train_network(inputs, label)
assert output_1[0].asnumpy() == output_2[0].asnumpy()
assert output_1[1].asnumpy() == output_2[1].asnumpy() == True
