forked from mindspore-Ecosystem/mindspore
40 lines
1.2 KiB
Python
40 lines
1.2 KiB
Python
import numpy as np
|
|
|
|
import mindspore.nn as nn
|
|
from mindspore import Tensor, context
|
|
from mindspore.nn import Momentum
|
|
from mindspore.nn import WithLossCell, TrainOneStepCell
|
|
from mindspore.ops import operations as P
|
|
from mindspore.parallel._cost_model_context import set_cost_model_context
|
|
|
|
|
|
class Net(nn.Cell):
|
|
def __init__(self, input_ch, out_ch):
|
|
super(Net, self).__init__()
|
|
self.dense = nn.Dense(input_ch, out_ch)
|
|
self.relu = P.ReLU()
|
|
|
|
def construct(self, x):
|
|
x = self.dense(x)
|
|
x = self.relu(x)
|
|
return x
|
|
|
|
|
|
def test_inference_phase():
|
|
context.set_auto_parallel_context(device_num=8, global_rank=0)
|
|
context.set_auto_parallel_context(parallel_mode="auto_parallel")
|
|
set_cost_model_context(run_phase=1)
|
|
|
|
net = Net(512, 128)
|
|
predict = Tensor(np.ones([64, 512]).astype(np.float32) * 0.001)
|
|
label = Tensor(np.ones([64, 128]).astype(np.float32))
|
|
|
|
loss = nn.SoftmaxCrossEntropyWithLogits()
|
|
optimizer = Momentum(params=net.trainable_params(), learning_rate=0.1, momentum=0.9)
|
|
net_with_loss = WithLossCell(net, loss)
|
|
train_network = TrainOneStepCell(net_with_loss, optimizer)
|
|
train_network.set_train()
|
|
train_network.set_auto_parallel()
|
|
|
|
_ = train_network(predict, label)
|