基于Pytorch的MLP实现基于Pytorch的MLP实现

import torch as pt
import torchvision as ptv
import numpy as np

train_set = ptv.datasets.MNIST("../../pytorch_database/mnist/train",train=True,transform=ptv.transforms.ToTensor(),download=True)
test_set = ptv.datasets.MNIST("../../pytorch_database/mnist/test",train=False,transform=ptv.transforms.ToTensor(),download=True)

train_dataset = pt.utils.data.DataLoader(train_set,batch_size=100)
test_dataset = pt.utils.data.DataLoader(test_set,batch_size=100)

class MLP(pt.nn.Module):
    def __init__(self):
        super(MLP,self).__init__()
        self.fc1 = pt.nn.Linear(784,512)
        self.fc2 = pt.nn.Linear(512,128)
        self.fc3 = pt.nn.Linear(128,10)
        
    def forward(self,din):
        din = din.view(-1,28*28)
        dout = pt.nn.functional.relu(self.fc1(din))
        dout = pt.nn.functional.relu(self.fc2(dout))
        return pt.nn.functional.softmax(self.fc3(dout))
model = MLP().cuda()
print(model)

MLP (
  (fc1): Linear (784 -> 512)
  (fc2): Linear (512 -> 128)
  (fc3): Linear (128 -> 10)
)

# loss func and optim
optimizer = pt.optim.SGD(model.parameters(),lr=0.01,momentum=0.9)
lossfunc = pt.nn.CrossEntropyLoss().cuda()

# accuarcy
def AccuarcyCompute(pred,label):
    pred = pred.cpu().data.numpy()
    label = label.cpu().data.numpy()
#     print(pred.shape(),label.shape())
    test_np = (np.argmax(pred,1) == label)
    test_np = np.float32(test_np)
    return np.mean(test_np)

# test accuarcy
# print(AccuarcyCompute(
#     np.array([[1,10,6],[0,2,5]],dtype=np.float32),
#     np.array([[1,2,8],[1,2,5]],dtype=np.float32)))

for x in range(4):
    for i,data in enumerate(train_dataset):
    
        optimizer.zero_grad()
    
        (inputs,labels) = data
        inputs = pt.autograd.Variable(inputs).cuda()
        labels = pt.autograd.Variable(labels).cuda()
    
        outputs = model(inputs)
    
        loss = lossfunc(outputs,labels)
        loss.backward()
    
        optimizer.step()
    
        if i % 100 == 0:
            print(i,":",AccuarcyCompute(outputs,labels))

0 : 0.9
100 : 0.84
200 : 0.82
300 : 0.88
400 : 0.9
500 : 0.92
0 : 0.93
100 : 0.91
200 : 0.9
300 : 0.91
400 : 0.9
500 : 0.91
0 : 0.93
100 : 0.91
200 : 0.94
300 : 0.91
400 : 0.93
500 : 0.92
0 : 0.96
100 : 0.94
200 : 0.95
300 : 0.91
400 : 0.93
500 : 0.94

accuarcy_list = []
for i,(inputs,labels) in enumerate(test_dataset):
    inputs = pt.autograd.Variable(inputs).cuda()
    labels = pt.autograd.Variable(labels).cuda()
    outputs = model(inputs)
    accuarcy_list.append(AccuarcyCompute(outputs,labels))
print(sum(accuarcy_list) / len(accuarcy_list))

0.936700002551

# only save paramters
pt.save(model.state_dict(),"../../pytorch_model/mlp/params/mlp_params.pt")

# save model
pt.save(model,"../../pytorch_model/mlp/model/mlp_model.pt")

/home/sky/virtualpython/pytorch0p2/lib/python3.5/site-packages/torch/serialization.py:147: UserWarning: Couldn't retrieve source code for container of type MLP. It won't be checked for correctness upon loading.
  "type " + obj.__name__ + ". It won't be checked "

test_save_net = MLP().cuda()
test_save_net.load_state_dict(pt.load("../../pytorch_model/mlp/params/mlp_params.pt"))
accuarcy_list = []
for i,(inputs,labels) in enumerate(test_dataset):
    inputs = pt.autograd.Variable(inputs).cuda()
    labels = pt.autograd.Variable(labels).cuda()
    outputs = model(inputs)
    accuarcy_list.append(AccuarcyCompute(outputs,labels))
print(sum(accuarcy_list) / len(accuarcy_list))

0.936700002551

test_save_model = pt.load("../../pytorch_model/mlp/model/mlp_model.pt")
accuarcy_list = []
for i,(inputs,labels) in enumerate(test_dataset):
    inputs = pt.autograd.Variable(inputs).cuda()
    labels = pt.autograd.Variable(labels).cuda()
    outputs = model(inputs)
    accuarcy_list.append(AccuarcyCompute(outputs,labels))
print(sum(accuarcy_list) / len(accuarcy_list))

0.936700002551