关于pytorch的出现nan如何解决

Question

import numpy as np
from torch.nn import functional as F
import h5py
import matplotlib.pyplot as plt
import lr_utils
import torch
import torch.nn as nn
import matplotlib.gridspec as grid_spec
train_set_x_orig , train_set_y , test_set_x_orig , test_set_y , classes = lr_utils.load_dataset()
train_x_flatten = train_set_x_orig.reshape(train_set_x_orig.shape[0], -1).T
test_x_flatten = test_set_x_orig.reshape(test_set_x_orig.shape[0], -1).T
train_x = train_x_flatten / 255
train_y = train_set_y
test_x = test_x_flatten / 255
test_y = test_set_y
print(train_x.shape,train_y.shape)
train_x=torch.from_numpy(train_x).float()
train_y=torch.from_numpy(train_y).float()
test_x=torch.from_numpy(test_x).float()
#test_y=torch.from_numpy(test_y).float()
W1=torch.randn([5,12288],requires_grad=True).float()
b1=torch.ones([1,1],requires_grad=True).float()
W2=torch.randn([1,5],requires_grad=True).float()
b2=torch.ones([1,1],requires_grad=True).float()
optimizer = torch.optim.SGD([W1,W2,b1,b2], 0.001)
cast=[]
print(test_x)
for step in range(10000):
z1=W1@train_x+b1
A1=torch.relu(z1)
z2=W2@A1+b2
A2=torch.sigmoid(z2)
print(A2.shape)
loss=F.binary_cross_entropy(train_y,A2)
loss.backward()
optimizer.step()
optimizer.zero_grad()
if step%100==0:
cast.append(loss)
testZ1=W1@test_x+b1
testA1=torch.relu(testZ1)
testZ2=W2@testA1+b2
testA2=torch.sigmoid(testZ2)
print(testA2.shape,testA2)
testA2=testA2.detach().numpy()
def predict(testA2):
for i in range(50):
if testA2[:,i]>0.5:
testA2[:,i]=1
else:
testA2[:,i]=0
return testA2
testA2=predict(testA2)
print(testA2,test_y)
acc2=np.mean(testA2==test_y)
print(acc2)
在做识别猫猫吴恩达深度学习时使用pytorch出现testA2的值都是nan，出现了问题，不知道代码哪里有错误。

Answer 1

debug吧，应该是数据或者读取数据出的问题