在调整loss计算的时候遇到了TypeError: only integer tensors of a single element can be converted to an index这个问题,原来的计算公式为:
self.loss_D = (self.loss_D_fake + self.loss_D_real) * 0.5
调整后的公式为:
# train.lcf[train.lcfCountD]为修正系数
self.loss_D = (self.loss_D_fake + self.loss_D_real) * 0.5 * train.lcf[train.lcfCountD]
但是运行时遇到报错:
Traceback (most recent call last):
File "train.py", line 59, in
model.optimize_parameters() # calculate loss functions, get gradients, update network weights
File "E:\imageProcessing\pix2pix_model.py", line 130, in optimize_parameters
self.backward_D() # calculate gradients for D
File "E:\imageProcessing\pix2pix_model.py", line 107, in backward_D
self.loss_D = (self.loss_D_fake + self.loss_D_real) * 0.5 * train.lcf[train.lcfCountD]
TypeError: only integer tensors of a single element can be converted to an index
但通过:
print(train.lcf[train.lcfCountD])
得知train.lcf[train.lcfCountD]的值为:
2.694736842105263
和前面乘的0.5是一样的,这是什么原因呢?再检查数据类型:
print(self.loss_D_fake + self.loss_D_real)
print((self.loss_D_fake + self.loss_D_real) * 0.5)
print(type(0.5))
print(type(train.lcf[train.lcfCountD]))
结果发现:
tensor(6.8438, device='cuda:0', grad_fn=)
tensor(3.4219, device='cuda:0', grad_fn=)
train.lcf[train.lcfCountD]数据居然是字符串!!!
通过强制类型转换:
print(type(train.lcf[train.lcfCountD]))
train.lcf[train.lcfCountD] = float(train.lcf[train.lcfCountD])
print(type(train.lcf[train.lcfCountD]))
结果终于正确:
至此这个问题已解决。
但是好奇为什么train.lcf[train.lcfCountD]是str呢,回溯它的来源:
引自train.py文件:
global lcf, lcfCountD, lcfCountG
lossCorrectFile = open('good_num_file.txt', 'r')
lcf = lossCorrectFile.readlines()
lcfCountD = 0
lcfCountG = 0
原来是文本读入的数据就是str型的,所以告诫了以后进行文本读写操作时,一定要注意数据类型啊!
DBzs 原创文章 9获赞 9访问量 599 关注 私信 展开阅读全文