CTC计算出来的loss非常大，大于60的loss, 不知道哪里出了问题

[joeqincai]

No description provided.

[ShangLe0607]

你好，请教一下，我运行create_lmdb_dataset.py出现not a valid image

戴志
/opt/crnn/images/9998.png is not a valid image
/opt/crnn/images/9999.png
窦君晶
/opt/crnn/images/9999.png is not a valid image
Created dataset with 0 samples
请问这是什么原因呢

[joeqincai]

说明这个图片没有读进来，你换其它格式的图片试试看------------------ 原始邮件 ------------------ 发件人: "ShangLe0607"<[email protected]> 发送时间: 2019年10月15日(星期二) 下午5:53 收件人: "AstarLight/Lets_OCR"<[email protected]>; 抄送: "joeqincai"<[email protected]>;"Author"<[email protected]>; 主题: Re: [AstarLight/Lets_OCR] CTC计算出来的loss非常大，大于60的loss, 不知道哪里出了问题 (#72) 你好，请教一下，我运行create_lmdb_dataset.py出现not a valid image 戴志 /opt/crnn/images/9998.png is not a valid image /opt/crnn/images/9999.png 窦君晶 /opt/crnn/images/9999.png is not a valid image Created dataset with 0 samples 请问这是什么原因呢 &mdash; You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[ShangLe0607]

@joeqincai 谢谢您，但是我试了试jpg格式还是同样的结果---、not valid image

[joeqincai]

你使用绝对路径去读一张图片试一下，我采用的是lmdb保存图片路径的方式，而不是保存图片字符串来做的

[ShangLe0607]

@joeqincai
改成了这种保存全路径的方式，还是同样的结果

[joeqincai]

def getitem(self, index):
assert index <= len(self), 'index range error'
index += 1
with self.env.begin(write=False) as txn:
img_key = 'image-%09d' % index
imgbuf = txn.get(img_key.encode())
imgAbsPath = imgbuf.decode()
# buf = six.BytesIO()
# buf.write(imgbuf)
# buf.seek(0)
try:
#img = Image.open(buf)
img = Image.open(imgAbsPath).convert('L')
except IOError:
print('Corrupted image for %d' % index)
return self[index + 1]
if self.transform is not None:
img = self.transform(img)
label_key = 'label-%09d' % index
label = txn.get(label_key.encode())
if self.target_transform is not None:
label = self.target_transform(label)
return (img, label)

[joeqincai]

def __getitem__(self, index): assert index <= len(self), 'index range error' index += 1 with self.env.begin(write=False) as txn: img_key = 'image-%09d' % index imgbuf = txn.get(img_key.encode()) imgAbsPath = imgbuf.decode() # buf = six.BytesIO() # buf.write(imgbuf) # buf.seek(0) try: #img = Image.open(buf) img = Image.open(imgAbsPath).convert('L') except IOError: print('Corrupted image for %d' % index) return self[index + 1] if self.transform is not None: img = self.transform(img) label_key = 'label-%09d' % index label = txn.get(label_key.encode()) if self.target_transform is not None: label = self.target_transform(label) return (img, label)

[joeqincai]

你需要先确定这个错误是在哪一句代码报错来的，然后再去尝试修复

[ShangLe0607]

我这个要得到test_lmdb和train_lmda的那个程序，结果是这样子，但确实能得到test_lmdb和train_lmda两个文件加,可是运行结果说是图片是无效，但整个程序并没有报错，

俞园翰
/opt/crnn/images/8996.jpg is not a valid image
/opt/crnn/images/8997.jpg
邵娴
/opt/crnn/images/8997.jpg is not a valid image
/opt/crnn/images/8998.jpg
苗茗
/opt/crnn/images/8998.jpg is not a valid image
/opt/crnn/images/8999.jpg
云姣谦
/opt/crnn/images/8999.jpg is not a valid image
Created dataset with 0 samples
[root@localhost lib]#

[joeqincai]

1、你要用编译器单步调试跟进去去看，在哪一句话报了这个错误；
2、能生成lmdb文件并不代表生成的就是你想要的