The error "divsion by zero"

[sunnydai]

Excuse me! I'm new to the chainer framework and python. When i train with the command "python train.py", it always occur the error ''divsion by zero', i am confused! Hope your responce

[SeitaroShinagawa]

Hi @sunnydai,
Thanks for your report. Could you show me the actual error you have observed?

[sunnydai]

Thank you for your responce, the error is as follow!

[SeitaroShinagawa]

Oh, I see. The error occurs in division by "len(self.dataset)"
It means "len(self.dataset)==0." Namely, data loading has failed.

Firstly, could you check your dataset path in "common/paths.py"?
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting#set-dataset-path

#common/paths.py
train_place2 = "/yourpath/place2/data_256/"
val_place2 = "/yourpath/place2/val_256/"
test_place2 = "/yourpath/test_256/"

Note that the paths should be replaced by actual your dataset paths.

[sunnydai]

I have set my own datasets path. But my datasets is not as large as the original one, I just select some of the pictures and i use 'python train.py' instead of 'python train.py -g 0'. I don't know if these two changes make the error?

[SeitaroShinagawa]

Option -g means the gpu id you want to use as below,
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/train.py#L22
There is no relationship between -g and loading dataset.

Is your mask dataset which is generated by generate_windows.py also correct?
Default train.py uses image size of 256 as below,
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/train.py#L92

#actual command to generate mask dataset
python generate_windows.py 256 1000

[sunnydai]

Yes, I have generate the mask.

[sunnydai]

It maybe the failure of the loading of VGG model because the 'completing model loaded' is missed. And i find that the VGG model has been download, I don't konw why the error occur?

[SeitaroShinagawa]

Actually, 'Completion model' represents the trained model with this completion (in-painting) network, not the VGG model. It has been successful because you got 'ok' in the line of 64.

[sunnydai]

Oh~In ‘train.py’ line 92, mask_path=''mask/256", if it should be mask_path=''mask/256/"?

[SeitaroShinagawa]

Let me confirm your path is correct again. Could you run the following commands?

#to check train data path
>>> import glob
>>> from common import paths
>>> train_keys = glob.glob(paths.train_place2+"*/*/*.jpg")
>>> print(len(train_keys))
# if successful, you will get "1434892" as return

#to check mask path
>>> import glob
>>> from common import paths
>>> mask_keys = glob.glob("mask/256/*.bmp")
>>> print(len(mask_keys))
# if successful, you will get "1000" as return

[SeitaroShinagawa]

mask_path=''mask/256" is correct. See here,
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/datasets/place2.py#L20
(Sorry for confusing you, the comment in this line is my mistake.)

[sunnydai]

Thanks for your patience! I have find the question. I make a mistake in "place2.py"~

[sunnydai]

Another error occur when i am training, when the evaluation process is needed, there is an error 'numpy.ndarray' object has no attribute 'get'

[SeitaroShinagawa]

Sorry for the inconvenience. It seems that you executed the code with CPU mode, right?
"get" changes a cuda.array to numpy.array.
You can solve it by removing .get() in this case.

[sunnydai]

Thank you very much! It is ok now!

[sunnydai]

I am sorry to trouble you again. After training the model, when i execute the "generate_result.py", I can not load the model that i have trained. The model are saved in "result".

[SeitaroShinagawa]

generate_result.py requires model path option, "--load_model" as below,
https://github.com/SeitaroShinagawa/chainer-partial_convolution_image_inpainting/blob/master/generate_result.py#L26

#example
python gnerate_result.py -g 0 --load_model your/model/path

[sunnydai]

Thank you very much~

[greysou1]

@sunnydai thank you for this thread. I followed the suggestions in this thread but still ran into the same issue. I changed my path and double-checked. I used a small dataset the "Small images (256 * 256) with easy directory structure" 21G one. It only has the train and val data set so I copied the val data and named it train_256.
Can you please help me with the dataset structure?

UPDATE: I changed my dataset paths to a different location and now running this

#to check train data path

import glob
from common import paths
train_keys = glob.glob(paths.train_place2+"//*.jpg")
print(len(train_keys))
#if successful, you will get "1434892" as return

returns 1803460, I assume thats because of the different dataset but that's better than 0 which I used to get before. But the ZeroDivisionError: division by zero error persists. Can you please help?