Single-Image-SVBRDF-Capture-rendering-loss
Single-Image-SVBRDF-Capture-rendering-loss copied to clipboard
valentin-deschaintre
Exception: input_dir does not exist
Hi, We are a student group looking into the paper and we are trying to make the code work and retrain the network on Google Colab (maybe with smaller dataset).
We added the untrained network on Google Colab, changed paths to match the input, output directories. We selected "train" as a parameter but we are getting this error:
Any idea what is happening?
here is the code beginning and folder structure
Best regards, Alin.
Hello,
What is the complete command line output when you try to run the method? What is the result if you add print(input_dir) line 360 (in the load_examples function)?
Out of curiosity, in which context are you looking into the paper?
Hi Valentin, This is the full output: error.pdf
This is the output after I add the print ("look here:" input:dir)
error with print (look here ).pdf
It is weird because the path is there with some training images...
We are master students following a basic ML course and we reviewed this project and the other 2 you have. The aim is to change some parameters (maybe a different optimizer) and retrain the network. We can maybe try with a real-life data set as well.
By the way is this network specifically using any pooling?
Ah! what is the course/university (this is simply out of curiosity)?
For the problem it seems to come from os.path.exists which doesn't manage to check existence on the drive or the path is incorrect (since you have just shown that input_dir is not none)
What happens if you run a new python console and simply run: import os os.path.exists('/content/drive/MyDrive/Colab Notebooks/Machine learning Programs/Algo 1 - Single image SVBRDF/inputExamples')
I tried a new file in Google Colab and mounted the drive and I tried the code you suggested. It came up as true. I tried the same code in the work file and it comes up as True as well... Maybe Google Colab is not the best for this task and there are some permission issues? Should we do it offline?
It's from Aarhus University Denmark - Technology Based Business Development :) Is it possible to contact you further on for clarifications regarding the paper? We are struggling with understanding some things.
And if you try to add to the code before the if statement:
print(input_dir is None) print(not os.path.exists(input_dir))
Which one comes back as True?
Yes you can contact me, I will try to help --in the limit of my available time though. To answer your previous question: there are not mean or max pooling in this architecture to reduce the resolution between convolution. The global feature track uses the average of the features (but I don't think this is what you meant?).
We also used your trained network to get an output from a 256 wall paper image.
With the parameter "eval" and we got these files:
There are the 4 feature maps, but what are the others? like the target and input ones?
Inputs is the input image (log corrected to visualize what the network received) and targets are the ground truth associated with the input image (this is only useful in test mode, in eval I replace it with black images since we don't have the ground truth parameters). I could have not save them but I didn't have the time to polish that much back then and it doesn't prevent usage.