[Week 5-Facade Parsing Using Deep Learning]
Team Members: furkan karababa, Onur Cankur, Javid Rajabov
This week we finished our progress report and submited it . Then we have focused on coding our basis model which is explained at last week’s blog post. We are currently working on our basis model.
Progress Report
We wrote methodology and experimental evaluation part of our report. In methodology part, we have explained how the our basis model we selected works. Then we mentioned how to improve this basis model. Details on how our basis model works and how to improve it is available in our progress report which you can see below.
In evaluation part, we have explained what datasets we will use on our basis model. Then we gave a detailed description of these datasets. For example, how many images are included in each data set, what labels etc … Finally, we explained the formula that we will use to calculate accuracy of our predictions. You can also see below details for Datasets and metric.
Running Basis Model
We started to work on the basis model we chose in the progress report.Our primary goal here was to run the source code with its own data set and check if it was working. We found 3 source code for the basis model we selected. The data set of each was different. We shared source codes among people in the group. Everyone has tried to run their own source code.
We have had a GPU problem when trying to run source codes, because the people who published the source code used the GPU when running the code. Otherwise, it took too long to run.To solve this problem, we used colaboratory, a GPU service provided by google for free. We downloaded the source code to google drive and then tried to run it using colaboratory.
Another problem we encountered was to install a dataset on google drive. The size of the dataset used by people who issued the source code was too large. This made it impossible for us to install google drive. To solve this problem, we have compressed the dataset as tar.gz. This makes it easier to drive the file. Then we opened the file tar.gz drive.
We managed to run the source code. Our next goal was to run the source code with our own datasets. But the format of the dataset we run is slightly different. We had to adapt our own dataset. We’re working on it right now. I’ve added the source code to the reference section, you can see it from there.
References
There are multiple models in this link. siftflow-fcn8s and voc-fcn8s are the source code we use. Here
Another source code we use. Here
