CNN for PV recognition in the urban environment

Authors: S. Roquette, M. Esguerra, A. Guerra and R. Castello (EPFL-LESOPB). Contacts: roberto.castello@epfl.ch

To submit a job to the cluster, the files in the cluster_files folder are used, model_main.py creates a new Unet model and it writes the weight and architecture in the directory pathaddress given a set of parameters passed as arguments: rotations, lights, dice, weight1, weight2, epochs -rotations is either a 1 or a 0 specifying if the input images contain or not rotations. -lights is either a 1 or a 0 specifying if the input images contain or not different lightnings.

• diceis 1 if the model uses dice loss function or 0 if it uses weighted cross entropy.
• weight1is the weight given to non-PV pixels errors if weighted cross entropy is used
• weight2is the weight given to PV pixels errors if weighted cross entropy is used
• epochs is the number of epochs the model will execute.

Images

For testing purposes, if you are running on the SCITAS cluster, you can simply access the folders SI_25_2015_1301-13, SI_25_2015_1301-32 and SI_25_2015_1261-33 located at:

/work/hyenergy/raw/SwissTopo/RGB_25cm/data_resized/

Please change the correspondent location in the data_loaders.py

Addition: the images originally used (50MB) for the students report can be downloaded using the following link:

https://enacshare.epfl.ch/fPbBRCfnDshpk3QAg6t7Kwz4VaEGyWN

In order to be able to run interactively, you want to download and save it in a Numpy_images/ folder, and run the Main.ipynb using settings memory = True, depending on your computer's power you will also want to set full_data = False.