Super Resolution GAN Project
Steps to connect to VM
N.B. Will only work if you've been approved to access the VM.
- Install G-Cloud CLI SDK
https://cloud.google.com/sdk/ - Connect to the instance
gcloud compute ssh ece285-tf-gpu-vm -- -L 8080:localhost:8080 - Open Jupyter notebook in browser by visiting:
localhost:8080
Installation Instructions All package requirements are in a pip requirements.txt file that can be installed with:
pip install -r requirements.txt
Loading dataset example utils.py contains scripts to import data and do some basic processing
import utils
path = glob.glob('./drive/My Drive/ECE285_Proj/datasets/VEDAI/*')
files = utils.scan_dataset(path)
training_set, testing_set = utils.create_subsets(files,'./drive/My Drive/ECE285_Proj/datasets/VEDAI/', use_validation = False)
rgb, infra = utils.read_VEDAI(training_set, path)
print('RGB shape: '+ str(rgb.shape))
print('Infra shape: '+ str(infra.shape))
training_data = normalize(combine_rgb_infra(rgb,infra))
print('Data shape: '+ str(training_data.shape))
OR
num_images = 50
files = utils.scan_dataset(data_dir, num_images)
training_set, testing_set = utils.create_subsets(files, data_dir, use_validation=False)
im_hr, im_lr, batch_idx = utils.load_data(0, training_set, data_dir, True, len(training_set))
Training
import DenseSRGAN
dir_pfx = './' # base file path
gan = DenseSRGAN.DenseSRGAN(dir_pfx,im_hr,im_lr,proj_pfx="OH",gpu_list=[1,3,5,7],dropout_rate=0.3,num_epochs_trained=450,weights_path='./weights/OH/')
gan.train(epochs=1000,verbose=False,bench_idx=2560,batch_size=16,save_interval=10,view_interval=2)
Predictions
# 1117,
ix = 10
img = gan.gen.predict(im_lr[ix:ix+1,:,:,:]).squeeze()
img = (img + 1)/2
plt.figure().suptitle('RGB+Infra', fontsize=20)
plt.subplot(1,2,1)
plt.imshow(im_hr[ix,:,:,:])
plt.subplot(1,2,2)
plt.imshow(img)