E3_Distill_ESRGAN

GAN Distillation on Enhanced Super Resolution GAN

Link to ESRGAN Tranining Codes: Click Here

Knowledge Distillation on Enhanced Super Resolution GAN to perform Super Resolution on model with much smaller number of variables. The Training Algorithm is inspired from https://arxiv.org/abs/1902.00159, with a custom loss function specific to the problem of image super resolution.

Results

ESRGAN

Latency: 17.117 Seconds

Mean PSNR Achieved: 28.2

Sample:

Input Image Shape: 180 x 320

Output image shape: 720 x 1280

PSNR of the Image: 30.462

Compressed ESRGAN

Latency: 0.4 Seconds

Mean PSNR Achieved: 25.3

Sample

Input Image Shape: 180 x 320

Output image shape: 720 x 1280

PSNR of the Image: 26.942

Student Model

The Residual in Residual Architecture of ESRGAN was followed. With much shallower trunk. Specifically,

Name of Node	Depth
Residual Dense Blocks(RDB)	2 Depthwise Convolutions
Residual in Residual Blocks(RRDB)	2 RDB units
Trunk	3 RRDB units
UpSample Layer	1 ConvTranspose unit with a stride length of 4

Size of Growth Channel (intermediate channel) used: 32

Trained Saved Model and TF Lite File

Specification of the Saved Model

Input Dimension: [None, 180, 320, 3]

Input Data Type: Float32

Output Dimension: [None, 180, 320, 3]

TensorFlow loadable link: https://github.com/captain-pool/GSOC/releases/download/2.0.0/compressed_esrgan.tar.gz

Specification of TF Lite

Input Dimension: [1, 180, 320, 3]

Input Data Type: Float32

Output Dimension: [1, 720, 1280, 3]

TensorFlow Lite: https://github.com/captain-pool/GSOC/releases/download/2.0.0/compressed_esrgan.tflite