Experiments | Tin's Note

Created on March 21|Last edited on March 30
Comment
﻿
Dataset: Link: https://surreyac-my.sharepoint.com/:u:/r/personal/th01167_surrey_ac_uk/Documents/Surrey/Applied_ML/CelebA-HQ-split.zip?csf=1&web=1&e=1HnXoo﻿
Since the dataset on SurreyLearn only contain 2665 images (missing 300 images for testing, and have no labels).
I have downloaded and split from the original CelebA-HQ (1024x1024) here:
It includes caption, and attributes as well for our advanced exps.
train_27000 and test_3000 are randomly split from original 30000 images.
train_2700 and test_300 are chosen from train_27000 and test_3000 => We can use this small sets for hyper-parameters tuning before scaling up to 27000.
Unconditional Results
Butterfly﻿
CelebA-HQ (27000 train, 100 epochs + 300 val)DDIM (100 inference steps):
Calculating FID between: 
Real images dir: data/CelebA-HQ-split/test_300 
Fake images dir: outputs/samples/ddim_fast 
Resizing images to 128x128 
Processing real images from: data/CelebA-HQ-split/test_300 
Resolving data files: 100%|███████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 7121.30it/s] 
Found 300 images 
Finished processing 5 images 
Processing fake images from: outputs/samples/ddim_fast 
Resolving data files: 100%|███████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 7144.19it/s] 
Downloading data: 100%|████████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 6677.02files/s] 
Generating train split: 300 examples [00:00, 6668.00 examples/s] 
Found 300 images 
Finished processing 5 images 
Calculating final FID score... 
FID Score: 73.0482 
DDPM (1000 inference steps):
Calculating FID between: 
Real images dir: data/CelebA-HQ-split/test_300 
Fake images dir: outputs/samples/ddpm_1000steps 
Resizing images to 128x128 
Processing real images from: data/CelebA-HQ-split/test_300 
Resolving data files: 100%|██████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 7362.52it/s] 
Found 300 images 
Finished processing 5 images 
Processing fake images from: outputs/samples/ddpm_1000steps 
Resolving data files: 100%|██████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 7540.73it/s] 
Downloading data: 100%|███████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 7584.96files/s] 
Generating train split: 300 examples [00:00, 6333.70 examples/s] 
Found 300 images 
Finished processing 5 images 
Calculating final FID score... 
FID Score: 65.6536 
﻿
CelebA-HQ (27000 train, 100 epochs + 300 val)﻿
CelebA-HQ (TA's 2700 train, DDPM 2000 steps, 100 epochs)Calculating FID between:
Real images dir: data/celeba_hq_split/test
Fake images dir: outputs/samples/ddpm2000_2700train_100epochs
Resizing images to 128x128
Processing real images from: data/celeba_hq_split/test
Resolving data files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 964947.24it/s]
Found 300 images
Processed 16/300 images...
Processed 176/300 images...
Finished processing 300 images
Processing fake images from: outputs/samples/ddpm2000_2700train_100epochs
Resolving data files: 100%|████████████████████████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 119541.25it/s]
Downloading data: 100%|██████████████████████████████████████████████████████████████████████████████████████████████████| 300/300 [00:00<00:00, 91485.47files/s]
Generating train split: 300 examples [00:00, 25934.52 examples/s]
Found 300 images
Processed 16/300 images...
Processed 176/300 images...
Finished processing 300 images
Calculating final FID score...
FID Score: 103.7588
﻿
Baseline (TA's 2700 train)﻿https://wandb.ai/tin-hoang/EEEM068_Diffusion_Models/runs/5i75qk6c?nw=nwusertinhoang﻿
﻿
﻿
FP32 (TA's 2700 train)﻿https://wandb.ai/tin-hoang/EEEM068_Diffusion_Models/runs/tq5fgl86?nw=nwusertinhoang﻿
﻿
More smooth (less background noise).
Better capture small details like hairs.
﻿
﻿
Run set145
﻿
﻿
Add a comment