Spaces:
Build error
Build error
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from diffusers import DDPMPipeline
|
2 |
+
image_pipe = DDPMPipeline.from_pretrained("google/ddpm-celebahq-256")
|
3 |
+
image_pipe.to("cuda")
|
4 |
+
images = image_pipe().images
|
5 |
+
image_pipe
|
6 |
+
from diffusers import UNet2DModel
|
7 |
+
repo_id = "google/ddpm-church-256"
|
8 |
+
model = UNet2DModel.from_pretrained(repo_id)
|
9 |
+
model
|
10 |
+
model.config
|
11 |
+
model_random = UNet2DModel(**model.config)
|
12 |
+
model_random.save_pretrained("my_model")
|
13 |
+
model_random = UNet2DModel.from_pretrained("my_model")
|
14 |
+
import torch
|
15 |
+
torch.manual_seed(0)
|
16 |
+
noisy_sample = torch.randn(
|
17 |
+
1, model.config.in_channels, model.config.sample_size, model.config.sample_size
|
18 |
+
)
|
19 |
+
noisy_sample.shape
|
20 |
+
with torch.no_grad():
|
21 |
+
noisy_residual = model(sample=noisy_sample, timestep=2).sample
|
22 |
+
noisy_residual.shape
|
23 |
+
from diffusers import DDPMScheduler
|
24 |
+
scheduler = DDPMScheduler.from_config(repo_id)
|
25 |
+
scheduler.config
|
26 |
+
scheduler.save_config("my_scheduler")
|
27 |
+
new_scheduler = DDPMScheduler.from_config("my_scheduler")
|
28 |
+
less_noisy_sample = scheduler.step(
|
29 |
+
model_output=noisy_residual, timestep=2, sample=noisy_sample
|
30 |
+
).prev_sample
|
31 |
+
less_noisy_sample.shape
|
32 |
+
import PIL.Image
|
33 |
+
import numpy as np
|
34 |
+
def display_sample(sample, i):
|
35 |
+
image_processed = sample.cpu().permute(0, 2, 3, 1)
|
36 |
+
image_processed = (image_processed + 1.0) * 127.5
|
37 |
+
image_processed = image_processed.numpy().astype(np.uint8)
|
38 |
+
image_pil = PIL.Image.fromarray(image_processed[0])
|
39 |
+
display(f"Image at step {i}")
|
40 |
+
display(image_pil)
|
41 |
+
model.to("cuda")
|
42 |
+
noisy_sample = noisy_sample.to("cuda")
|
43 |
+
import tqdm
|
44 |
+
sample = noisy_sample
|
45 |
+
for i, t in enumerate(tqdm.tqdm(scheduler.timesteps)):
|
46 |
+
with torch.no_grad():
|
47 |
+
residual = model(sample, t).sample
|
48 |
+
sample = scheduler.step(residual, t, sample).prev_sample
|
49 |
+
if (i + 1) % 50 == 0:
|
50 |
+
display_sample(sample, i + 1)
|
51 |
+
from diffusers import DDIMScheduler
|
52 |
+
scheduler = DDIMScheduler.from_config(repo_id)
|
53 |
+
scheduler.set_timesteps(num_inference_steps=50)
|
54 |
+
import tqdm
|
55 |
+
sample = noisy_sample
|
56 |
+
for i, t in enumerate(tqdm.tqdm(scheduler.timesteps)):
|
57 |
+
with torch.no_grad():
|
58 |
+
residual = model(sample, t).sample
|
59 |
+
sample = scheduler.step(residual, t, sample).prev_sample
|
60 |
+
if (i + 1) % 10 == 0:
|
61 |
+
display_sample(sample, i + 1)
|