Update app.py
Browse files
app.py
CHANGED
@@ -10,7 +10,7 @@ from glide_text2im.model_creation import (
|
|
10 |
)
|
11 |
has_cuda = th.cuda.is_available()
|
12 |
device = th.device('cpu' if not has_cuda else 'cuda')
|
13 |
-
#
|
14 |
options = model_and_diffusion_defaults()
|
15 |
options['use_fp16'] = has_cuda
|
16 |
options['timestep_respacing'] = '100' # use 100 diffusion steps for fast sampling
|
@@ -21,7 +21,8 @@ if has_cuda:
|
|
21 |
model.to(device)
|
22 |
model.load_state_dict(load_checkpoint('base', device))
|
23 |
print('total base parameters', sum(x.numel() for x in model.parameters()))
|
24 |
-
|
|
|
25 |
options_up = model_and_diffusion_defaults_upsampler()
|
26 |
options_up['use_fp16'] = has_cuda
|
27 |
options_up['timestep_respacing'] = 'fast27' # use 27 diffusion steps for very fast sampling
|
@@ -38,31 +39,15 @@ def show_images(batch: th.Tensor):
|
|
38 |
reshaped = scaled.permute(2, 0, 3, 1).reshape([batch.shape[2], -1, 3])
|
39 |
display(Image.fromarray(reshaped.numpy()))
|
40 |
# Sampling parameters
|
41 |
-
prompt = "
|
42 |
batch_size = 1
|
43 |
guidance_scale = 3.0
|
44 |
|
45 |
# Tune this parameter to control the sharpness of 256x256 images.
|
46 |
# A value of 1.0 is sharper, but sometimes results in grainy artifacts.
|
47 |
upsample_temp = 0.997
|
48 |
-
|
49 |
-
samples = diffusion.p_sample_loop(
|
50 |
-
model_fn,
|
51 |
-
(full_batch_size, 3, options["image_size"], options["image_size"]),
|
52 |
-
device=device,
|
53 |
-
clip_denoised=True,
|
54 |
-
progress=True,
|
55 |
-
model_kwargs=model_kwargs,
|
56 |
-
cond_fn=None,
|
57 |
-
)[:batch_size]
|
58 |
-
model.del_cache()
|
59 |
-
|
60 |
-
# Show the output
|
61 |
-
show_images(samples)
|
62 |
-
|
63 |
-
|
64 |
import gradio as gr
|
65 |
-
def
|
66 |
# Set the prompt text
|
67 |
prompt = prompt
|
68 |
|
@@ -71,43 +56,54 @@ def generate_upsampled_image_from_text(prompt):
|
|
71 |
##############################
|
72 |
|
73 |
# Create the text tokens to feed to the model.
|
74 |
-
tokens =
|
75 |
-
tokens, mask =
|
76 |
-
tokens,
|
77 |
)
|
78 |
|
79 |
-
# Create the
|
80 |
-
|
81 |
-
|
82 |
-
|
|
|
83 |
|
84 |
-
|
|
|
85 |
tokens=th.tensor(
|
86 |
-
[tokens] * batch_size, device=device
|
87 |
),
|
88 |
mask=th.tensor(
|
89 |
-
[mask] * batch_size,
|
90 |
dtype=th.bool,
|
91 |
device=device,
|
92 |
),
|
93 |
)
|
94 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
# Sample from the base model.
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
up_shape,
|
101 |
-
noise=th.randn(up_shape, device=device) * upsample_temp,
|
102 |
device=device,
|
103 |
clip_denoised=True,
|
104 |
progress=True,
|
105 |
model_kwargs=model_kwargs,
|
106 |
cond_fn=None,
|
107 |
)[:batch_size]
|
108 |
-
|
109 |
|
110 |
# Show the output
|
111 |
-
show_images(
|
112 |
-
demo = gr.Interface(fn =
|
113 |
-
demo.launch()
|
|
|
10 |
)
|
11 |
has_cuda = th.cuda.is_available()
|
12 |
device = th.device('cpu' if not has_cuda else 'cuda')
|
13 |
+
# a base model.
|
14 |
options = model_and_diffusion_defaults()
|
15 |
options['use_fp16'] = has_cuda
|
16 |
options['timestep_respacing'] = '100' # use 100 diffusion steps for fast sampling
|
|
|
21 |
model.to(device)
|
22 |
model.load_state_dict(load_checkpoint('base', device))
|
23 |
print('total base parameters', sum(x.numel() for x in model.parameters()))
|
24 |
+
|
25 |
+
# Create an upsampler model.
|
26 |
options_up = model_and_diffusion_defaults_upsampler()
|
27 |
options_up['use_fp16'] = has_cuda
|
28 |
options_up['timestep_respacing'] = 'fast27' # use 27 diffusion steps for very fast sampling
|
|
|
39 |
reshaped = scaled.permute(2, 0, 3, 1).reshape([batch.shape[2], -1, 3])
|
40 |
display(Image.fromarray(reshaped.numpy()))
|
41 |
# Sampling parameters
|
42 |
+
prompt = ""
|
43 |
batch_size = 1
|
44 |
guidance_scale = 3.0
|
45 |
|
46 |
# Tune this parameter to control the sharpness of 256x256 images.
|
47 |
# A value of 1.0 is sharper, but sometimes results in grainy artifacts.
|
48 |
upsample_temp = 0.997
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
49 |
import gradio as gr
|
50 |
+
def generate_image_from_text(prompt):
|
51 |
# Set the prompt text
|
52 |
prompt = prompt
|
53 |
|
|
|
56 |
##############################
|
57 |
|
58 |
# Create the text tokens to feed to the model.
|
59 |
+
tokens = model.tokenizer.encode(prompt)
|
60 |
+
tokens, mask = model.tokenizer.padded_tokens_and_mask(
|
61 |
+
tokens, options['text_ctx']
|
62 |
)
|
63 |
|
64 |
+
# Create the classifier-free guidance tokens (empty)
|
65 |
+
full_batch_size = batch_size * 2
|
66 |
+
uncond_tokens, uncond_mask = model.tokenizer.padded_tokens_and_mask(
|
67 |
+
[], options['text_ctx']
|
68 |
+
)
|
69 |
|
70 |
+
# Pack the tokens together into model kwargs.
|
71 |
+
model_kwargs = dict(
|
72 |
tokens=th.tensor(
|
73 |
+
[tokens] * batch_size + [uncond_tokens] * batch_size, device=device
|
74 |
),
|
75 |
mask=th.tensor(
|
76 |
+
[mask] * batch_size + [uncond_mask] * batch_size,
|
77 |
dtype=th.bool,
|
78 |
device=device,
|
79 |
),
|
80 |
)
|
81 |
|
82 |
+
# Create a classifier-free guidance sampling function
|
83 |
+
def model_fn(x_t, ts, **kwargs):
|
84 |
+
half = x_t[: len(x_t) // 2]
|
85 |
+
combined = th.cat([half, half], dim=0)
|
86 |
+
model_out = model(combined, ts, **kwargs)
|
87 |
+
eps, rest = model_out[:, :3], model_out[:, 3:]
|
88 |
+
cond_eps, uncond_eps = th.split(eps, len(eps) // 2, dim=0)
|
89 |
+
half_eps = uncond_eps + guidance_scale * (cond_eps - uncond_eps)
|
90 |
+
eps = th.cat([half_eps, half_eps], dim=0)
|
91 |
+
return th.cat([eps, rest], dim=1)
|
92 |
+
|
93 |
# Sample from the base model.
|
94 |
+
model.del_cache()
|
95 |
+
samples = diffusion.p_sample_loop(
|
96 |
+
model_fn,
|
97 |
+
(full_batch_size, 3, options["image_size"], options["image_size"]),
|
|
|
|
|
98 |
device=device,
|
99 |
clip_denoised=True,
|
100 |
progress=True,
|
101 |
model_kwargs=model_kwargs,
|
102 |
cond_fn=None,
|
103 |
)[:batch_size]
|
104 |
+
model.del_cache()
|
105 |
|
106 |
# Show the output
|
107 |
+
show_images(samples)
|
108 |
+
demo = gr.Interface(fn =generate_image_from_text,inputs ="text",outputs ="image")
|
109 |
+
demo.launch()
|