Spaces:
Runtime error
Runtime error
add stable diffusion painter for gradio app
Browse files- app.py +1 -1
- climategan/trainer.py +22 -17
- climategan/utils.py +2 -2
- climategan_wrapper.py +476 -0
- inferences.py +0 -108
app.py
CHANGED
@@ -6,7 +6,7 @@ import gradio as gr
|
|
6 |
import googlemaps
|
7 |
from skimage import io
|
8 |
from urllib import parse
|
9 |
-
from
|
10 |
|
11 |
|
12 |
def predict(api_key):
|
|
|
6 |
import googlemaps
|
7 |
from skimage import io
|
8 |
from urllib import parse
|
9 |
+
from climategan_wrapper import ClimateGAN
|
10 |
|
11 |
|
12 |
def predict(api_key):
|
climategan/trainer.py
CHANGED
@@ -223,7 +223,7 @@ class Trainer:
|
|
223 |
bin_value=-1,
|
224 |
half=False,
|
225 |
xla=False,
|
226 |
-
cloudy=
|
227 |
auto_resize_640=False,
|
228 |
ignore_event=set(),
|
229 |
return_masks=False,
|
@@ -308,24 +308,29 @@ class Trainer:
|
|
308 |
if xla:
|
309 |
xm.mark_step()
|
310 |
|
|
|
|
|
311 |
if numpy:
|
312 |
with Timer(store=stores.get("numpy", [])):
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
321 |
-
|
322 |
-
|
323 |
-
|
324 |
-
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
|
|
|
|
|
|
329 |
if return_masks:
|
330 |
output_data["mask"] = (
|
331 |
((mask > bin_value) * 255).cpu().numpy().astype(np.uint8)
|
|
|
223 |
bin_value=-1,
|
224 |
half=False,
|
225 |
xla=False,
|
226 |
+
cloudy=True,
|
227 |
auto_resize_640=False,
|
228 |
ignore_event=set(),
|
229 |
return_masks=False,
|
|
|
308 |
if xla:
|
309 |
xm.mark_step()
|
310 |
|
311 |
+
output_data = {}
|
312 |
+
|
313 |
if numpy:
|
314 |
with Timer(store=stores.get("numpy", [])):
|
315 |
+
if "flood" not in ignore_event:
|
316 |
+
# normalize to 0-1
|
317 |
+
flood = normalize(flood).cpu()
|
318 |
+
# convert to numpy
|
319 |
+
flood = flood.permute(0, 2, 3, 1).numpy()
|
320 |
+
# convert to 0-255 uint8
|
321 |
+
flood = (flood * 255).astype(np.uint8)
|
322 |
+
output_data["flood"] = flood
|
323 |
+
if "wildfire" not in ignore_event:
|
324 |
+
wildfire = normalize(wildfire).cpu()
|
325 |
+
wildfire = wildfire.permute(0, 2, 3, 1).numpy()
|
326 |
+
wildfire = (wildfire * 255).astype(np.uint8)
|
327 |
+
output_data["wildfire"] = wildfire
|
328 |
+
if "smog" not in ignore_event:
|
329 |
+
smog = normalize(smog).cpu()
|
330 |
+
smog = smog.permute(0, 2, 3, 1).numpy()
|
331 |
+
smog = (smog * 255).astype(np.uint8)
|
332 |
+
output_data["smog"] = smog
|
333 |
+
|
334 |
if return_masks:
|
335 |
output_data["mask"] = (
|
336 |
((mask > bin_value) * 255).cpu().numpy().astype(np.uint8)
|
climategan/utils.py
CHANGED
@@ -922,9 +922,9 @@ class Timer:
|
|
922 |
self.store = store
|
923 |
self.precision = precision
|
924 |
self.ignore = ignore
|
925 |
-
self.cuda = cuda
|
926 |
|
927 |
-
if cuda:
|
928 |
self._start_event = torch.cuda.Event(enable_timing=True)
|
929 |
self._end_event = torch.cuda.Event(enable_timing=True)
|
930 |
|
|
|
922 |
self.store = store
|
923 |
self.precision = precision
|
924 |
self.ignore = ignore
|
925 |
+
self.cuda = cuda if cuda is not None else torch.cuda.is_available()
|
926 |
|
927 |
+
if self.cuda:
|
928 |
self._start_event = torch.cuda.Event(enable_timing=True)
|
929 |
self._end_event = torch.cuda.Event(enable_timing=True)
|
930 |
|
climategan_wrapper.py
ADDED
@@ -0,0 +1,476 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# based on https://huggingface.co/spaces/NimaBoscarino/climategan/blob/main/inferences.py # noqa: E501
|
2 |
+
# thank you @NimaBoscarino
|
3 |
+
|
4 |
+
import re
|
5 |
+
from pathlib import Path
|
6 |
+
from uuid import uuid4
|
7 |
+
|
8 |
+
import numpy as np
|
9 |
+
import torch
|
10 |
+
from diffusers import StableDiffusionInpaintPipeline
|
11 |
+
from PIL import Image
|
12 |
+
from skimage.color import rgba2rgb
|
13 |
+
from skimage.transform import resize
|
14 |
+
|
15 |
+
from climategan.trainer import Trainer
|
16 |
+
|
17 |
+
|
18 |
+
def concat_events(output_dict, events, i=None, axis=1):
|
19 |
+
"""
|
20 |
+
Concatenates the `i`th data in `output_dict` according to the keys listed
|
21 |
+
in `events` on dimension `axis`.
|
22 |
+
|
23 |
+
Args:
|
24 |
+
output_dict (dict[Union[list[np.array], np.array]]): A dictionary mapping
|
25 |
+
events to their corresponding data :
|
26 |
+
{k: [HxWxC]} (for i != None) or {k: BxHxWxC}.
|
27 |
+
events (list[str]): output_dict's keys to concatenate.
|
28 |
+
axis (int, optional): Concatenation axis. Defaults to 1.
|
29 |
+
"""
|
30 |
+
cs = [e for e in events if e in output_dict]
|
31 |
+
if i is not None:
|
32 |
+
return uint8(np.concatenate([output_dict[c][i] for c in cs], axis=axis))
|
33 |
+
return uint8(np.concatenate([output_dict[c] for c in cs], axis=axis))
|
34 |
+
|
35 |
+
|
36 |
+
def clear(folder):
|
37 |
+
"""
|
38 |
+
Deletes all the images without the inference separator "---" in their name.
|
39 |
+
|
40 |
+
Args:
|
41 |
+
folder (Union[str, Path]): The folder to clear.
|
42 |
+
"""
|
43 |
+
for i in list(Path(folder).iterdir()):
|
44 |
+
if i.is_file() and "---" in i.stem:
|
45 |
+
i.unlink()
|
46 |
+
|
47 |
+
|
48 |
+
def uint8(array, rescale=False):
|
49 |
+
"""
|
50 |
+
convert an array to np.uint8 (does not rescale or anything else than changing dtype)
|
51 |
+
Args:
|
52 |
+
array (np.array): array to modify
|
53 |
+
Returns:
|
54 |
+
np.array(np.uint8): converted array
|
55 |
+
"""
|
56 |
+
if rescale:
|
57 |
+
if array.min() < 0:
|
58 |
+
if array.min() >= -1 and array.max() <= 1:
|
59 |
+
array = (array + 1) / 2
|
60 |
+
else:
|
61 |
+
raise ValueError(
|
62 |
+
f"Data range mismatch for image: ({array.min()}, {array.max()})"
|
63 |
+
)
|
64 |
+
if array.max() <= 1:
|
65 |
+
array = array * 255
|
66 |
+
return array.astype(np.uint8)
|
67 |
+
|
68 |
+
|
69 |
+
def resize_and_crop(img, to=640):
|
70 |
+
"""
|
71 |
+
Resizes an image so that it keeps the aspect ratio and the smallest dimensions
|
72 |
+
is `to`, then crops this resized image in its center so that the output is `to x to`
|
73 |
+
without aspect ratio distortion
|
74 |
+
Args:
|
75 |
+
img (np.array): np.uint8 255 image
|
76 |
+
Returns:
|
77 |
+
np.array: [0, 1] np.float32 image
|
78 |
+
"""
|
79 |
+
# resize keeping aspect ratio: smallest dim is 640
|
80 |
+
h, w = img.shape[:2]
|
81 |
+
if h < w:
|
82 |
+
size = (to, int(to * w / h))
|
83 |
+
else:
|
84 |
+
size = (int(to * h / w), to)
|
85 |
+
|
86 |
+
r_img = resize(img, size, preserve_range=True, anti_aliasing=True)
|
87 |
+
r_img = uint8(r_img)
|
88 |
+
|
89 |
+
# crop in the center
|
90 |
+
H, W = r_img.shape[:2]
|
91 |
+
|
92 |
+
top = (H - to) // 2
|
93 |
+
left = (W - to) // 2
|
94 |
+
|
95 |
+
rc_img = r_img[top : top + to, left : left + to, :]
|
96 |
+
|
97 |
+
return rc_img / 255.0
|
98 |
+
|
99 |
+
|
100 |
+
def to_m1_p1(img):
|
101 |
+
"""
|
102 |
+
rescales a [0, 1] image to [-1, +1]
|
103 |
+
Args:
|
104 |
+
img (np.array): float32 numpy array of an image in [0, 1]
|
105 |
+
i (int): Index of the image being rescaled
|
106 |
+
Raises:
|
107 |
+
ValueError: If the image is not in [0, 1]
|
108 |
+
Returns:
|
109 |
+
np.array(np.float32): array in [-1, +1]
|
110 |
+
"""
|
111 |
+
if img.min() >= 0 and img.max() <= 1:
|
112 |
+
return (img.astype(np.float32) - 0.5) * 2
|
113 |
+
raise ValueError(f"Data range mismatch for image: ({img.min()}, {img.max()})")
|
114 |
+
|
115 |
+
|
116 |
+
# No need to do any timing in this, since it's just for the HF Space
|
117 |
+
class ClimateGAN:
|
118 |
+
def __init__(self, model_path) -> None:
|
119 |
+
"""
|
120 |
+
A wrapper for the ClimateGAN model that you can use to generate
|
121 |
+
events from images or folders containing images.
|
122 |
+
|
123 |
+
Args:
|
124 |
+
model_path (Union[str, Path]): Where to load the Masker from
|
125 |
+
"""
|
126 |
+
torch.set_grad_enabled(False)
|
127 |
+
self.target_size = 640
|
128 |
+
self.trainer = Trainer.resume_from_path(
|
129 |
+
model_path,
|
130 |
+
setup=True,
|
131 |
+
inference=True,
|
132 |
+
new_exp=None,
|
133 |
+
)
|
134 |
+
self.trainer.G.half()
|
135 |
+
self._stable_diffusion_is_setup = False
|
136 |
+
|
137 |
+
def _setup_stable_diffusion(self):
|
138 |
+
"""
|
139 |
+
Sets up the stable diffusion pipeline for in-painting.
|
140 |
+
Make sure you have accepted the license on the model's card
|
141 |
+
https://huggingface.co/CompVis/stable-diffusion-v1-4
|
142 |
+
"""
|
143 |
+
try:
|
144 |
+
self.sdip_pipeline = StableDiffusionInpaintPipeline.from_pretrained(
|
145 |
+
"runwayml/stable-diffusion-inpainting",
|
146 |
+
revision="fp16",
|
147 |
+
torch_dtype=torch.float16,
|
148 |
+
safety_checker=None,
|
149 |
+
).to(self.trainer.device)
|
150 |
+
self._stable_diffusion_is_setup = True
|
151 |
+
except Exception as e:
|
152 |
+
print(
|
153 |
+
"\nCould not load stable diffusion model. "
|
154 |
+
+ "Please make sure you have accepted the license on the model's"
|
155 |
+
+ " card https://huggingface.co/CompVis/stable-diffusion-v1-4\n"
|
156 |
+
)
|
157 |
+
raise e
|
158 |
+
|
159 |
+
def _preprocess_image(self, img):
|
160 |
+
# rgba to rgb
|
161 |
+
data = img if img.shape[-1] == 3 else uint8(rgba2rgb(img) * 255)
|
162 |
+
|
163 |
+
# to args.target_size
|
164 |
+
data = resize_and_crop(data, self.target_size)
|
165 |
+
|
166 |
+
# resize() produces [0, 1] images, rescale to [-1, 1]
|
167 |
+
data = to_m1_p1(data)
|
168 |
+
return data
|
169 |
+
|
170 |
+
# Does all three inferences at the moment.
|
171 |
+
def infer_single(
|
172 |
+
self,
|
173 |
+
orig_image,
|
174 |
+
painter="both",
|
175 |
+
prompt="An HD picture of a street with dirty water after a heavy flood",
|
176 |
+
concats=[
|
177 |
+
"input",
|
178 |
+
"masked_input",
|
179 |
+
"climategan_flood",
|
180 |
+
"stable_flood",
|
181 |
+
"stable_copy_flood",
|
182 |
+
],
|
183 |
+
):
|
184 |
+
"""
|
185 |
+
Infers the image with the ClimateGAN model.
|
186 |
+
Importantly (and unlike self.infer_preprocessed_batch), the image is
|
187 |
+
pre-processed by self._preprocess_image before going through the networks.
|
188 |
+
|
189 |
+
Output dict contains the following keys:
|
190 |
+
- "input": The input image
|
191 |
+
- "mask": The mask used to generate the flood (from ClimateGAN's Masker)
|
192 |
+
- "masked_input": The input image with the mask applied
|
193 |
+
- "climategan_flood": The flooded image generated by ClimateGAN's Painter
|
194 |
+
on the masked input (only if "painter" is "climategan" or "both").
|
195 |
+
- "stable_flood": The flooded image in-painted by the stable diffusion model
|
196 |
+
from the mask and the input image (only if "painter" is "stable_diffusion"
|
197 |
+
or "both").
|
198 |
+
- "stable_copy_flood": The flooded image in-painted by the stable diffusion
|
199 |
+
model with its original context pasted back in:
|
200 |
+
y = m * flooded + (1-m) * input
|
201 |
+
(only if "painter" is "stable_diffusion" or "both").
|
202 |
+
|
203 |
+
Args:
|
204 |
+
orig_image (Union[str, np.array]): image to infer on. Can be a path to
|
205 |
+
an image which will be read.
|
206 |
+
painter (str, optional): Which painter to use: "climategan",
|
207 |
+
"stable_diffusion" or "both". Defaults to "both".
|
208 |
+
prompt (str, optional): The prompt used to guide the diffusion. Defaults
|
209 |
+
to "An HD picture of a street with dirty water after a heavy flood".
|
210 |
+
concats (list, optional): List of keys in `output` to concatenate together
|
211 |
+
in a new `{original_stem}_concat` image written. Defaults to:
|
212 |
+
["input", "masked_input", "climategan_flood", "stable_flood",
|
213 |
+
"stable_copy_flood"].
|
214 |
+
|
215 |
+
Returns:
|
216 |
+
dict: a dictionary containing the output images {k: HxWxC}. C is omitted
|
217 |
+
for masks (HxW).
|
218 |
+
"""
|
219 |
+
image_array = (
|
220 |
+
np.array(Image.open(orig_image))
|
221 |
+
if isinstance(orig_image, str)
|
222 |
+
else orig_image
|
223 |
+
)
|
224 |
+
image = self._preprocess_image(image_array)
|
225 |
+
output_dict = self.infer_preprocessed_batch(
|
226 |
+
image[None, ...], painter, prompt, concats
|
227 |
+
)
|
228 |
+
return {k: v[0] for k, v in output_dict.items()}
|
229 |
+
|
230 |
+
def infer_preprocessed_batch(
|
231 |
+
self,
|
232 |
+
images,
|
233 |
+
painter="both",
|
234 |
+
prompt="An HD picture of a street with dirty water after a heavy flood",
|
235 |
+
concats=[
|
236 |
+
"input",
|
237 |
+
"masked_input",
|
238 |
+
"climategan_flood",
|
239 |
+
"stable_flood",
|
240 |
+
"stable_copy_flood",
|
241 |
+
],
|
242 |
+
):
|
243 |
+
"""
|
244 |
+
Infers ClimateGAN predictions on a batch of preprocessed images.
|
245 |
+
It assumes that each image in the batch has been preprocessed with
|
246 |
+
self._preprocess_image().
|
247 |
+
|
248 |
+
Output dict contains the following keys:
|
249 |
+
- "input": The input image
|
250 |
+
- "mask": The mask used to generate the flood (from ClimateGAN's Masker)
|
251 |
+
- "masked_input": The input image with the mask applied
|
252 |
+
- "climategan_flood": The flooded image generated by ClimateGAN's Painter
|
253 |
+
on the masked input (only if "painter" is "climategan" or "both").
|
254 |
+
- "stable_flood": The flooded image in-painted by the stable diffusion model
|
255 |
+
from the mask and the input image (only if "painter" is "stable_diffusion"
|
256 |
+
or "both").
|
257 |
+
- "stable_copy_flood": The flooded image in-painted by the stable diffusion
|
258 |
+
model with its original context pasted back in:
|
259 |
+
y = m * flooded + (1-m) * input
|
260 |
+
(only if "painter" is "stable_diffusion" or "both").
|
261 |
+
|
262 |
+
Args:
|
263 |
+
images (np.array): A batch of input images BxHxWx3
|
264 |
+
painter (str, optional): Which painter to use: "climategan",
|
265 |
+
"stable_diffusion" or "both". Defaults to "both".
|
266 |
+
prompt (str, optional): The prompt used to guide the diffusion. Defaults
|
267 |
+
to "An HD picture of a street with dirty water after a heavy flood".
|
268 |
+
concats (list, optional): List of keys in `output` to concatenate together
|
269 |
+
in a new `{original_stem}_concat` image written. Defaults to:
|
270 |
+
["input", "masked_input", "climategan_flood", "stable_flood",
|
271 |
+
"stable_copy_flood"].
|
272 |
+
|
273 |
+
Returns:
|
274 |
+
dict: a dictionary containing the output images
|
275 |
+
"""
|
276 |
+
assert painter in [
|
277 |
+
"both",
|
278 |
+
"stable_diffusion",
|
279 |
+
"climategan",
|
280 |
+
], f"Unknown painter: {painter}"
|
281 |
+
|
282 |
+
ignore_event = set()
|
283 |
+
if painter == "climategan":
|
284 |
+
ignore_event.add("flood")
|
285 |
+
|
286 |
+
# Retrieve numpy events as a dict {event: array[BxHxWxC]}
|
287 |
+
outputs = self.trainer.infer_all(
|
288 |
+
images,
|
289 |
+
numpy=True,
|
290 |
+
bin_value=0.5,
|
291 |
+
half=True,
|
292 |
+
ignore_event=ignore_event,
|
293 |
+
return_masks=True,
|
294 |
+
)
|
295 |
+
|
296 |
+
outputs["input"] = uint8(images, True)
|
297 |
+
# from Bx1xHxW to BxHxWx1
|
298 |
+
outputs["masked_input"] = outputs["input"] * (
|
299 |
+
outputs["mask"].squeeze(1)[..., None] == 0
|
300 |
+
)
|
301 |
+
|
302 |
+
if painter in {"both", "climategan"}:
|
303 |
+
outputs["climategan_flood"] = outputs.pop("flood")
|
304 |
+
else:
|
305 |
+
del outputs["flood"]
|
306 |
+
|
307 |
+
if painter != "climategan":
|
308 |
+
if not self._stable_diffusion_is_setup:
|
309 |
+
print("Setting up stable diffusion in-painting pipeline")
|
310 |
+
self._setup_stable_diffusion()
|
311 |
+
|
312 |
+
mask = outputs["mask"].squeeze(1)
|
313 |
+
input_images = (
|
314 |
+
torch.tensor(images).permute(0, 3, 1, 2).to(self.trainer.device)
|
315 |
+
)
|
316 |
+
input_mask = torch.tensor(mask[:, None, ...] > 0).to(self.trainer.device)
|
317 |
+
floods = self.sdip_pipeline(
|
318 |
+
prompt=[prompt] * images.shape[0],
|
319 |
+
image=input_images,
|
320 |
+
mask_image=input_mask,
|
321 |
+
height=640,
|
322 |
+
width=640,
|
323 |
+
num_inference_steps=50,
|
324 |
+
)
|
325 |
+
|
326 |
+
bin_mask = mask[..., None] > 0
|
327 |
+
flood = np.stack([np.array(i) for i in floods.images])
|
328 |
+
copy_flood = flood * bin_mask + uint8(images, True) * (1 - bin_mask)
|
329 |
+
outputs["stable_flood"] = flood
|
330 |
+
outputs["stable_copy_flood"] = copy_flood
|
331 |
+
|
332 |
+
if concats:
|
333 |
+
outputs["concat"] = concat_events(outputs, concats, axis=2)
|
334 |
+
|
335 |
+
return {k: v.squeeze(1) if v.shape[1] == 1 else v for k, v in outputs.items()}
|
336 |
+
|
337 |
+
def infer_folder(
|
338 |
+
self,
|
339 |
+
folder_path,
|
340 |
+
painter="both",
|
341 |
+
prompt="An HD picture of a street with dirty water after a heavy flood",
|
342 |
+
batch_size=4,
|
343 |
+
concats=[
|
344 |
+
"input",
|
345 |
+
"masked_input",
|
346 |
+
"climategan_flood",
|
347 |
+
"stable_flood",
|
348 |
+
"stable_copy_flood",
|
349 |
+
],
|
350 |
+
write=True,
|
351 |
+
overwrite=False,
|
352 |
+
):
|
353 |
+
"""
|
354 |
+
Infers the images in a folder with the ClimateGAN model, batching images for
|
355 |
+
inference according to the batch_size.
|
356 |
+
|
357 |
+
Images must end in .jpg, .jpeg or .png (not case-sensitive).
|
358 |
+
Images must not contain the separator ("---") in their name.
|
359 |
+
|
360 |
+
Images will be written to disk in the same folder as the input images, with
|
361 |
+
a name that depends on its data, potentially the prompt and a random
|
362 |
+
identifier in case multiple inferences are run in the folder.
|
363 |
+
|
364 |
+
Output dict contains the following keys:
|
365 |
+
- "input": The input image
|
366 |
+
- "mask": The mask used to generate the flood (from ClimateGAN's Masker)
|
367 |
+
- "masked_input": The input image with the mask applied
|
368 |
+
- "climategan_flood": The flooded image generated by ClimateGAN's Painter
|
369 |
+
on the masked input (only if "painter" is "climategan" or "both").
|
370 |
+
- "stable_flood": The flooded image in-painted by the stable diffusion model
|
371 |
+
from the mask and the input image (only if "painter" is "stable_diffusion"
|
372 |
+
or "both").
|
373 |
+
- "stable_copy_flood": The flooded image in-painted by the stable diffusion
|
374 |
+
model with its original context pasted back in:
|
375 |
+
y = m * flooded + (1-m) * input
|
376 |
+
(only if "painter" is "stable_diffusion" or "both").
|
377 |
+
|
378 |
+
Args:
|
379 |
+
folder_path (Union[str, Path]): Where to read images from.
|
380 |
+
painter (str, optional): Which painter to use: "climategan",
|
381 |
+
"stable_diffusion" or "both". Defaults to "both".
|
382 |
+
prompt (str, optional): The prompt used to guide the diffusion. Defaults
|
383 |
+
to "An HD picture of a street with dirty water after a heavy flood".
|
384 |
+
batch_size (int, optional): Size of inference batches. Defaults to 4.
|
385 |
+
concats (list, optional): List of keys in `output` to concatenate together
|
386 |
+
in a new `{original_stem}_concat` image written. Defaults to:
|
387 |
+
["input", "masked_input", "climategan_flood", "stable_flood",
|
388 |
+
"stable_copy_flood"].
|
389 |
+
write (bool, optional): Whether or not to write the outputs to the input
|
390 |
+
folder.Defaults to True.
|
391 |
+
overwrite (Union[bool, str], optional): Whether to overwrite the images or
|
392 |
+
not. If a string is provided, it will be included in the name.
|
393 |
+
Defaults to False.
|
394 |
+
|
395 |
+
Returns:
|
396 |
+
dict: a dictionary containing the output images
|
397 |
+
"""
|
398 |
+
folder_path = Path(folder_path).expanduser().resolve()
|
399 |
+
assert folder_path.exists(), f"Folder {str(folder_path)} does not exist"
|
400 |
+
assert folder_path.is_dir(), f"{str(folder_path)} is not a directory"
|
401 |
+
im_paths = [
|
402 |
+
p
|
403 |
+
for p in folder_path.iterdir()
|
404 |
+
if p.suffix.lower() in [".jpg", ".png", ".jpeg"] and "---" not in p.name
|
405 |
+
]
|
406 |
+
assert im_paths, f"No images found in {str(folder_path)}"
|
407 |
+
ims = [self._preprocess_image(np.array(Image.open(p))) for p in im_paths]
|
408 |
+
batches = [
|
409 |
+
np.stack(ims[i : i + batch_size]) for i in range(0, len(ims), batch_size)
|
410 |
+
]
|
411 |
+
inferences = [
|
412 |
+
self.infer_preprocessed_batch(b, painter, prompt, concats) for b in batches
|
413 |
+
]
|
414 |
+
|
415 |
+
outputs = {
|
416 |
+
k: [i for e in inferences for i in e[k]] for k in inferences[0].keys()
|
417 |
+
}
|
418 |
+
|
419 |
+
if write:
|
420 |
+
self.write(outputs, im_paths, painter, overwrite, prompt)
|
421 |
+
|
422 |
+
return outputs
|
423 |
+
|
424 |
+
def write(
|
425 |
+
self,
|
426 |
+
outputs,
|
427 |
+
im_paths,
|
428 |
+
painter="both",
|
429 |
+
overwrite=False,
|
430 |
+
prompt="",
|
431 |
+
):
|
432 |
+
"""
|
433 |
+
Writes the outputs of the inference to disk, in the input folder.
|
434 |
+
|
435 |
+
Images will be named like:
|
436 |
+
f"{original_stem}---{overwrite_prefix}_{painter_type}_{output_type}.{suffix}"
|
437 |
+
`painter_type` is either "climategan" or f"stable_diffusion_{prompt}"
|
438 |
+
|
439 |
+
Args:
|
440 |
+
outputs (_type_): The inference procedure's output dict.
|
441 |
+
im_paths (list[Path]): The list of input images paths.
|
442 |
+
painter (str, optional): Which painter was used. Defaults to "both".
|
443 |
+
overwrite (bool, optional): Whether to overwrite the images or not.
|
444 |
+
If a string is provided, it will be included in the name.
|
445 |
+
If False, a random identifier will be added to the name.
|
446 |
+
Defaults to False.
|
447 |
+
prompt (str, optional): The prompt used to guide the diffusion. Defaults
|
448 |
+
to "".
|
449 |
+
"""
|
450 |
+
prompt = re.sub("[^0-9a-zA-Z]+", "", prompt).lower()
|
451 |
+
overwrite_prefix = ""
|
452 |
+
if not overwrite:
|
453 |
+
overwrite_prefix = str(uuid4())[:8]
|
454 |
+
print("Writing events with prefix", overwrite_prefix)
|
455 |
+
else:
|
456 |
+
if isinstance(overwrite, str):
|
457 |
+
overwrite_prefix = overwrite
|
458 |
+
print("Writing events with prefix", overwrite_prefix)
|
459 |
+
|
460 |
+
# for each image, for each event/data type
|
461 |
+
for i, im_path in enumerate(im_paths):
|
462 |
+
for event, ims in outputs.items():
|
463 |
+
painter_prefix = ""
|
464 |
+
if painter == "climategan" and event == "flood":
|
465 |
+
painter_prefix = "climategan"
|
466 |
+
elif (
|
467 |
+
painter in {"stable_diffusion", "both"} and event == "stable_flood"
|
468 |
+
):
|
469 |
+
painter_prefix = f"_stable_{prompt}"
|
470 |
+
elif painter == "both" and event == "climategan_flood":
|
471 |
+
painter_prefix = ""
|
472 |
+
|
473 |
+
im = ims[i]
|
474 |
+
im = Image.fromarray(uint8(im))
|
475 |
+
imstem = f"{im_path.stem}---{overwrite_prefix}{painter_prefix}_{event}"
|
476 |
+
im.save(im_path.parent / (imstem + im_path.suffix))
|
inferences.py
DELETED
@@ -1,108 +0,0 @@
|
|
1 |
-
# based on https://huggingface.co/spaces/NimaBoscarino/climategan/blob/main/inferences.py # noqa: E501
|
2 |
-
# thank you @NimaBoscarino
|
3 |
-
|
4 |
-
import torch
|
5 |
-
from skimage.color import rgba2rgb
|
6 |
-
from skimage.transform import resize
|
7 |
-
import numpy as np
|
8 |
-
|
9 |
-
from climategan.trainer import Trainer
|
10 |
-
|
11 |
-
|
12 |
-
def uint8(array):
|
13 |
-
"""
|
14 |
-
convert an array to np.uint8 (does not rescale or anything else than changing dtype)
|
15 |
-
Args:
|
16 |
-
array (np.array): array to modify
|
17 |
-
Returns:
|
18 |
-
np.array(np.uint8): converted array
|
19 |
-
"""
|
20 |
-
return array.astype(np.uint8)
|
21 |
-
|
22 |
-
|
23 |
-
def resize_and_crop(img, to=640):
|
24 |
-
"""
|
25 |
-
Resizes an image so that it keeps the aspect ratio and the smallest dimensions
|
26 |
-
is `to`, then crops this resized image in its center so that the output is `to x to`
|
27 |
-
without aspect ratio distortion
|
28 |
-
Args:
|
29 |
-
img (np.array): np.uint8 255 image
|
30 |
-
Returns:
|
31 |
-
np.array: [0, 1] np.float32 image
|
32 |
-
"""
|
33 |
-
# resize keeping aspect ratio: smallest dim is 640
|
34 |
-
h, w = img.shape[:2]
|
35 |
-
if h < w:
|
36 |
-
size = (to, int(to * w / h))
|
37 |
-
else:
|
38 |
-
size = (int(to * h / w), to)
|
39 |
-
|
40 |
-
r_img = resize(img, size, preserve_range=True, anti_aliasing=True)
|
41 |
-
r_img = uint8(r_img)
|
42 |
-
|
43 |
-
# crop in the center
|
44 |
-
H, W = r_img.shape[:2]
|
45 |
-
|
46 |
-
top = (H - to) // 2
|
47 |
-
left = (W - to) // 2
|
48 |
-
|
49 |
-
rc_img = r_img[top : top + to, left : left + to, :]
|
50 |
-
|
51 |
-
return rc_img / 255.0
|
52 |
-
|
53 |
-
|
54 |
-
def to_m1_p1(img):
|
55 |
-
"""
|
56 |
-
rescales a [0, 1] image to [-1, +1]
|
57 |
-
Args:
|
58 |
-
img (np.array): float32 numpy array of an image in [0, 1]
|
59 |
-
i (int): Index of the image being rescaled
|
60 |
-
Raises:
|
61 |
-
ValueError: If the image is not in [0, 1]
|
62 |
-
Returns:
|
63 |
-
np.array(np.float32): array in [-1, +1]
|
64 |
-
"""
|
65 |
-
if img.min() >= 0 and img.max() <= 1:
|
66 |
-
return (img.astype(np.float32) - 0.5) * 2
|
67 |
-
raise ValueError(f"Data range mismatch for image: ({img.min()}, {img.max()})")
|
68 |
-
|
69 |
-
|
70 |
-
# No need to do any timing in this, since it's just for the HF Space
|
71 |
-
class ClimateGAN:
|
72 |
-
def __init__(self, model_path) -> None:
|
73 |
-
torch.set_grad_enabled(False)
|
74 |
-
self.target_size = 640
|
75 |
-
self.trainer = Trainer.resume_from_path(
|
76 |
-
model_path,
|
77 |
-
setup=True,
|
78 |
-
inference=True,
|
79 |
-
new_exp=None,
|
80 |
-
)
|
81 |
-
|
82 |
-
# Does all three inferences at the moment.
|
83 |
-
def inference(self, orig_image):
|
84 |
-
image = self._preprocess_image(orig_image)
|
85 |
-
|
86 |
-
# Retrieve numpy events as a dict {event: array[BxHxWxC]}
|
87 |
-
outputs = self.trainer.infer_all(
|
88 |
-
image,
|
89 |
-
numpy=True,
|
90 |
-
bin_value=0.5,
|
91 |
-
)
|
92 |
-
|
93 |
-
return (
|
94 |
-
outputs["flood"].squeeze(),
|
95 |
-
outputs["wildfire"].squeeze(),
|
96 |
-
outputs["smog"].squeeze(),
|
97 |
-
)
|
98 |
-
|
99 |
-
def _preprocess_image(self, img):
|
100 |
-
# rgba to rgb
|
101 |
-
data = img if img.shape[-1] == 3 else uint8(rgba2rgb(img) * 255)
|
102 |
-
|
103 |
-
# to args.target_size
|
104 |
-
data = resize_and_crop(data, self.target_size)
|
105 |
-
|
106 |
-
# resize() produces [0, 1] images, rescale to [-1, 1]
|
107 |
-
data = to_m1_p1(data)
|
108 |
-
return data
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|