Spaces:
Running
Running
Upload inference.py
Browse files- inference.py +330 -0
inference.py
ADDED
@@ -0,0 +1,330 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import cv2
|
3 |
+
import os
|
4 |
+
import numpy as np
|
5 |
+
import shutil
|
6 |
+
from models.anime_gan import GeneratorV1
|
7 |
+
from models.anime_gan_v2 import GeneratorV2
|
8 |
+
from models.anime_gan_v3 import GeneratorV3
|
9 |
+
from utils.common import load_checkpoint, RELEASED_WEIGHTS
|
10 |
+
from utils.image_processing import resize_image, normalize_input, denormalize_input
|
11 |
+
from utils import read_image, is_image_file
|
12 |
+
from tqdm import tqdm
|
13 |
+
# from torch.cuda.amp import autocast
|
14 |
+
|
15 |
+
try:
|
16 |
+
import matplotlib.pyplot as plt
|
17 |
+
except ImportError:
|
18 |
+
plt = None
|
19 |
+
|
20 |
+
try:
|
21 |
+
import moviepy.video.io.ffmpeg_writer as ffmpeg_writer
|
22 |
+
from moviepy.video.io.VideoFileClip import VideoFileClip
|
23 |
+
except ImportError:
|
24 |
+
ffmpeg_writer = None
|
25 |
+
VideoFileClip = None
|
26 |
+
|
27 |
+
|
28 |
+
VALID_FORMATS = {
|
29 |
+
'jpeg', 'jpg', 'jpe',
|
30 |
+
'png', 'bmp',
|
31 |
+
}
|
32 |
+
|
33 |
+
def auto_load_weight(weight, version=None, map_location=None):
|
34 |
+
"""Auto load Generator version from weight."""
|
35 |
+
weight_name = os.path.basename(weight).lower()
|
36 |
+
if version is not None:
|
37 |
+
version = version.lower()
|
38 |
+
assert version in {"v1", "v2", "v3"}, f"Version {version} does not exist"
|
39 |
+
# If version is provided, use it.
|
40 |
+
cls = {
|
41 |
+
"v1": GeneratorV1,
|
42 |
+
"v2": GeneratorV2,
|
43 |
+
"v3": GeneratorV3
|
44 |
+
}[version]
|
45 |
+
else:
|
46 |
+
# Try to get class by name of weight file
|
47 |
+
# For convenenice, weight should start with classname
|
48 |
+
# e.g: Generatorv2_{anything}.pt
|
49 |
+
if weight_name in RELEASED_WEIGHTS:
|
50 |
+
version = RELEASED_WEIGHTS[weight_name][0]
|
51 |
+
return auto_load_weight(weight, version=version, map_location=map_location)
|
52 |
+
|
53 |
+
elif weight_name.startswith("generatorv2"):
|
54 |
+
cls = GeneratorV2
|
55 |
+
elif weight_name.startswith("generatorv3"):
|
56 |
+
cls = GeneratorV3
|
57 |
+
elif weight_name.startswith("generator"):
|
58 |
+
cls = GeneratorV1
|
59 |
+
else:
|
60 |
+
raise ValueError((f"Can not get Model from {weight_name}, "
|
61 |
+
"you might need to explicitly specify version"))
|
62 |
+
model = cls()
|
63 |
+
load_checkpoint(model, weight, strip_optimizer=True, map_location=map_location)
|
64 |
+
model.eval()
|
65 |
+
return model
|
66 |
+
|
67 |
+
|
68 |
+
class Predictor:
|
69 |
+
def __init__(self, weight='hayao', device='cpu', amp=True):
|
70 |
+
# if not torch.cuda.is_available():
|
71 |
+
# device = 'cpu'
|
72 |
+
# # Amp not working on cpu
|
73 |
+
# amp = False
|
74 |
+
|
75 |
+
self.amp = False # Automatic Mixed Precision
|
76 |
+
#self.device_type = 'cuda' if device.startswith('cuda') else 'cpu'
|
77 |
+
self.device_type = 'cpu'
|
78 |
+
self.device = torch.device(device)
|
79 |
+
self.G = auto_load_weight(weight, map_location=device)
|
80 |
+
self.G.to(self.device)
|
81 |
+
|
82 |
+
def transform_and_show(
|
83 |
+
self,
|
84 |
+
image_path,
|
85 |
+
figsize=(18, 10),
|
86 |
+
save_path=None
|
87 |
+
):
|
88 |
+
image = resize_image(read_image(image_path))
|
89 |
+
anime_img = self.transform(image)
|
90 |
+
anime_img = anime_img.astype('uint8')
|
91 |
+
|
92 |
+
fig = plt.figure(figsize=figsize)
|
93 |
+
fig.add_subplot(1, 2, 1)
|
94 |
+
# plt.title("Input")
|
95 |
+
plt.imshow(image)
|
96 |
+
plt.axis('off')
|
97 |
+
fig.add_subplot(1, 2, 2)
|
98 |
+
# plt.title("Anime style")
|
99 |
+
plt.imshow(anime_img[0])
|
100 |
+
plt.axis('off')
|
101 |
+
plt.tight_layout()
|
102 |
+
plt.show()
|
103 |
+
if save_path is not None:
|
104 |
+
plt.savefig(save_path)
|
105 |
+
|
106 |
+
def transform(self, image, denorm=True):
|
107 |
+
'''
|
108 |
+
Transform a image to animation
|
109 |
+
|
110 |
+
@Arguments:
|
111 |
+
- image: np.array, shape = (Batch, width, height, channels)
|
112 |
+
|
113 |
+
@Returns:
|
114 |
+
- anime version of image: np.array
|
115 |
+
'''
|
116 |
+
with torch.no_grad():
|
117 |
+
image = self.preprocess_images(image)
|
118 |
+
# image = image.to(self.device)
|
119 |
+
# with autocast(self.device_type, enabled=self.amp):
|
120 |
+
# print(image.dtype, self.G)
|
121 |
+
fake = self.G(image)
|
122 |
+
fake = fake.detach().cpu().numpy()
|
123 |
+
# Channel last
|
124 |
+
fake = fake.transpose(0, 2, 3, 1)
|
125 |
+
|
126 |
+
if denorm:
|
127 |
+
fake = denormalize_input(fake, dtype=np.uint8)
|
128 |
+
return fake
|
129 |
+
|
130 |
+
def transform_image(self,image):
|
131 |
+
# if not is_image_file(save_path):
|
132 |
+
# raise ValueError(f"{save_path} is not valid")
|
133 |
+
|
134 |
+
# image = read_image(file_path)
|
135 |
+
#
|
136 |
+
# if image is None:
|
137 |
+
# raise ValueError(f"Could not get image from {file_path}")
|
138 |
+
|
139 |
+
anime_img = self.transform(resize_image(image))[0]
|
140 |
+
return anime_img
|
141 |
+
# cv2.imwrite(save_path, anime_img[..., ::-1])
|
142 |
+
# print(f"Anime image saved to {save_path}")
|
143 |
+
|
144 |
+
def transform_in_dir(self, img_dir, dest_dir, max_images=0, img_size=(512, 512)):
|
145 |
+
'''
|
146 |
+
Read all images from img_dir, transform and write the result
|
147 |
+
to dest_dir
|
148 |
+
|
149 |
+
'''
|
150 |
+
os.makedirs(dest_dir, exist_ok=True)
|
151 |
+
|
152 |
+
files = os.listdir(img_dir)
|
153 |
+
files = [f for f in files if self.is_valid_file(f)]
|
154 |
+
print(f'Found {len(files)} images in {img_dir}')
|
155 |
+
|
156 |
+
if max_images:
|
157 |
+
files = files[:max_images]
|
158 |
+
|
159 |
+
for fname in tqdm(files):
|
160 |
+
image = cv2.imread(os.path.join(img_dir, fname))[:,:,::-1]
|
161 |
+
image = resize_image(image)
|
162 |
+
anime_img = self.transform(image)[0]
|
163 |
+
ext = fname.split('.')[-1]
|
164 |
+
fname = fname.replace(f'.{ext}', '')
|
165 |
+
cv2.imwrite(os.path.join(dest_dir, f'{fname}.jpg'), anime_img[..., ::-1])
|
166 |
+
|
167 |
+
def transform_video_yuan(self, input_path, output_path, batch_size=4, start=0, end=0):
|
168 |
+
|
169 |
+
'''
|
170 |
+
Transform a video to animation version
|
171 |
+
https://github.com/lengstrom/fast-style-transfer/blob/master/evaluate.py#L21
|
172 |
+
'''
|
173 |
+
# Force to None
|
174 |
+
end = end or None
|
175 |
+
|
176 |
+
if not os.path.isfile(input_path):
|
177 |
+
raise FileNotFoundError(f'{input_path} does not exist')
|
178 |
+
|
179 |
+
output_dir = "/".join(output_path.split("/")[:-1])
|
180 |
+
os.makedirs(output_dir, exist_ok=True)
|
181 |
+
is_gg_drive = '/drive/' in output_path
|
182 |
+
temp_file = ''
|
183 |
+
#output_file = open(output_path, 'wb')
|
184 |
+
|
185 |
+
if is_gg_drive:
|
186 |
+
# Writing directly into google drive can be inefficient
|
187 |
+
temp_file = f'tmp_anime.{output_path.split(".")[-1]}'
|
188 |
+
|
189 |
+
def transform_and_write(frames, count, writer):
|
190 |
+
anime_images = self.transform(frames)
|
191 |
+
for i in range(0, count):
|
192 |
+
img = np.clip(anime_images[i], 0, 255)
|
193 |
+
writer.write_frame(img)
|
194 |
+
|
195 |
+
video_clip = VideoFileClip(input_path, audio=False)
|
196 |
+
if start or end:
|
197 |
+
video_clip = video_clip.subclip(start, end)
|
198 |
+
|
199 |
+
video_writer = ffmpeg_writer.FFMPEG_VideoWriter(
|
200 |
+
output_path,
|
201 |
+
video_clip.size, video_clip.fps, codec="libx264",
|
202 |
+
# preset="medium", bitrate="2000k",
|
203 |
+
audiofile=input_path, threads=None,
|
204 |
+
ffmpeg_params=None)
|
205 |
+
|
206 |
+
total_frames = round(video_clip.fps * video_clip.duration)
|
207 |
+
print(f'Transfroming video {input_path}, {total_frames} frames, size: {video_clip.size}')
|
208 |
+
|
209 |
+
batch_shape = (batch_size, video_clip.size[1], video_clip.size[0], 3)
|
210 |
+
frame_count = 0
|
211 |
+
frames = np.zeros(batch_shape, dtype=np.float32)
|
212 |
+
for frame in tqdm(video_clip.iter_frames()):
|
213 |
+
try:
|
214 |
+
frames[frame_count] = frame
|
215 |
+
frame_count += 1
|
216 |
+
if frame_count == batch_size:
|
217 |
+
transform_and_write(frames, frame_count, video_writer)
|
218 |
+
frame_count = 0
|
219 |
+
except Exception as e:
|
220 |
+
print(e)
|
221 |
+
break
|
222 |
+
|
223 |
+
|
224 |
+
# The last frames
|
225 |
+
if frame_count != 0:
|
226 |
+
transform_and_write(frames, frame_count, video_writer)
|
227 |
+
|
228 |
+
if temp_file:
|
229 |
+
# move to output path
|
230 |
+
shutil.move(temp_file, output_path)
|
231 |
+
|
232 |
+
print(f'Animation video saved to {output_path}')
|
233 |
+
video_writer.close()
|
234 |
+
#output_file.close()
|
235 |
+
|
236 |
+
def transform_video(self, input_path, output_path, batch_size=4, start=0, end=0):
|
237 |
+
end = end or None
|
238 |
+
|
239 |
+
if not os.path.isfile(input_path):
|
240 |
+
raise FileNotFoundError(f'{input_path} does not exist')
|
241 |
+
|
242 |
+
output_dir = "/".join(output_path.split("/")[:-1])
|
243 |
+
os.makedirs(output_dir, exist_ok=True)
|
244 |
+
is_gg_drive = '/drive/' in output_path
|
245 |
+
temp_file = ''
|
246 |
+
|
247 |
+
if is_gg_drive:
|
248 |
+
temp_file = f'tmp_anime.{output_path.split(".")[-1]}'
|
249 |
+
|
250 |
+
def transform_and_write(frames, count, writer):
|
251 |
+
anime_images = self.transform(frames)
|
252 |
+
for i in range(count):
|
253 |
+
img = np.clip(anime_images[i], 0, 255).astype(np.uint8)
|
254 |
+
writer.write(img)
|
255 |
+
|
256 |
+
video_capture = cv2.VideoCapture(input_path)
|
257 |
+
frame_width = int(video_capture.get(cv2.CAP_PROP_FRAME_WIDTH))
|
258 |
+
frame_height = int(video_capture.get(cv2.CAP_PROP_FRAME_HEIGHT))
|
259 |
+
fps = int(video_capture.get(cv2.CAP_PROP_FPS))
|
260 |
+
frame_count = int(video_capture.get(cv2.CAP_PROP_FRAME_COUNT))
|
261 |
+
|
262 |
+
if start or end:
|
263 |
+
start_frame = int(start * fps)
|
264 |
+
end_frame = int(end * fps) if end else frame_count
|
265 |
+
video_capture.set(cv2.CAP_PROP_POS_FRAMES, start_frame)
|
266 |
+
frame_count = end_frame - start_frame
|
267 |
+
|
268 |
+
video_writer = cv2.VideoWriter(
|
269 |
+
output_path, cv2.VideoWriter_fourcc(*'mp4v'), fps, (frame_width, frame_height))
|
270 |
+
|
271 |
+
print(f'Transforming video {input_path}, {frame_count} frames, size: ({frame_width}, {frame_height})')
|
272 |
+
|
273 |
+
batch_shape = (batch_size, frame_height, frame_width, 3)
|
274 |
+
frames = np.zeros(batch_shape, dtype=np.uint8)
|
275 |
+
frame_idx = 0
|
276 |
+
|
277 |
+
try:
|
278 |
+
for _ in tqdm(range(frame_count)):
|
279 |
+
ret, frame = video_capture.read()
|
280 |
+
if not ret:
|
281 |
+
break
|
282 |
+
frames[frame_idx] = frame
|
283 |
+
frame_idx += 1
|
284 |
+
if frame_idx == batch_size:
|
285 |
+
transform_and_write(frames, frame_idx, video_writer)
|
286 |
+
frame_idx = 0
|
287 |
+
except Exception as e:
|
288 |
+
print(e)
|
289 |
+
finally:
|
290 |
+
video_capture.release()
|
291 |
+
video_writer.release()
|
292 |
+
|
293 |
+
if temp_file:
|
294 |
+
shutil.move(temp_file, output_path)
|
295 |
+
|
296 |
+
print(f'Animation video saved to {output_path}')
|
297 |
+
def preprocess_images(self, images):
|
298 |
+
'''
|
299 |
+
Preprocess image for inference
|
300 |
+
|
301 |
+
@Arguments:
|
302 |
+
- images: np.ndarray
|
303 |
+
|
304 |
+
@Returns
|
305 |
+
- images: torch.tensor
|
306 |
+
'''
|
307 |
+
images = images.astype(np.float32)
|
308 |
+
|
309 |
+
# Normalize to [-1, 1]
|
310 |
+
images = normalize_input(images)
|
311 |
+
images = torch.from_numpy(images)
|
312 |
+
|
313 |
+
images = images.to(self.device)
|
314 |
+
|
315 |
+
# Add batch dim
|
316 |
+
if len(images.shape) == 3:
|
317 |
+
images = images.unsqueeze(0)
|
318 |
+
|
319 |
+
# channel first
|
320 |
+
images = images.permute(0, 3, 1, 2)
|
321 |
+
|
322 |
+
return images
|
323 |
+
|
324 |
+
|
325 |
+
@staticmethod
|
326 |
+
def is_valid_file(fname):
|
327 |
+
ext = fname.split('.')[-1]
|
328 |
+
return ext in VALID_FORMATS
|
329 |
+
|
330 |
+
|