forked from Beinabih/Pytorch-HeadTrip
/
dream.py
467 lines (373 loc) · 14.3 KB
/
dream.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
"""Summary
"""
import argparse
import glob
import os
import random
import cv2
import matplotlib.pyplot as plt
import numpy as np
import scipy.ndimage as nd
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import tqdm
from PIL import Image
from torch.autograd import Variable
from torchvision import models
from config import load_config
from depth_inference import MiDaS
from spyNet import calc_opflow
from utils import clip, convert, deprocess, get_octaves, preprocess, warp
class Dreamer:
def __init__(self, img_p, outpath, config):
"""Summary
Args:
model (TYPE): Description
batchsize (TYPE): Description
img_p (TYPE): Description
outpath (TYPE): Description
config (TYPE): Description
"""
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.config = config
self.img_p = img_p
self.outpath = outpath
self.init_model()
self.norm_str = 5
self.octave_list = [1.1, 1.2, 1.3, 1.4, 1.5]
self.num_octaves_para = config["num_octaves"]
self.octave_scale = config["octave_scale"]
self.at_layer_para = config["at_layer"]
self.lr = config["lr"]
self.random = config["random"]
self.no_class = config["no_class"]
self.ch_list = config["channel_list"]
self.img_list = sorted(glob.glob(img_p))
self.depth = config["use_depth"]
self.depth_w = config["depth_str"]
self.depth_model = MiDaS(False)
self.loss = nn.BCEWithLogitsLoss()
self.transform = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),
]
)
if self.random:
self.random_para()
if self.no_class:
self.layers = list(self.model.features.children())
self.model = nn.Sequential(*self.layers[: (self.at_layer_para + 1)])
self.norm_str = 1
print(
self.config["num_iterations"],
self.octave_scale,
self.config["num_octaves"],
self.lr,
)
def init_model(self):
"""initializes the model with the config file"""
if self.config["model"] == "resnet":
network = models.resnext50_32x4d(pretrained=True)
elif self.config["model"] == "vgg19":
network = models.vgg19(pretrained=True)
elif self.config["model"] == "densenet":
network = models.densenet121(pretrained=True)
elif self.config["model"] == "inception":
network = models.inception_v3(pretrained=True)
elif self.config["model"] == "mobile":
network = models.mobilenet_v2(pretrained=True)
elif self.config["model"] == "shuffle":
network = models.shufflenet_v2_x0_5(pretrained=True)
elif self.config["model"] == "squeeze":
network = models.squeezenet1_1(pretrained=True)
elif self.config["model"] == "resnetx":
network = models.resnext101_32x8d(pretrained=True)
elif self.config["model"] == "masnet":
network = models.mnasnet1_0(pretrained=True)
elif self.config["model"] == "googlenet":
network = models.googlenet(pretrained=True)
elif self.config["model"] == "alexnet":
network = models.alexnet(pretrained=True)
else:
print("Invalid Model")
network.eval()
self.model = network.to(self.device)
if self.config["fp16"]:
scaler = torch.cuda.amp.GradScaler()
self.model = self.model.half()
# amp.register_float_function(torch, "batch_norm")
# self.model = amp.initialize(self.model, opt_level="O2")
def forward(self, model, image, z, d_img=None, mask=None):
"""Summary
Args:
model (TYPE): Description
image (TYPE): Description
z (TYPE): Description
d_img (None, optional): Description
mask (None, optional): Description
Returns:
TYPE: Description
"""
model.zero_grad()
if self.config["fp16"]:
with torch.cuda.amp.autocast():
out = model(image)
else:
out = model(image)
if self.config["guided"]:
target = self.get_target(self.config, z, out)
loss = -self.loss(out, target)
else:
loss = out.norm()
loss.backward()
avg_grad = np.abs(image.grad.data.cpu().numpy()).mean()
norm_lr = self.lr / avg_grad
grad = image.grad.data
dream_grad = grad * (norm_lr * self.norm_str)
if self.depth:
d_img = torch.from_numpy(d_img)
d_img = d_img[0, 0].to(self.device)
dream_grad *= d_img * self.depth_w
if mask is not None:
mask = torch.from_numpy(mask)
dream_grad *= mask.to(self.device)
image.data += dream_grad
image.data = clip(image.data)
image.grad.data.zero_()
return image
def get_target(self, config, z, out):
"""Summary
Args:
config (Dictionary): Config File
z (Integer): Iteration Value
out (tensor): Model Output
Returns:
Tensor: Target Tensor for guided dreaming
"""
target = torch.zeros((1, 1000)).to(self.device)
if config["max_output"]:
out = out.float()
if config["pyramid_max"]:
if z == 0:
self.channel = out.argmax()
target[0, self.channel] = 100
else:
out[0, self.channel] = 0
self.channel = out.argmax()
target = torch.zeros((1, 1000)).to(self.device)
target[0, self.channel] = 100
else:
self.channel = out.argmax()
target[0, self.channel] = 100
else:
for ch in config["channel_list"]:
target[0, ch] = 100
return target
def dream(self, image, model, d_img=None, mask=None):
"""Updates the image to maximize outputs for n iterations
Args:
image (TYPE): Description
model (TYPE): Description
d_img (None, optional): Description
mask (None, optional): Description
Returns:
TYPE: Description
"""
Tensor = (
torch.cuda.FloatTensor if torch.cuda.is_available else torch.FloatTensor
)
image = Variable(Tensor(image), requires_grad=True)
for n in range(self.config["num_iterations"]):
image = self.forward(model, image, n, d_img, mask)
return image.cpu().data.numpy()
def deep_dream(self, image, model, i, seq, mask=None):
"""Main deep dream method
Args:
image (TYPE): Description
model (TYPE): Description
i (TYPE): Description
seq (TYPE): Description
mask (None, optional): Description
Returns:
TYPE: Description
"""
image_p = image.unsqueeze(0).cpu().detach().numpy()
args = []
octaves = get_octaves(image_p, self.config["num_octaves"], self.octave_scale)
if self.depth:
d_img = self.depth_model.inference(convert(image_p))
d_img = d_img / np.max(d_img)
if self.config["invert_depth"]:
d_img = 1 - d_img
if self.config["use_threshold"]:
d_img[d_img < self.config["th_val"]] = 0
d_img = np.expand_dims(d_img, 0)
d_img = np.expand_dims(d_img, 0)
d_img_octaves = get_octaves(
d_img, self.config["num_octaves"], self.octave_scale
)
d_img_octaves = d_img_octaves[::-1]
args.append(d_img_octaves)
if mask is not None:
mask = np.transpose(mask, (2, 0, 1))
mask = np.expand_dims(mask, 0)
octaves_mask = get_octaves(
mask, self.config["num_octaves"], self.octave_scale
)
octaves_mask = octaves_mask[::-1]
args.append(octaves_mask)
kernel = np.ones((5, 5), np.uint8)
self.detail = np.zeros_like(octaves[-1])
for octave, octave_base in enumerate(tqdm.tqdm(octaves[::-1], desc="Dreaming")):
if octave > 0:
# Upsample detail to new octave dimension
self.detail = nd.zoom(
self.detail,
np.array(octave_base.shape) / np.array(self.detail.shape),
order=1,
)
input_image = octave_base + self.detail
dreamed_image = self.dream(
input_image, model, *map(lambda x: x[octave], args)
)
self.detail = dreamed_image - octave_base
return input_image
def save_img(self, img, suffix, iter_):
"""Summary
Args:
img (numpy array): Output Image
suffix (string): filename suffix
iter_ (integer): the iteration value
"""
img = deprocess(img)
img = np.clip(img, 0, 1)
file_name = self.img_list[self.config["start_position"] + iter_]
file_name = file_name.split("/")[-1]
plt.imsave(self.outpath + "/{}{}".format(suffix, file_name), img)
def get_opflow_image(self, img1, dream_img, img2):
"""Calculates the optical flow with opencv and the spynet
Args:
img1 (TYPE): Description
dream_img (TYPE): Description
img2 (TYPE): Description
Returns:
TYPE: Description
"""
img1 = np.float32(img1)
dream_img = np.float32(dream_img)
img2 = np.float32(img2)
h, w, c = img1.shape
if self.config["use_spynet"]:
flow = calc_opflow(np.uint8(img1), np.uint8(img2))
flow = np.transpose(np.float32(flow), (1, 2, 0))
else:
grayImg1 = cv2.cvtColor(img1, cv2.COLOR_RGB2GRAY)
grayImg2 = cv2.cvtColor(img2, cv2.COLOR_RGB2GRAY)
flow = cv2.calcOpticalFlowFarneback(
grayImg1,
grayImg2,
pyr_scale=0.5,
levels=3,
winsize=15,
iterations=3,
poly_n=3,
poly_sigma=1.2,
flags=0,
flow=1,
)
inv_flow = flow
flow = -flow
flow[:, :, 0] += np.arange(w)
flow[:, :, 1] += np.arange(h)[:, np.newaxis]
halludiff = cv2.addWeighted(img2, 0.1, dream_img, 0.9, 0) - img1
halludiff = cv2.remap(halludiff, flow, None, cv2.INTER_LINEAR)
hallu_flow = img2 + halludiff
magnitude, angle = cv2.cartToPolar(inv_flow[..., 0], inv_flow[..., 1])
norm_mag = cv2.normalize(magnitude, None, 0, 255, cv2.NORM_MINMAX)
ret, mask = cv2.threshold(norm_mag, 6, 255, cv2.THRESH_BINARY)
flow_mask = mask.astype(np.uint8).reshape((h, w, 1))
blendstatic = 0.1
background_blendimg = cv2.addWeighted(
img2, (1 - blendstatic), dream_img, blendstatic, 0
)
background_masked = cv2.bitwise_and(
background_blendimg, background_blendimg, mask=cv2.bitwise_not(flow_mask)
)
return hallu_flow, background_masked
def random_para(self):
"""chooses random parameters"""
self.config["num_iterations"] = random.randint(2, 14)
if not self.config["guided"]:
self.at_layer_para = random.randint(10, 38)
self.config["num_octaves"] = random.randint(30, 40)
self.lr = random.choice([0.01, 0.009, 0.008, 0.02, 0.03, 0.007])
self.octave_scale = random.choice(self.octave_list)
def dream_single(self):
"""Dreams independent frames"""
for i, path in enumerate(self.img_list):
img1 = Image.open(path)
d_img = self.deep_dream(self.transform(img1), self.model, i, seq="first")
self.save_img(d_img, "", i)
def dream_seq(self):
"""Dreams a sequence with optical flow"""
for i, path in enumerate(self.img_list[self.config["start_position"] :]):
if i == 0:
img1 = Image.open(path)
d_img = self.deep_dream(
self.transform(img1), self.model, i, seq="first"
)
self.save_img(d_img, "", i)
d_img = convert(d_img)
flow_iter = 0
# the iterations needs to be reduced
self.config["num_iterations"] -= 5
if i > 0:
img2 = Image.open(path)
feature_img, background_masked = self.get_opflow_image(
img1, d_img, img2
)
feature_img = np.clip(feature_img, 0, 255)
background_masked[background_masked > 0] = 1 - (flow_iter * 0.1) # 0.5
background_masked[background_masked == 0] = flow_iter * 0.1
d_img = self.deep_dream(
self.transform(np.uint8(feature_img)),
self.model,
i,
seq="first",
mask=background_masked,
)
# change position
img1 = img2
self.save_img(d_img, "", i)
d_img = convert(d_img)
flow_iter += 1
flow_iter = 0 if flow_iter > 5 else flow_iter
def start_dreamer(config):
"""
Args:
config (Dictionary): The config file
"""
pretrained = config["pretrained"]
# Load image
if os.path.isdir(config["input"]):
img_p = config["input"] + "/*"
elif os.path.isfile(config["input"]):
img_p = config["input"]
else:
raise Exception("Wrong Input")
outpath = config["outpath"]
os.makedirs(outpath, exist_ok=True)
dreamer = Dreamer(img_p, outpath, config)
if config["seq"]:
dreamer.dream_seq()
else:
dreamer.dream_single()
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--config", default="", type=str)
opt = parser.parse_args()
config = load_config()
start_dreamer(config)