File size: 36,140 Bytes
2cafca2
 
973a4da
 
 
 
 
2cafca2
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
973a4da
 
 
 
 
 
 
2cafca2
 
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
 
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
 
 
 
973a4da
2cafca2
 
 
 
 
 
 
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
 
a5df893
2cafca2
 
 
 
6a0e8ed
973a4da
2cafca2
 
 
 
 
6a0e8ed
 
 
 
 
 
 
973a4da
 
2cafca2
 
 
 
 
 
 
 
 
 
 
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
 
 
 
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
973a4da
2cafca2
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2cafca2
973a4da
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a5df893
731be70
2cafca2
 
 
 
 
a5df893
2cafca2
 
973a4da
 
 
 
2cafca2
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
973a4da
2cafca2
 
 
 
 
731be70
2cafca2
 
 
 
 
 
 
 
 
 
 
731be70
2cafca2
 
 
 
973a4da
2cafca2
 
973a4da
2cafca2
 
 
 
 
 
 
 
731be70
2cafca2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
973a4da
 
a5df893
973a4da
 
 
a5df893
973a4da
 
 
 
 
 
9a76dcd
6a0e8ed
 
 
 
 
 
 
973a4da
 
2cafca2
 
 
 
 
 
973a4da
2cafca2
 
 
a5df893
 
2cafca2
 
 
a5df893
2cafca2
 
 
 
973a4da
2cafca2
 
 
 
 
 
 
 
 
 
 
 
973a4da
2cafca2
973a4da
 
 
 
 
 
2cafca2
731be70
 
 
2cafca2
731be70
973a4da
 
731be70
973a4da
 
731be70
 
 
2cafca2
731be70
 
973a4da
 
 
 
 
 
 
2cafca2
731be70
 
 
2cafca2
731be70
973a4da
 
731be70
973a4da
 
731be70
 
 
2cafca2
731be70
973a4da
 
 
2cafca2
731be70
973a4da
 
731be70
 
2cafca2
731be70
 
2cafca2
973a4da
 
 
 
731be70
973a4da
2cafca2
973a4da
 
 
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
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
#!/usr/bin/env python

import numpy as np
import pandas as pd

from sklearn.svm import SVC
from sklearn.decomposition import PCA
from sklearn.linear_model import LogisticRegression, LinearRegression
from sklearn.model_selection import train_test_split

from tqdm import tqdm
import random
from os.path import join
import os
import pickle

import torch

import matplotlib.pyplot as plt
import PIL
from PIL import Image, ImageColor

import sys
sys.path.append('backend')
from color_annotations import extract_color
from networks_stylegan3 import *
sys.path.append('.')

import dnnlib 
import legacy


def hex2rgb(hex_value):
    h = hex_value.strip("#") 
    rgb = tuple(int(h[i:i+2], 16) for i in (0, 2, 4))
    return rgb

def rgb2hsv(r, g, b):
    # Normalize R, G, B values
    r, g, b = r / 255.0, g / 255.0, b / 255.0
    
    # h, s, v = hue, saturation, value
    max_rgb = max(r, g, b)    
    min_rgb = min(r, g, b)   
    difference = max_rgb-min_rgb 
    
    # if max_rgb and max_rgb are equal then h = 0
    if max_rgb == min_rgb:
        h = 0
    
    # if max_rgb==r then h is computed as follows
    elif max_rgb == r:
        h = (60 * ((g - b) / difference) + 360) % 360
    
    # if max_rgb==g then compute h as follows
    elif max_rgb == g:
        h = (60 * ((b - r) / difference) + 120) % 360
    
    # if max_rgb=b then compute h
    elif max_rgb == b:
        h = (60 * ((r - g) / difference) + 240) % 360
    
    # if max_rgb==zero then s=0
    if max_rgb == 0:
        s = 0
    else:
        s = (difference / max_rgb) * 100
    
    # compute v
    v = max_rgb * 100
    # return rounded values of H, S and V
    return tuple(map(round, (h, s, v)))
 
class DisentanglementBase:
    def __init__(self, repo_folder, model, annotations, df, space, colors_list, compute_s=False, variable='H1', categorical=True):
        self.device = 'cuda' if torch.cuda.is_available() else 'cpu'
        print('Using device', self.device)
        self.repo_folder = repo_folder
        self.model = model.to(self.device)
        self.annotations = annotations
        self.df = df
        self.space = space
        self.categorical = categorical
        self.variable = variable
        
        self.layers = ['input', 'L0_36_512', 'L1_36_512', 'L2_36_512', 'L3_52_512',
                       'L4_52_512', 'L5_84_512', 'L6_84_512', 'L7_148_512', 'L8_148_512', 
                       'L9_148_362', 'L10_276_256', 'L11_276_181', 'L12_276_128', 
                       'L13_256_128', 'L14_256_3']
        self.layers_shapes = [4, 512, 512, 512, 512, 512, 512, 512, 512, 512, 512, 362, 256, 181, 128, 128]
        self.decoding_layers = 16
        self.colors_list = colors_list
        
        self.to_hsv()
        if compute_s:
            self.get_s_space()
        
    def to_hsv(self):
        """
        The tohsv function takes the top 3 colors of each image and converts them to HSV values.
        It then adds these values as new columns in the dataframe.
        
        :param self: Allow the function to access the dataframe
        :return: The dataframe with the new columns added
        :doc-author: Trelent
        """
        print('Adding HSV encoding')
        self.df['H1'] = self.df['top1col'].map(lambda x: rgb2hsv(*hex2rgb(x))[0])
        self.df['H2'] = self.df['top2col'].map(lambda x: rgb2hsv(*hex2rgb(x))[0])
        self.df['H3'] = self.df['top3col'].map(lambda x: rgb2hsv(*hex2rgb(x))[0])
        
        self.df['S1'] = self.df['top1col'].map(lambda x: rgb2hsv(*hex2rgb(x))[1])
        self.df['S2'] = self.df['top2col'].map(lambda x: rgb2hsv(*hex2rgb(x))[1])
        self.df['S3'] = self.df['top3col'].map(lambda x: rgb2hsv(*hex2rgb(x))[1])
        
        self.df['V1'] = self.df['top1col'].map(lambda x: rgb2hsv(*hex2rgb(x))[2])
        self.df['V2'] = self.df['top2col'].map(lambda x: rgb2hsv(*hex2rgb(x))[2])
        self.df['V3'] = self.df['top3col'].map(lambda x: rgb2hsv(*hex2rgb(x))[2])
        
        print('Adding RGB encoding')
        self.df['R1'] = self.df['top1col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[0])
        self.df['R2'] = self.df['top2col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[0])
        self.df['R3'] = self.df['top3col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[0])
        
        self.df['G1'] = self.df['top1col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[1])
        self.df['G2'] = self.df['top2col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[1])
        self.df['G3'] = self.df['top3col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[1])
        
        self.df['B1'] = self.df['top1col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[2])
        self.df['B2'] = self.df['top2col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[2])
        self.df['B3'] = self.df['top3col'].map(lambda x: ImageColor.getcolor(x, 'RGB')[2])
    
    def get_s_space(self):
        """
        The get_s_space function takes the w_vectors from the annotations dictionary and uses them to generate s_vectors.
        The s_space is a space of vectors that are generated by passing each w vector through each layer of the model.
        This allows us to see how much information about a particular class is contained in different layers.
        
        :param self: Bind the method to a class
        :return: A list of lists of s vectors
        :doc-author: Trelent
        """
        print('Getting S space from W')
        ss = []
        for w in tqdm(self.annotations['w_vectors']):
            w_torch = torch.from_numpy(w).to(self.device)
            W = w_torch.expand((16, -1)).unsqueeze(0)
            s = []
            for i,layer in enumerate(self.layers):
                s.append(getattr(self.model.synthesis, layer).affine(W[0, i].unsqueeze(0)).cpu().numpy())

            ss.append(s)
        self.annotations['s_vectors'] = ss
        annotations_file = join(self.repo_folder, 'data/textile_annotated_files/seeds0000-100000_S.pkl')
        print('Storing s for future use here:', annotations_file)
        with open(annotations_file, 'wb') as f:
            pickle.dump(self.annotations, f)

    def get_encoded_latent(self):
        # ... (existing code for getX)
        if self.space.lower() == 'w':
            X = np.array(self.annotations['w_vectors']).reshape((len(self.annotations['w_vectors']), 512))
        elif self.space.lower() == 'z':
            X = np.array(self.annotations['z_vectors']).reshape((len(self.annotations['z_vectors']), 512))
        elif self.space.lower() == 's':
            concat_v = []
            for i in range(len(self.annotations['w_vectors'])):
                concat_v.append(np.concatenate(self.annotations['s_vectors'][i], axis=1))
            X = np.array(concat_v)
            X = X[:, 0, :]
        else:
            Exception("Sorry, option not available, select among Z, W, S")
            
        print('Shape embedding:', X.shape)
        return X
    
    def get_train_val(self, extremes=False):
        y = np.array(self.df[self.variable].values)
        X = self.get_encoded_latent()[:y.shape[0], :]
        if self.categorical:
            bins = [(x-1) * 360 / (len(self.colors_list) - 1)  if x != 1 
                    else 1 for x in range(len(self.colors_list) + 1)]
            bins[0] = 0
            
            y_cat = pd.cut(y, 
                            bins=bins,
                            labels=self.colors_list,
                            include_lowest=True
                            )
            print(y_cat.value_counts())
            
            y_h_cat[y_s == 0] = 'Gray'
            y_h_cat[y_s == 100] = 'Gray'
            y_h_cat[y_v == 0] = 'Gray'
            y_h_cat[y_v == 100] = 'Gray'
            
            print(y_cat.value_counts())
            x_train, x_val, y_train, y_val = train_test_split(X, y_cat, test_size=0.2)
        else:
            if extremes:
                # Calculate the number of elements to consider (10% of array size)
                num_elements = int(0.2 * len(y))
                # Get indices of the top num_elements maximum values
                top_indices = np.argpartition(array, -num_elements)[-num_elements:]
                bottom_indices = np.argpartition(array, -num_elements)[:num_elements]
                y_ext = y[top_indices + bottom_indices, :]
                X_ext = X[top_indices + bottom_indices, :]
                x_train, x_val, y_train, y_val = train_test_split(X_ext, y_ext, test_size=0.2)
            else:
                x_train, x_val, y_train, y_val = train_test_split(X, y, test_size=0.2)
        return x_train, x_val, y_train, y_val
    
    def InterFaceGAN_separation_vector(self, method='LR', C=0.1):
        """
        Method from InterfaceGAN
        The get_separation_space function takes in a type_bin, annotations, and df.
        It then samples 100 of the most representative abstracts for that type_bin and 100 of the least representative abstracts for that type_bin.
        It then trains an SVM or logistic regression model on these 200 samples to find a separation space between them. 
        The function returns this separation space as well as how many nodes are important in this separation space.
        
        :param type_bin: Select the type of abstracts to be used for training
        :param annotations: Access the z_vectors
        :param df: Get the abstracts that are used for training
        :param samples: Determine how many samples to take from the top and bottom of the distribution
        :param method: Specify the classifier to use
        :param C: Control the regularization strength
        :return: The weights of the linear classifier
        :doc-author: Trelent
        """
        x_train, x_val, y_train, y_val = self.get_train_val()
        
        if self.categorical:
            if method == 'SVM':
                svc = SVC(gamma='auto', kernel='linear', random_state=0, C=C)
                svc.fit(x_train, y_train)
                print('Val performance SVM', np.round(svc.score(x_val, y_val), 2))
                return svc.coef_ / np.linalg.norm(svc.coef_)
            elif method == 'LR':
                clf = LogisticRegression(random_state=0, C=C)
                clf.fit(x_train, y_train)
                print('Val performance logistic regression', np.round(clf.score(x_val, y_val), 2))
                return clf.coef_ / np.linalg.norm(clf.coef_)
        else:
            clf = LinearRegression()
            clf.fit(x_train, y_train)
            print('Val performance linear regression', np.round(clf.score(x_val, y_val), 2))
            return clf.coef_ / np.linalg.norm(clf.coef_)
            
    def get_original_position_latent(self, positive_idxs, negative_idxs):
        # ... (existing code for get_original_pos)
        separation_vectors = []
        for i in range(len(self.colors_list)):
            if self.space.lower() == 's':
                current_idx = 0
                vectors = []
                for j, (leng, layer) in enumerate(zip(self.layers_shapes, self.layers)):
                    arr = np.zeros(leng)
                    for positive_idx in positive_idxs[i]:
                        if positive_idx >= current_idx and positive_idx < current_idx + leng:
                            arr[positive_idx - current_idx] = 1
                    for negative_idx in negative_idxs[i]:
                        if negative_idx >= current_idx and negative_idx < current_idx + leng:
                            arr[negative_idx - current_idx] = 1
                        arr = arr / (np.linalg.norm(arr) + 0.000001)
                    vectors.append(arr)
                    current_idx += leng
            elif self.space.lower() == 'z' or self.space.lower() == 'w':
                vectors = np.zeros(512)
                vectors[positive_idxs[i]] = 1
                vectors[negative_idxs[i]] = -1
                vectors = vectors / (np.linalg.norm(vectors) + 0.000001)
            else:
                raise Exception("""This space is not allowed in this function, 
                                    select among Z, W, S""")
            separation_vectors.append(vectors)
            
        return separation_vectors    
    
    def StyleSpace_separation_vector(self, sign=True, num_factors=20, cutout=0.25):
        """ Formula from StyleSpace Analysis """
        x_train, x_val, y_train, y_val = self.get_train_val()
        
        positive_idxs = []
        negative_idxs = []
        for color in self.colors_list:
            x_col = x_train[np.where(y_train == color)]
            mp = np.mean(x_train, axis=0)
            sp = np.std(x_train, axis=0)
            de = (x_col - mp) / sp
            meu = np.mean(de, axis=0)
            seu = np.std(de, axis=0)
            if sign:
                thetau = meu / seu
                positive_idx = np.argsort(thetau)[-num_factors//2:]
                negative_idx = np.argsort(thetau)[:num_factors//2]
                
            else:
                thetau = np.abs(meu) / seu
                positive_idx = np.argsort(thetau)[-num_factors:]
                negative_idx = []
                

            if cutout:
                beyond_cutout = np.where(np.abs(thetau) > cutout)
                positive_idx = np.intersect1d(positive_idx, beyond_cutout)
                negative_idx = np.intersect1d(negative_idx, beyond_cutout)
                
                if len(positive_idx) == 0 and len(negative_idx) == 0:
                    print('No values found above the current cutout', cutout, 'for color', color, '.\n Disentangled vector will be all zeros.' )
                
            positive_idxs.append(positive_idx)
            negative_idxs.append(negative_idx)
        
        separation_vectors = self.get_original_position_latent(positive_idxs, negative_idxs)
        return separation_vectors

    def GANSpace_separation_vectors(self, num_components):
        x_train, x_val, y_train, y_val = self.get_train_val()
        if self.space.lower() == 'w':
            pca = PCA(n_components=num_components)

            dims_pca = pca.fit_transform(x_train.T)
            dims_pca /= np.linalg.norm(dims_pca, axis=0)
            
            return dims_pca
        
        else:
            raise("""This space is not allowed in this function, 
                     only W""")
    
    def generate_images(self, seed, separation_vector=None, lambd=0):
        """
        The generate_original_image function takes in a latent vector and the model,
        and returns an image generated from that latent vector.
        
        
        :param z: Generate the image
        :param model: Generate the image
        :return: A pil image
        :doc-author: Trelent
        """
        G = self.model.to(self.device) # type: ignore
        # Labels.
        label = torch.zeros([1, G.c_dim], device=self.device)
        if self.space.lower() == 'z':
            vec = self.annotations['z_vectors'][seed]
            Z = torch.from_numpy(vec.copy()).to(self.device)
            if separation_vector is not None:
                change = torch.from_numpy(separation_vector.copy()).unsqueeze(0).to(self.device)
                Z = torch.add(Z, change, alpha=lambd)
            img = G(Z, label, truncation_psi=1, noise_mode='const')
        elif self.space.lower() == 'w':
            vec = self.annotations['w_vectors'][seed]
            W = torch.from_numpy(np.repeat(vec, self.decoding_layers, axis=0)
                                 .reshape(1, self.decoding_layers, vec.shape[1]).copy()).to(self.device)
            if separation_vector is not None:
                change = torch.from_numpy(separation_vector.copy()).unsqueeze(0).to(self.device)
                W = torch.add(W, change, alpha=lambd)
            img = G.synthesis(W, noise_mode='const')
        else:
            raise Exception("""This space is not allowed in this function, 
                            select either W or Z or use generate_flexible_images""")
            
        img = (img.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
        return PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB')

    def forward_from_style(self, x, styles, layer):
        dtype = torch.float16 if (getattr(self.model.synthesis, layer).use_fp16 and self.device=='cuda') else torch.float32
        
        if getattr(self.model.synthesis, layer).is_torgb:
            weight_gain = 1 / np.sqrt(getattr(self.model.synthesis, layer).in_channels * (getattr(self.model.synthesis, layer).conv_kernel ** 2))
            styles = styles * weight_gain
        
        input_gain = getattr(self.model.synthesis, layer).magnitude_ema.rsqrt().to(dtype)
        
        # Execute modulated conv2d.
        x = modulated_conv2d(x=x.to(dtype), w=getattr(self.model.synthesis, layer).weight.to(dtype), s=styles.to(dtype),
        padding=getattr(self.model.synthesis, layer).conv_kernel-1, 
                        demodulate=(not getattr(self.model.synthesis, layer).is_torgb), 
                        input_gain=input_gain.to(dtype))
        
        # Execute bias, filtered leaky ReLU, and clamping.
        gain = 1 if getattr(self.model.synthesis, layer).is_torgb else np.sqrt(2)
        slope = 1 if getattr(self.model.synthesis, layer).is_torgb else 0.2
        
        x = filtered_lrelu.filtered_lrelu(x=x, fu=getattr(self.model.synthesis, layer).up_filter, fd=getattr(self.model.synthesis, layer).down_filter, 
                                            b=getattr(self.model.synthesis, layer).bias.to(x.dtype),
                                            up=getattr(self.model.synthesis, layer).up_factor, down=getattr(self.model.synthesis, layer).down_factor, 
                                            padding=getattr(self.model.synthesis, layer).padding,
                                            gain=gain, slope=slope, clamp=getattr(self.model.synthesis, layer).conv_clamp)
        return x
    
    def generate_flexible_images(self, seed, separation_vector=None, lambd=0):
        if self.space.lower() != 's':
            raise Exception("""This space is not allowed in this function, 
                            select S or use generate_images""")
            
        vec = self.annotations['w_vectors'][seed]
        w_torch = torch.from_numpy(vec).to(self.device)
        W = w_torch.expand((self.decoding_layers, -1)).unsqueeze(0)
        x = self.model.synthesis.input(W[0,0].unsqueeze(0))
        for i, layer in enumerate(self.layers[1:]):
            style = getattr(self.model.synthesis, layer).affine(W[0, i].unsqueeze(0))
            if separation_vector is not None:
                change = torch.from_numpy(separation_vector[i+1].copy()).unsqueeze(0).to(self.device)
                style = torch.add(style, change, alpha=lambd)
            x = self.forward_from_style(x, style, layer)
        
        if self.model.synthesis.output_scale != 1:
                x = x * self.model.synthesis.output_scale

        img = (x.permute(0, 2, 3, 1) * 127.5 + 128).clamp(0, 255).to(torch.uint8)
        img = PIL.Image.fromarray(img[0].cpu().numpy(), 'RGB')
            
        return img

    def generate_changes(self, seed, separation_vector, min_epsilon=-3, max_epsilon=3, count=5, savefig=True, feature=None, method=None, save_separately=False):
        """
        The regenerate_images function takes a model, z, and decision_boundary as input.  It then
        constructs an inverse rotation/translation matrix and passes it to the generator.  The generator
        expects this matrix as an inverse to avoid potentially failing numerical operations in the network.
        The function then generates images using G(z_0, label) where z_0 is a linear combination of z and the decision boundary.
        
        :param model: Pass in the model to be used for image generation
        :param z: Generate the starting point of the line
        :param decision_boundary: Generate images along the direction of the decision boundary
        :param min_epsilon: Set the minimum value of lambda
        :param max_epsilon: Set the maximum distance from the original image to generate
        :param count: Determine the number of images that are generated
        :return: A list of images and a list of lambdas
        :doc-author: Trelent
        """
            
        os.makedirs(join(self.repo_folder, 'figures'), exist_ok=True)
        lambdas = np.linspace(min_epsilon, max_epsilon, count)
        images = []
        # Generate images.
        for _, lambd in enumerate(lambdas):
            if self.space.lower() == 's':
                images.append(self.generate_flexible_images(seed, separation_vector=separation_vector, lambd=lambd))
            elif self.space.lower() in ['z', 'w']:
                images.append(self.generate_images(seed, separation_vector=separation_vector, lambd=lambd))
        
        if savefig:
            fig, axs = plt.subplots(1, len(images), figsize=(90,20))
            title = 'Disentanglement method: '+ method + ', on feature: ' + feature + ' on space: ' + self.space + ', image seed: ' + str(seed)
            name = '_'.join([method, feature, self.space, str(seed), str(lambdas[-1])])
            fig.suptitle(title, fontsize=20)
                
            for i, (image, lambd) in enumerate(zip(images, lambdas)):
                axs[i].imshow(image)
                axs[i].set_title(np.round(lambd, 2))
            plt.tight_layout()
            plt.savefig(join(self.repo_folder, 'figures', 'examples_new', name+'.jpg'))
            plt.close()
            
            if save_separately:
                for i, (image, lambd) in enumerate(zip(images, lambdas)):
                    plt.imshow(image)
                    plt.tight_layout()
                    plt.savefig(join(self.repo_folder, 'figures', 'examples_new', name + '_' + str(lambd) + '.jpg'))
                    plt.close()
            
        return images, lambdas
    
    def get_verification_score(self, separation_vector, feature_id, samples=10, lambd=1, savefig=False, feature=None, method=None):
        items = random.sample(range(100000), samples)
        if self.categorical:
            if feature_id == 0:
                hue_low = 0
                hue_high = 1
            elif feature_id == 1:
                hue_low = 1
                hue_high = (feature_id - 1) * 360 / (len(self.colors_list) - 1)  
            else:
                hue_low = (feature_id - 1) * 360 / (len(self.colors_list) - 1) 
                hue_high = feature_id * 360 / (len(self.colors_list) - 1)
        
            matches = 0
        
            for seed in tqdm(items):
                images, lambdas = self.generate_changes(seed, separation_vector, min_epsilon=-lambd, max_epsilon=lambd, count=3, savefig=savefig, feature=feature, method=method)
                try:
                    colors_negative = extract_color(images[0], 5, 1, None)
                    h0, s0, v0 = rgb2hsv(*hex2rgb(colors_negative[0]))
                
                    colors_orig = extract_color(images[1], 5, 1, None)
                    h1, s1, v1 = rgb2hsv(*hex2rgb(colors_orig[0])) 
                    
                    colors_positive = extract_color(images[2], 5, 1, None)
                    h2, s2, v2 = rgb2hsv(*hex2rgb(colors_positive[0]))
                    
                    if h1 > hue_low and h1 < hue_high:
                        samples -= 1
                    else:
                        if (h0 > hue_low and h0 < hue_high) or (h2 > hue_low and h2 < hue_high):
                            matches += 1
                
                except Exception as e:
                    print(e)
            
            return np.round(matches / samples, 2)
            
        else:
            increase = 0
        
            for seed in tqdm(items):
                images, lambdas = self.generate_changes(seed, separation_vector, min_epsilon=-lambd, 
                                                        max_epsilon=lambd, count=3, savefig=savefig, 
                                                        feature=feature, method=method)
                try:
                    colors_negative = extract_color(images[0], 5, 1, None)
                    r0, g0, b0 = hex2rgb(colors_negative[0])
                    h0, s0, v0 = rgb2hsv(*hex2rgb(colors_negative[0]))
                
                    colors_orig = extract_color(images[1], 5, 1, None)
                    r1, g1, b1 = hex2rgb(colors_orig[0])
                    h1, s1, v1 = rgb2hsv(*hex2rgb(colors_orig[0]))
                    
                    colors_positive = extract_color(images[2], 5, 1, None)
                    r2, g2, b2 = hex2rgb(colors_positive[0])
                    h2, s2, v2 = rgb2hsv(*hex2rgb(colors_positive[0]))
                    
                    if 's' in self.variable.lower():
                        increase += max(0, s2 - s1)
                    elif 'v' in self.variable.lower():
                        increase += max(0, v2 - v1)
                    elif 'r' in self.variable.lower():
                        increase += max(0, r2 - r1)
                    elif 'g' in self.variable.lower():
                        increase += max(0, g2 - g1)
                    elif 'b' in self.variable.lower():
                        increase += max(0, b2 - b1)
                    else:
                        raise('Continous variable not allowed, choose between RGB or SV')
                except Exception as e:
                    print(e)
            
            return np.round(increase / samples, 2)
            

def continous_experiment(name, var, repo_folder, model, annotations, df, space, colors_list, kwargs):
    scores = []
    print(f'Launching {name} experiment')
    disentanglemnet_exp = DisentanglementBase(repo_folder, model, annotations, df, space=space, colors_list=colors_list, compute_s=False, variable=var, categorical=False)
    for extr in kwargs['extremes']:                
        separation_vector = disentanglemnet_exp.InterFaceGAN_separation_vector()
        print(f'Generating images with variations for {name}')
        for s in range(30):
            seed = random.randint(0,100000)
            for eps in kwargs['max_lambda']:
                disentanglemnet_exp.generate_changes(seed, separation_vector, min_epsilon=-eps, max_epsilon=eps, savefig=True, feature=name, method= 'InterFaceGAN_' + str(extr))
                                    
        print('Finally obtaining verification score')
        for verif in kwargs['lambda_verif']:    
            score = disentanglemnet_exp.get_verification_score(separation_vector, 0, samples=kwargs['samples'], lambd=verif, savefig=False, feature=name, method='InterFaceGAN_' + str(extr))
            print(f'Score for method InterfaceGAN on {name}:', score)
                                
            scores.append([space, 'InterFaceGAN', name, score, 'extremes method:' + str(extr) + 'verification lambda:' + str(verif), ', '.join(list(separation_vector.astype(str)))])
            
        score_df = pd.DataFrame(scores, columns=['space', 'method', 'color', 'score', 'kwargs', 'vector'])
        print(score_df)
        score_df.to_csv(join(repo_folder, f'data/scores_{name}.csv')) 
            
def main():
    repo_folder = '.'
    annotations_file = join(repo_folder, 'data/textile_annotated_files/seeds0000-1000000.pkl')
    with open(annotations_file, 'rb') as f:
        annotations = pickle.load(f)

    df_file = join(repo_folder, 'data/textile_annotated_files/top_three_colours_00000-730003.csv')
    df = pd.read_csv(df_file).fillna('#000000')

    model_file = join(repo_folder, 'data/textile_model_files/network-snapshot-005000.pkl')
    with dnnlib.util.open_url(model_file) as f:
        model = legacy.load_network_pkl(f)['G_ema'] # type: ignore


    # colors_list = ['Red', 'Orange', 'Yellow', 'Yellow Green', 'Chartreuse Green',
    #                'Kelly Green', 'Green Blue Seafoam', 'Cyan Blue',
    #                'Warm Blue', 'Indigo', 'Purple Magenta', 'Magenta Pink']
    # colors_list = ['Gray', 'Red Orange', 'Yellow', 'Green', 'Light Blue',
    #                'Blue', 'Purple', 'Pink']
    colors_list = ['Gray', 'Red', 'Yellow', 'Green', 'Cyan',
                   'Blue', 'Magenta']
    
    scores = []
    kwargs = {'CL method':['LR', 'SVM'], 'C':[0.1, 1], 'sign':[True, False], 
              'num_factors':[1, 5, 10, 20], 'cutout': [None], 'max_lambda':[18, 6], 
              'samples':30, 'lambda_verif':[14, 7], 'extremes':[True, False]}
    continuous = False
    specific_examples = [53139, 99376, 16, 99585, 40851, 70, 17703, 44, 52628,
                        99884, 52921, 46180, 19995, 40920,  554]
    
    if specific_examples is not None:
        disentanglemnet_exp = DisentanglementBase(repo_folder, model, annotations, df, space='w', colors_list=colors_list, compute_s=False)

        # separation_vectors = disentanglemnet_exp.StyleSpace_separation_vector(sign=True, num_factors=10, cutout=None)
        separation_vectors = disentanglemnet_exp.InterFaceGAN_separation_vector(method='LR', C=0.1)
        for specific_example in specific_examples:
            seed = specific_example
            for i, color in enumerate(colors_list):
                disentanglemnet_exp.generate_changes(seed, separation_vectors[i], min_epsilon=-18, max_epsilon=18, savefig=True, save_separately=True, feature=color, method='InterFaceGAN' + '_' + str('LR') + '_' + str(0.1) + '_' + str(None))
        
        return
       
    for space in ['w', ]: #'z', 's'
        print('Launching experiment with space:', space)
        
        if continuous:
            continous_experiment('Saturation', 'S1', repo_folder, model, annotations, df, space, colors_list, kwargs) 
            continous_experiment('Value', 'V1', repo_folder, model, annotations, df, space, colors_list, kwargs) 
            continous_experiment('Red', 'R1', repo_folder, model, annotations, df, space, colors_list, kwargs) 
            continous_experiment('Green', 'G1', repo_folder, model, annotations, df, space, colors_list, kwargs) 
            continous_experiment('Blue', 'B1', repo_folder, model, annotations, df, space, colors_list, kwargs) 
            break
        
        print('Launching Hue experiment')
        variable = 'H1'
        disentanglemnet_exp = DisentanglementBase(repo_folder, model, annotations, df, space=space, colors_list=colors_list, compute_s=False, variable=variable)

        for method in ['StyleSpace', 'InterFaceGAN',]: #'GANSpace'
            if space != 's' and method == 'InterFaceGAN':
                print('Now obtaining separation vector for using InterfaceGAN')
                for met in kwargs['CL method']:
                    for c in kwargs['C']:
                        separation_vectors = disentanglemnet_exp.InterFaceGAN_separation_vector(method=met, C=c)
                        for i, color in enumerate(colors_list):
                            print(f'Generating images with variations for color {color}')
                            for s in range(30):
                                seed = random.randint(0,100000)
                                for eps in kwargs['max_lambda']:
                                    disentanglemnet_exp.generate_changes(seed, separation_vectors[i], min_epsilon=-eps, max_epsilon=eps, savefig=True, feature=color, method=str(method) + '_' + str(met) + '_' + str(c) + '_' + str(len(colors_list)) + '_' + str(variable))
                                
                            print('Finally obtaining verification score')
                            for verif in kwargs['lambda_verif']:    
                                score = disentanglemnet_exp.get_verification_score(separation_vectors[i], i, samples=kwargs['samples'], lambd=verif, savefig=False, feature=color, method=method)
                                print('Score for method', method, 'on space', space, 'for color', color, ':', score)
                            
                                scores.append([space, method, color, score, 'classification method:' + met + ', regularization: ' + str(c) + ', verification lambda:' + str(verif), ', '.join(list(separation_vectors[i].astype(str)))])
                        score_df = pd.DataFrame(scores, columns=['space', 'method', 'color', 'score', 'kwargs', 'vector'])
                        print(score_df)
                        score_df.to_csv(join(repo_folder, f'data/scores_InterfaceGAN_{variable}_{len(colors_list)}.csv'))
                    
                        
            elif method == 'StyleSpace':
                print('Now obtaining separation vector for using StyleSpace')
                for sign in kwargs['sign']:
                    for num_factors in kwargs['num_factors']:
                        for cutout in kwargs['cutout']:
                            separation_vectors = disentanglemnet_exp.StyleSpace_separation_vector(sign=sign, num_factors=num_factors, cutout=cutout)
                            for i, color in enumerate(colors_list):
                                print(f'Generating images with variations for color {color}')
                                for s in range(30):
                                    seed = random.randint(0,100000)
                                    for eps in kwargs['max_lambda']:
                                        disentanglemnet_exp.generate_changes(seed, separation_vectors[i], min_epsilon=-eps, max_epsilon=eps, savefig=True, feature=color, method=method + '_' + str(num_factors) + '_' + str(cutout) + '_' + str(sign)  + '_' + str(len(colors_list)) + '_' + str(variable))
                                            
                                print('Finally obtaining verification score')
                                for verif in kwargs['lambda_verif']:    
                                    score = disentanglemnet_exp.get_verification_score(separation_vectors[i], i, samples=kwargs['samples'], lambd=verif, savefig=False, feature=color, method=method)
                                    print('Score for method', method, 'on space', space, 'for color', color, ':', score)
                                        
                                    scores.append([space, method, color, score, 'using sign:' + str(sign) + ', number of factors: ' + str(num_factors) + ', using cutout: ' + str(cutout) + ', verification lambda:' + str(verif), ', '.join(list(separation_vectors[i].astype(str)))])
                            score_df = pd.DataFrame(scores, columns=['space', 'method', 'color', 'score', 'kwargs', 'vector'])
                            print(score_df)
                            score_df.to_csv(join(repo_folder, f'data/scores_StyleSpace_{variable}_{len(colors_list)}.csv'))
                            
            if space == 'w' and method == 'GANSpace':
                print('Now obtaining separation vector for using GANSpace')
                separation_vectors = disentanglemnet_exp.GANSpace_separation_vectors(100)
                print(separation_vectors.shape)
                for s in range(30):
                    print('Generating images with variations')
                    seed = random.randint(0,100000)
                    for i in range(100):
                        for eps in kwargs['max_lambda']:
                            disentanglemnet_exp.generate_changes(seed, separation_vectors.T[i], min_epsilon=-eps, max_epsilon=eps, savefig=True, feature='dimension_' + str(i), method=method)
                                        
                score = None
                scores.append([space, method, 'PCA', score, '100', ', '.join(list(separation_vectors.T[i].astype(str)))])
            else:
                print('Skipping', method, 'on space', space)
                continue
            
    score_df = pd.DataFrame(scores, columns=['space', 'method', 'color', 'score', 'kwargs', 'vector'])
    print(score_df)
    score_df.to_csv(join(repo_folder, 'data/scores_{}.csv'.format(pd.to_datetime.now().strftime("%Y-%m-%d_%H%M%S"))))
 
if __name__ == "__main__":
    main()