File size: 19,900 Bytes
15369ca
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import ast
import io
import math
import statistics
import string

import cairosvg
import clip
import cv2
import kagglehub
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from more_itertools import chunked
from PIL import Image, ImageFilter
from transformers import (
    AutoProcessor,
    BitsAndBytesConfig,
    PaliGemmaForConditionalGeneration,
)

svg_constraints = kagglehub.package_import('metric/svg-constraints')


class ParticipantVisibleError(Exception):
    pass


def score(
    solution: pd.DataFrame, submission: pd.DataFrame, row_id_column_name: str, random_seed: int = 0
) -> float:
    """Calculates a fidelity score by comparing generated SVG images to target text descriptions.

    Parameters
    ----------
    solution : pd.DataFrame
        A DataFrame containing target questions, choices, and answers about an SVG image.
    submission : pd.DataFrame
        A DataFrame containing generated SVG strings. Must have a column named 'svg'.
    row_id_column_name : str
        The name of the column containing row identifiers. This column is removed before scoring.
    random_seed : int
        A seed to set the random state.

    Returns
    -------
    float
        The mean fidelity score (a value between 0 and 1) representing the average similarity between the generated SVGs and their descriptions.
        A higher score indicates better fidelity.

    Raises
    ------
    ParticipantVisibleError
        If the 'svg' column in the submission DataFrame is not of string type or if validation of the SVG fails.

    Examples
    --------
    >>> import pandas as pd
    >>> solution = pd.DataFrame({
    ...     'id': ["abcde"],
    ...     'question': ['["Is there a red circle?", "What shape is present?"]'],
    ...     'choices': ['[["yes", "no"], ["square", "circle", "triangle", "hexagon"]]'],
    ...     'answer': ['["yes", "circle"]'],
    ... })
    >>> submission = pd.DataFrame({
    ...     'id': ["abcde"],
    ...     'svg': ['<svg viewBox="0 0 100 100"><circle cx="50" cy="50" r="40" fill="red"/></svg>'],
    ... })
    >>> score(solution, submission, 'row_id', random_seed=42)
    0...
    """
    # Convert solution fields to list dtypes and expand
    for colname in ['question', 'choices', 'answer']:
        solution[colname] = solution[colname].apply(ast.literal_eval)
    solution = solution.explode(['question', 'choices', 'answer'])

    # Validate
    if not pd.api.types.is_string_dtype(submission.loc[:, 'svg']):
        raise ParticipantVisibleError('svg must be a string.')

    # Check that SVG code meets defined constraints
    constraints = svg_constraints.SVGConstraints()
    try:
        for svg in submission.loc[:, 'svg']:
            constraints.validate_svg(svg)
    except:
        raise ParticipantVisibleError('SVG code violates constraints.')

    # Score
    vqa_evaluator = VQAEvaluator()
    aesthetic_evaluator = AestheticEvaluator()

    results = []
    rng = np.random.RandomState(random_seed)
    try:
        df = solution.merge(submission, on='id')
        for i, (_, group) in enumerate(df.loc[
            :, ['id', 'question', 'choices', 'answer', 'svg']
        ].groupby('id')):
            questions, choices, answers, svg = [
                group[col_name].to_list()
                for col_name in group.drop('id', axis=1).columns
            ]
            svg = svg[0]  # unpack singleton from list
            group_seed = rng.randint(0, np.iinfo(np.int32).max)
            image_processor = ImageProcessor(image=svg_to_png(svg), seed=group_seed).apply()
            image = image_processor.image.copy()
            aesthetic_score = aesthetic_evaluator.score(image)
            vqa_score = vqa_evaluator.score(questions, choices, answers, image)
            image_processor.reset().apply_random_crop_resize().apply_jpeg_compression(quality=90)
            ocr_score = vqa_evaluator.ocr(image_processor.image)
            instance_score = (
                harmonic_mean(vqa_score, aesthetic_score, beta=0.5) * ocr_score
            )
            results.append(instance_score)

    except:
        raise ParticipantVisibleError('SVG failed to score.')

    fidelity = statistics.mean(results)
    return float(fidelity)


class VQAEvaluator:
    """Evaluates images based on their similarity to a given text description using multiple choice questions."""

    def __init__(self):
        self.quantization_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type='nf4',
            bnb_4bit_use_double_quant=True,
            bnb_4bit_compute_dtype=torch.float16,
        )
        self.letters = string.ascii_uppercase
        self.model_path = kagglehub.model_download(
            'google/paligemma-2/transformers/paligemma2-10b-mix-448'
        )
        self.processor = AutoProcessor.from_pretrained(self.model_path)
        self.model = PaliGemmaForConditionalGeneration.from_pretrained(
            self.model_path,
            low_cpu_mem_usage=True,
            quantization_config=self.quantization_config,
        ).to('cuda')

    def score(self, questions, choices, answers, image, n=4):
        scores = []
        batches = (chunked(qs, n) for qs in [questions, choices, answers])
        for question_batch, choice_batch, answer_batch in zip(*batches, strict=True):
            scores.extend(
                self.score_batch(
                    image,
                    question_batch,
                    choice_batch,
                    answer_batch,
                )
            )
        return statistics.mean(scores)

    def score_batch(
        self,
        image: Image.Image,
        questions: list[str],
        choices_list: list[list[str]],
        answers: list[str],
    ) -> list[float]:
        """Evaluates the image based on multiple choice questions and answers.

        Parameters
        ----------
        image : PIL.Image.Image
            The image to evaluate.
        questions : list[str]
            List of questions about the image.
        choices_list : list[list[str]]
            List of lists of possible answer choices, corresponding to each question.
        answers : list[str]
            List of correct answers from the choices, corresponding to each question.

        Returns
        -------
        list[float]
            List of scores (values between 0 and 1) representing the probability of the correct answer for each question.
        """
        prompts = [
            self.format_prompt(question, choices)
            for question, choices in zip(questions, choices_list, strict=True)
        ]
        batched_choice_probabilities = self.get_choice_probability(
            image, prompts, choices_list
        )

        scores = []
        for i, _ in enumerate(questions):
            choice_probabilities = batched_choice_probabilities[i]
            answer = answers[i]
            answer_probability = 0.0
            for choice, prob in choice_probabilities.items():
                if choice == answer:
                    answer_probability = prob
                    break
            scores.append(answer_probability)

        return scores

    def format_prompt(self, question: str, choices: list[str]) -> str:
        prompt = f'<image>answer en Question: {question}\nChoices:\n'
        for i, choice in enumerate(choices):
            prompt += f'{self.letters[i]}. {choice}\n'
        return prompt

    def mask_choices(self, logits, choices_list):
        """Masks logits for the first token of each choice letter for each question in the batch."""
        batch_size = logits.shape[0]
        masked_logits = torch.full_like(logits, float('-inf'))

        for batch_idx in range(batch_size):
            choices = choices_list[batch_idx]
            for i in range(len(choices)):
                letter_token = self.letters[i]

                first_token = self.processor.tokenizer.encode(
                    letter_token, add_special_tokens=False
                )[0]
                first_token_with_space = self.processor.tokenizer.encode(
                    ' ' + letter_token, add_special_tokens=False
                )[0]

                if isinstance(first_token, int):
                    masked_logits[batch_idx, first_token] = logits[
                        batch_idx, first_token
                    ]
                if isinstance(first_token_with_space, int):
                    masked_logits[batch_idx, first_token_with_space] = logits[
                        batch_idx, first_token_with_space
                    ]

        return masked_logits

    def get_choice_probability(self, image, prompts, choices_list) -> list[dict]:
        inputs = self.processor(
            images=[image] * len(prompts),
            text=prompts,
            return_tensors='pt',
            padding='longest',
        ).to('cuda')

        with torch.no_grad():
            outputs = self.model(**inputs)
            logits = outputs.logits[:, -1, :]  # Logits for the last (predicted) token
            masked_logits = self.mask_choices(logits, choices_list)
            probabilities = torch.softmax(masked_logits, dim=-1)

        batched_choice_probabilities = []
        for batch_idx in range(len(prompts)):
            choice_probabilities = {}
            choices = choices_list[batch_idx]
            for i, choice in enumerate(choices):
                letter_token = self.letters[i]
                first_token = self.processor.tokenizer.encode(
                    letter_token, add_special_tokens=False
                )[0]
                first_token_with_space = self.processor.tokenizer.encode(
                    ' ' + letter_token, add_special_tokens=False
                )[0]

                prob = 0.0
                if isinstance(first_token, int):
                    prob += probabilities[batch_idx, first_token].item()
                if isinstance(first_token_with_space, int):
                    prob += probabilities[batch_idx, first_token_with_space].item()
                choice_probabilities[choice] = prob

            # Renormalize probabilities for each question
            total_prob = sum(choice_probabilities.values())
            if total_prob > 0:
                renormalized_probabilities = {
                    choice: prob / total_prob
                    for choice, prob in choice_probabilities.items()
                }
            else:
                renormalized_probabilities = (
                    choice_probabilities  # Avoid division by zero if total_prob is 0
                )
            batched_choice_probabilities.append(renormalized_probabilities)

        return batched_choice_probabilities

    def ocr(self, image, free_chars=4):
        inputs = (
            self.processor(
                text='<image>ocr\n',
                images=image,
                return_tensors='pt',
            )
            .to(torch.float16)
            .to(self.model.device)
        )
        input_len = inputs['input_ids'].shape[-1]

        with torch.inference_mode():
            outputs = self.model.generate(**inputs, max_new_tokens=32, do_sample=False)
            outputs = outputs[0][input_len:]
            decoded = self.processor.decode(outputs, skip_special_tokens=True)

        num_char = len(decoded)

        # Exponentially decreasing towards 0.0 if more than free_chars detected
        return min(1.0, math.exp(-num_char + free_chars))


class AestheticPredictor(nn.Module):
    def __init__(self, input_size):
        super().__init__()
        self.input_size = input_size
        self.layers = nn.Sequential(
            nn.Linear(self.input_size, 1024),
            nn.Dropout(0.2),
            nn.Linear(1024, 128),
            nn.Dropout(0.2),
            nn.Linear(128, 64),
            nn.Dropout(0.1),
            nn.Linear(64, 16),
            nn.Linear(16, 1),
        )

    def forward(self, x):
        return self.layers(x)


class AestheticEvaluator:
    def __init__(self):
        self.model_path = 'improved-aesthetic-predictor/sac+logos+ava1-l14-linearMSE.pth'
        self.clip_model_path = 'ViT-L/14'
        self.predictor, self.clip_model, self.preprocessor = self.load()

    def load(self):
        """Loads the aesthetic predictor model and CLIP model."""
        state_dict = torch.load(self.model_path, weights_only=True, map_location='cuda')

        # CLIP embedding dim is 768 for CLIP ViT L 14
        predictor = AestheticPredictor(768)
        predictor.load_state_dict(state_dict)
        predictor.to('cuda')
        predictor.eval()
        clip_model, preprocessor = clip.load(self.clip_model_path, device='cuda')

        return predictor, clip_model, preprocessor

    def score(self, image: Image.Image) -> float:
        """Predicts the CLIP aesthetic score of an image."""
        image = self.preprocessor(image).unsqueeze(0).to('cuda')

        with torch.no_grad():
            image_features = self.clip_model.encode_image(image)
            # l2 normalize
            image_features /= image_features.norm(dim=-1, keepdim=True)
            image_features = image_features.cpu().detach().numpy()

        score = self.predictor(torch.from_numpy(image_features).to('cuda').float())

        return score.item() / 10.0  # scale to [0, 1]


def harmonic_mean(a: float, b: float, beta: float = 1.0) -> float:
    """
    Calculate the harmonic mean of two values, weighted using a beta parameter.

    Args:
        a: First value (e.g., precision)
        b: Second value (e.g., recall)
        beta: Weighting parameter

    Returns:
        Weighted harmonic mean
    """
    # Handle zero values to prevent division by zero
    if a <= 0 or b <= 0:
        return 0.0
    return (1 + beta**2) * (a * b) / (beta**2 * a + b)


def svg_to_png(svg_code: str, size: tuple = (384, 384)) -> Image.Image:
    """
    Converts an SVG string to a PNG image using CairoSVG.

    If the SVG does not define a `viewBox`, it will add one using the provided size.

    Parameters
    ----------
    svg_code : str
        The SVG string to convert.
    size : tuple[int, int], default=(384, 384)
        The desired size of the output PNG image (width, height).

    Returns
    -------
    PIL.Image.Image
        The generated PNG image.
    """
    # Ensure SVG has proper size attributes
    if 'viewBox' not in svg_code:
        svg_code = svg_code.replace('<svg', f'<svg viewBox="0 0 {size[0]} {size[1]}"')

    # Convert SVG to PNG
    png_data = cairosvg.svg2png(bytestring=svg_code.encode('utf-8'))
    return Image.open(io.BytesIO(png_data)).convert('RGB').resize(size)


class ImageProcessor:
    def __init__(self, image: Image.Image, seed=None):
        """Initialize with either a path to an image or a PIL Image object."""
        self.image = image
        self.original_image = self.image.copy()
        if seed is not None:
            self.rng = np.random.RandomState(seed)
        else:
            self.rng = np.random

    def reset(self):
        self.image = self.original_image.copy()
        return self
    
    def visualize_comparison(
        self,
        original_name='Original',
        processed_name='Processed',
        figsize=(10, 5),
        show=True,
    ):
        """Display original and processed images side by side."""
        fig, (ax1, ax2) = plt.subplots(1, 2, figsize=figsize)
        ax1.imshow(np.asarray(self.original_image))
        ax1.set_title(original_name)
        ax1.axis('off')

        ax2.imshow(np.asarray(self.image))
        ax2.set_title(processed_name)
        ax2.axis('off')

        title = f'{original_name} vs {processed_name}'
        fig.suptitle(title)
        fig.tight_layout()
        if show:
            plt.show()
        return fig

    def apply_median_filter(self, size=3):
        """Apply median filter to remove outlier pixel values.

        Args:
            size: Size of the median filter window.
        """
        self.image = self.image.filter(ImageFilter.MedianFilter(size=size))
        return self

    def apply_bilateral_filter(self, d=9, sigma_color=75, sigma_space=75):
        """Apply bilateral filter to smooth while preserving edges.

        Args:
            d: Diameter of each pixel neighborhood
            sigma_color: Filter sigma in the color space
            sigma_space: Filter sigma in the coordinate space
        """
        # Convert PIL Image to numpy array for OpenCV
        img_array = np.asarray(self.image)

        # Apply bilateral filter
        filtered = cv2.bilateralFilter(img_array, d, sigma_color, sigma_space)

        # Convert back to PIL Image
        self.image = Image.fromarray(filtered)
        return self

    def apply_fft_low_pass(self, cutoff_frequency=0.5):
        """Apply low-pass filter in the frequency domain using FFT.

        Args:
            cutoff_frequency: Normalized cutoff frequency (0-1).
                Lower values remove more high frequencies.
        """
        # Convert to numpy array, ensuring float32 for FFT
        img_array = np.array(self.image, dtype=np.float32)

        # Process each color channel separately
        result = np.zeros_like(img_array)
        for i in range(3):  # For RGB channels
            # Apply FFT
            f = np.fft.fft2(img_array[:, :, i])
            fshift = np.fft.fftshift(f)

            # Create a low-pass filter mask
            rows, cols = img_array[:, :, i].shape
            crow, ccol = rows // 2, cols // 2
            mask = np.zeros((rows, cols), np.float32)
            r = int(min(crow, ccol) * cutoff_frequency)
            center = [crow, ccol]
            x, y = np.ogrid[:rows, :cols]
            mask_area = (x - center[0]) ** 2 + (y - center[1]) ** 2 <= r * r
            mask[mask_area] = 1

            # Apply mask and inverse FFT
            fshift_filtered = fshift * mask
            f_ishift = np.fft.ifftshift(fshift_filtered)
            img_back = np.fft.ifft2(f_ishift)
            img_back = np.real(img_back)

            result[:, :, i] = img_back

        # Clip to 0-255 range and convert to uint8 after processing all channels
        result = np.clip(result, 0, 255).astype(np.uint8)

        # Convert back to PIL Image
        self.image = Image.fromarray(result)
        return self

    def apply_jpeg_compression(self, quality=85):
        """Apply JPEG compression.

        Args:
            quality: JPEG quality (0-95). Lower values increase compression.
        """
        buffer = io.BytesIO()
        self.image.save(buffer, format='JPEG', quality=quality)
        buffer.seek(0)
        self.image = Image.open(buffer)
        return self

    def apply_random_crop_resize(self, crop_percent=0.05):
        """Randomly crop and resize back to original dimensions.

        Args:
            crop_percent: Percentage of image to crop (0-0.4).
        """
        width, height = self.image.size
        crop_pixels_w = int(width * crop_percent)
        crop_pixels_h = int(height * crop_percent)

        left = self.rng.randint(0, crop_pixels_w + 1)
        top = self.rng.randint(0, crop_pixels_h + 1)
        right = width - self.rng.randint(0, crop_pixels_w + 1)
        bottom = height - self.rng.randint(0, crop_pixels_h + 1)

        self.image = self.image.crop((left, top, right, bottom))
        self.image = self.image.resize((width, height), Image.BILINEAR)
        return self

    def apply(self):
        """Apply an ensemble of defenses."""
        return (
            self.apply_random_crop_resize(crop_percent=0.03)
            .apply_jpeg_compression(quality=95)
            .apply_median_filter(size=9)
            .apply_fft_low_pass(cutoff_frequency=0.5)
            .apply_bilateral_filter(d=5, sigma_color=75, sigma_space=75)
            .apply_jpeg_compression(quality=92)
        )