File size: 7,391 Bytes
acec273
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from pycparser.ply.yacc import token
from ultralytics import YOLO
from transformers import TrOCRProcessor, VisionEncoderDecoderModel, AutoModelForCausalLM, pipeline, AutoModelForMaskedLM
from PIL import Image
import numpy as np
import pandas as pd
from nltk.translate import bleu_score
from nltk.translate.bleu_score import SmoothingFunction
import torch

yolo_weights_path = "final_wts.pt"

device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_available() else 'cpu'

processor = TrOCRProcessor.from_pretrained('microsoft/trocr-large-handwritten')
trocr_model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-large-handwritten').to(device)
trocr_model.config.num_beams = 1

yolo_model = YOLO(yolo_weights_path).to('mps')
unmasker_large = pipeline('fill-mask', model='roberta-large', device=device)
roberta_model = AutoModelForMaskedLM.from_pretrained("roberta-large").to(device)

print(f'TrOCR and YOLO Models loaded on {device}')


-------------------------------------------------------


CONFIDENCE_THRESHOLD = 0.72
BLEU_THRESHOLD = 0.6


def inference(image_path, debug=False, return_texts='final'):
    def get_cropped_images(image_path):
        results = yolo_model(image_path, save=True)
        patches = []
        ys = []
        for box in sorted(results[0].boxes, key=lambda x: x.xywh[0][1]):
            image = Image.open(image_path).convert("RGB")
            x_center, y_center, w, h  = box.xywh[0].cpu().numpy()
            x, y = x_center - w / 2, y_center - h / 2
            cropped_image = image.crop((x, y, x + w, y + h))
            patches.append(cropped_image)
            ys.append(y)
        bounding_box_path = results[0].save_dir + results[0].path[results[0].path.rindex('/'):-4] + '.jpg'
        return patches, ys, bounding_box_path

    def get_model_output(images):
        pixel_values = processor(images=images, return_tensors="pt").pixel_values.to(device)
        output = trocr_model.generate(pixel_values, return_dict_in_generate=True, output_logits=True, max_new_tokens=30)
        generated_texts = processor.batch_decode(output.sequences, skip_special_tokens=True)
        generated_tokens = [processor.tokenizer.convert_ids_to_tokens(seq) for seq in output.sequences]
        stacked_logits = torch.stack(output.logits, dim=1)
        return generated_texts, stacked_logits, generated_tokens

    def get_scores(logits):
        scores = logits.softmax(-1).max(-1).values.mean(-1)
        return scores

    def post_process_texts(generated_texts):
        for i in range(len(generated_texts)):
            if len(generated_texts[i]) > 2 and generated_texts[i][:2] == '# ':
                generated_texts[i] = generated_texts[i][2:]
                
            if len(generated_texts[i]) > 2 and generated_texts[i][-2:] == ' #':
                generated_texts[i] = generated_texts[i][:-2]
        return generated_texts

    def get_qualified_texts(generated_texts, scores, y, logits, tokens):
        qualified_texts = []
        for text, score, y_i, logits_i, tokens_i in zip(generated_texts, scores, y, logits, tokens):
            if score > CONFIDENCE_THRESHOLD:
                qualified_texts.append({
                    'text': text,
                    'score': score,
                    'y': y_i,
                    'logits': logits_i,
                    'tokens': tokens_i
                })
        return qualified_texts

    def get_adjacent_bleu_scores(qualified_texts):
        def get_bleu_score(hypothesis, references):
            weights = [0.5, 0.5]
            smoothing = SmoothingFunction()
            return bleu_score.sentence_bleu(references, hypothesis, weights=weights,
                                            smoothing_function=smoothing.method1)

        for i in range(len(qualified_texts)):
            hyp = qualified_texts[i]['text'].split()
            bleu = 0
            if i < len(qualified_texts) - 1:
                ref = qualified_texts[i + 1]['text'].split()
                bleu = get_bleu_score(hyp, [ref])
            qualified_texts[i]['bleu'] = bleu
        return qualified_texts

    def remove_overlapping_texts(qualified_texts):
        final_texts = []
        new = True
        for i in range(len(qualified_texts)):
            if new:
                final_texts.append(qualified_texts[i])
            else:
                if final_texts[-1]['score'] < qualified_texts[i]['score']:
                    final_texts[-1] = qualified_texts[i]
            new = qualified_texts[i]['bleu'] < BLEU_THRESHOLD
        return final_texts

    cropped_images, y, bounding_box_path = get_cropped_images(image_path)
    if debug:
        print('Number of cropped images:', len(cropped_images))
    generated_texts, logits, gen_tokens = get_model_output(cropped_images)
    normalised_scores = get_scores(logits)
    if return_texts == 'generated':
        return pd.DataFrame({
            'text': generated_texts,
            'score': normalised_scores,
            'y': y,
        })
    generated_texts = post_process_texts(generated_texts)
    if return_texts == 'post_processed':
        return pd.DataFrame({
            'text': generated_texts,
            'score': normalised_scores,
            'y': y
        })
    qualified_texts = get_qualified_texts(generated_texts, normalised_scores, y, logits, gen_tokens)
    if return_texts == 'qualified':
        return pd.DataFrame(qualified_texts)
    qualified_texts = get_adjacent_bleu_scores(qualified_texts)
    if return_texts == 'qualified_with_bleu':
        return pd.DataFrame(qualified_texts)
    final_texts = remove_overlapping_texts(qualified_texts)
    final_texts_df = pd.DataFrame(final_texts, columns=['text', 'score', 'y'])
    final_tokens = [text['tokens'] for text in final_texts]
    final_logits = [text['logits'] for text in final_texts]
    if return_texts == 'final':
        return final_texts_df
    
    return final_texts_df, bounding_box_path, final_tokens, final_logits, generated_texts


image_path = "raw_dataset/g06-037h.png"
df, bounding_path, tokens, logits, gen_texts = inference(image_path, debug=False, return_texts='final_v2')

----------------------------------------------------------------


def get_new_logits(tokens):
    inputs = tokens.reshape(1, -1)
    # Get the logits from the model
    with torch.no_grad():
        outputs = roberta_model(input_ids=inputs, attention_mask=torch.ones(inputs.shape).to(device))
        logits = outputs.logits


    logits_flattened = logits.reshape(-1, slogits.shape[-1])
    print(processor.batch_decode([logits_flattened.argmax(-1)], skip_special_tokens=True))
    return logits.reshape(tokens.shape + (logits.shape[-1],))


slogits = torch.stack([logit for logit in logits], dim=0)
tokens = slogits.argmax(-1)
confidence = slogits.softmax(-1).max(-1).values
indices = torch.where(confidence < 0.5)
# put 50264(mask) when confidence < 0.5
for i, j in zip(indices[0], indices[1]):
    if i != 6:
        continue
    tokens[i, j] = torch.tensor(50264)

new_logits = get_new_logits(tokens)


----------------------------------------------------------------


for i, j in zip(indices[0], indices[1]):
    slogits[i, j] = slogits[i, j] * 0.1 + new_logits[i, j] * 0.5

logits_flattened = slogits.reshape(-1, slogits.shape[-1])
processor.batch_decode([logits_flattened.argmax(-1)], skip_special_tokens=True)