File size: 7,537 Bytes
5486f6f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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

from dataclasses import dataclass, field
from typing import Optional
import pandas as pd
import os

import torch
from transformers import VisionEncoderDecoderModel, TrOCRProcessor, Seq2SeqTrainer, Seq2SeqTrainingArguments, default_data_collator, EarlyStoppingCallback
from peft import LoraConfig, get_peft_model

from data import AphaPenDataset
import evaluate
from sklearn.model_selection import train_test_split

from src.calibrator import EncoderDecoderCalibrator
from src.loss import MarginLoss, KLRegularization
from src.similarity import CERSimilarity
from datetime import datetime
import torch.nn.functional as F

os.environ["WANDB_PROJECT"] = "Alphapen-TrOCR"

# # Step 1: Load the dataset
train_df_path = "/mnt/data1/Datasets/AlphaPen/" + "training_data.csv"
test_df_path = "/mnt/data1/Datasets/AlphaPen/" + "testing_data.csv"

#train_df = pd.read_csv(train_df_path)
#train_df.dropna(inplace=True)

train_df = pd.read_csv(test_df_path)[:4000]
train_df.dropna(inplace=True)

test_df = pd.read_csv(test_df_path)[4000:]
test_df.dropna(inplace=True)

# we reset the indices to start from zero
train_df.reset_index(drop=True, inplace=True)
test_df.reset_index(drop=True, inplace=True)
root_dir = "/mnt/data1/Datasets/OCR/Alphapen/clean_data/final_cropped_rotated_"

model_name = "microsoft/trocr-large-handwritten"

processor = TrOCRProcessor.from_pretrained(model_name)
train_dataset = AphaPenDataset(root_dir=root_dir, df=train_df, processor=processor)
eval_dataset = AphaPenDataset(root_dir=root_dir, df=test_df, processor=processor)

# Step 2: Load the model
model = VisionEncoderDecoderModel.from_pretrained(model_name)

# set special tokens used for creating the decoder_input_ids from the labels
model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
model.config.pad_token_id = processor.tokenizer.pad_token_id
# make sure vocab size is set correctly
model.config.vocab_size = model.config.decoder.vocab_size
# for peft
model.vocab_size = model.config.decoder.vocab_size

# set beam search parameters
model.config.eos_token_id = processor.tokenizer.sep_token_id
model.config.max_length = 64
model.config.early_stopping = True
model.config.no_repeat_ngram_size = 3
model.config.length_penalty = 2.0
model.config.num_beams = 4

# LoRa
lora_config = LoraConfig(
    r=1,
    lora_alpha=8,
    lora_dropout=0.1,
    target_modules=[
        'query',
        'key',
        'value',
        'intermediate.dense',
        'output.dense',
        #'wte',
        #'wpe',
        #'c_attn',
        #'c_proj',
        #'q_attn',
        #'c_fc'
    ],
)
model = get_peft_model(model, lora_config)

tokenizer = processor.tokenizer
# sim = CERSimilarity(tokenizer)
# loss = MarginLoss(sim, beta=0.1, num_samples=60)
# reg = KLRegularization(model)
# calibrator = EncoderDecoderCalibrator(model, loss, reg, 15, 15)

# # Step 3: Define the training arguments
training_args = Seq2SeqTrainingArguments(
    predict_with_generate=True,
    evaluation_strategy="steps",
    per_device_train_batch_size=8,
    per_device_eval_batch_size=8,
    bf16=True,
    bf16_full_eval=True,
    output_dir="./",
    logging_steps=100,
    save_steps=20000,
    eval_steps=500,
    # report_to="wandb",
    optim="adamw_torch_fused",
    lr_scheduler_type="cosine",
    gradient_accumulation_steps=2,
    learning_rate=1.0e-4,
    max_steps=10000,
    run_name=f"trocr-LoRA-{datetime.now().strftime('%Y-%m-%d-%H-%M-%s')}",
)

# Step 4: Define a metric

cer_metric = evaluate.load("cer")

def compute_cer(pred, target):
    return cer_metric.compute(predictions=[pred], references=[target])['cer']

def generate_candidates(model, pixel_values, num_candidates=10):
    return model.generate(
        pixel_values,
        num_return_sequences=num_candidates,
        num_beams=num_candidates,
        output_scores=True,
        return_dict_in_generate=True
    )

def rank_loss(positive_scores, negative_scores):
    return F.relu(1 - positive_scores + negative_scores).mean()

def margin_loss(positive_scores, negative_scores, margin=0.1):
    return F.relu(margin - positive_scores + negative_scores).mean()

def calibration_loss(model, pixel_values, ground_truth, processor, loss_type='margin'):
    candidates = generate_candidates(model, pixel_values)
    candidate_sequences = processor.batch_decode(candidates.sequences, skip_special_tokens=True)
    
    ground_truth = processor.decode(ground_truth, skip_special_tokens=True)
    
    similarities = [1 - compute_cer(cand, ground_truth) for cand in candidate_sequences]
    
    positive_pairs = []
    negative_pairs = []
    
    for i in range(len(similarities)):
        for j in range(i + 1, len(similarities)):
            if similarities[i] > similarities[j]:
                positive_pairs.append((i, j))
            else:
                negative_pairs.append((i, j))
    
    if not positive_pairs or not negative_pairs:
        return torch.tensor(0.0, device=pixel_values.device)
    
    positive_scores = candidates.sequences_scores[torch.tensor(positive_pairs)[:, 0]]
    negative_scores = candidates.sequences_scores[torch.tensor(negative_pairs)[:, 1]]
    
    if loss_type == 'rank':
        return rank_loss(positive_scores, negative_scores)
    elif loss_type == 'margin':
        return margin_loss(positive_scores, negative_scores)
    else:
        raise ValueError("Invalid loss type. Choose 'rank' or 'margin'.")

class CalibratedTrainer(Seq2SeqTrainer):
    def __init__(self, *args, **kwargs):
        self.processor = kwargs.pop('processor', None)
        self.calibration_weight = kwargs.pop('calibration_weight', 0.1)
        self.calibration_loss_type = kwargs.pop('calibration_loss_type', 'margin')
        super().__init__(*args, **kwargs)

    def compute_loss(self, model, inputs, return_outputs=False):
        labels = inputs.pop("labels")
        pixel_values = inputs['pixel_values']
        
        outputs = model.generate(**inputs, return_dict_in_generate=True, output_logits=True)
        
        logits = outputs.logits
        print(logits)

        # Original cross-entropy loss
        ce_loss = F.cross_entropy(logits.view(-1, logits.size(-1)), labels.view(-1), ignore_index=-100)

        # Calibration loss
        cal_loss = calibration_loss(model, pixel_values, labels, self.processor, self.calibration_loss_type)

        # Combine losses
        total_loss = ce_loss + self.calibration_weight * cal_loss

        return (total_loss, outputs) if return_outputs else total_loss



def compute_metrics(pred):
    # accuracy_metric = evaluate.load("precision")
    cer_metric = evaluate.load("cer")

    labels_ids = pred.label_ids
    pred_ids = pred.predictions

    pred_str = processor.batch_decode(pred_ids, skip_special_tokens=True)
    labels_ids[labels_ids == -100] = processor.tokenizer.pad_token_id
    label_str = processor.batch_decode(labels_ids, skip_special_tokens=True)

    cer = cer_metric.compute(predictions=pred_str, references=label_str)
    # accuracy = accuracy_metric.compute(predictions=pred_ids.tolist(), references=labels_ids.tolist())

    return {"cer": cer}

# # Step 5: Define the Trainer
# Step 5: Define the Trainer
trainer = CalibratedTrainer(
    model=model,
    tokenizer=processor.feature_extractor,
    args=training_args,
    # compute_metrics=compute_metrics,
    train_dataset=train_dataset,
    eval_dataset=eval_dataset,
    data_collator=default_data_collator,
    processor=processor,
    calibration_weight=0.1,
    calibration_loss_type='margin'  # or 'rank'
)


trainer.train()