finetune_phi3_vision.py

from datasets import Dataset, DatasetDict, Image
import pandas as pd
import os

import torch
from peft import LoraConfig
from transformers import AutoProcessor, BitsAndBytesConfig
from transformers import AutoModelForCausalLM, AutoModelForVision2Seq
from datetime import datetime
import evaluate
# Define train and test size.
TRAIN_SAMPLES = 1000
TEST_SAMPLES = 200
TEST_SIZE = 0.166 #

# Define the directory containing the images.
df_path = "/mnt/data1/Datasets/AlphaPen/" + "training_data.csv"
df = pd.read_csv(df_path)
df.dropna(inplace=True)
df["id"] = range(df.shape[0])
df["query"] = "What is shown in this image?"

root_dir = "/mnt/data1/Datasets/OCR/Alphapen/clean_data/final_cropped_rotated_"
image_paths = [root_dir + img for img in df.filename]

# New batch
df_path_2 = "/mnt/data1/Datasets/AlphaPen/" + "training_b2.csv"
df_2 = pd.read_csv(df_path_2)
df_2.dropna(inplace=True)
df_2["id"] = range(df_2.shape[0])
df_2["query"] = "What is shown in this image?"

root_dir_2 = "/mnt/data1/Datasets/OCR/Alphapen/DataBatch2/clean_data/cropped_data/cropped_"
image_paths_2 = [root_dir_2 + img for img in df_2.filename]
# Create a list of other columns such as id, query, and answer.
ids = range(df.shape[0] + df_2.shape[0])
queries = df['query'].tolist() + df_2['query'].tolist()
answers = df['text'].tolist() + df_2['text'].tolist()

# Create the dataset dictionary.
dataset_dict = {
    'id': ids,
    'image': image_paths + image_paths_2,
    'query': queries,
    'answers': answers
}

# Create the dataset.
dataset = Dataset.from_dict(dataset_dict)

# Cast the 'image' column to Image type.
dataset = dataset.cast_column("image", Image())

# Split the dataset into train and test.
split_dataset = dataset.train_test_split(test_size=TEST_SIZE, shuffle=False)

train_dataset = split_dataset["train"]
eval_dataset  = split_dataset["test"]
print(len(train_dataset))
# Push the dataset on Hugging Face Hub.
# split_dataset.push_to_hub("NSTiwari/DocumentIDEFICS_QA")

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

# Define model ID
# model_id = "microsoft/Phi-3-vision-128k-instruct"
model_id =  "HuggingFaceM4/idefics2-8b"

DEVICE = "cuda:0"
USE_LORA = False
USE_QLORA = True

processor = AutoProcessor.from_pretrained(
    model_id,
    do_image_splitting=False
)

# print(processor.tokenizer.additional_special_tokens.index("<image>"))
if USE_QLORA or USE_LORA:
    lora_config = LoraConfig(
        r=64,
        lora_alpha=16,
        lora_dropout=0.1,
    #     target_modules= [
    #     "q_proj",
    #     "k_proj",
    #     "v_proj",
    #     "o_proj",
    #     "gate_proj",
    #     "up_proj",
    #     # "down_proj",
    # ], 
        target_modules = '.*(text_model|modality_projection|perceiver_resampler).*(down_proj|gate_proj|up_proj|k_proj|q_proj|v_proj|o_proj).*$',
        use_dora=False if USE_QLORA else True,
        init_lora_weights="gaussian"
    )
    if USE_QLORA:
        bnb_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=torch.float16
        )
    model = AutoModelForVision2Seq.from_pretrained(
        model_id,
        torch_dtype=torch.float16,
        quantization_config=bnb_config if USE_QLORA else None,
        trust_remote_code=True
    )
    model.config.decoder_start_token_id = processor.tokenizer.cls_token_id
    model.config.pad_token_id = processor.tokenizer.pad_token_id
    model.config.max_length= 128
    model.add_adapter(lora_config)
    model.enable_adapters()
else:
    model = AutoModelForVision2Seq.from_pretrained(
        model_id,
        torch_dtype=torch.float16,
        _attn_implementation="flash_attention_2", # Need GPUs like A100 or H100.
        trust_remote_code=True
    ).to(DEVICE)



import random

class MyDataCollator:
    def __init__(self, processor):
        self.processor = processor
        self.image_token_id = processor.tokenizer.additional_special_tokens_ids[
            processor.tokenizer.additional_special_tokens.index("<image>")
        ]

    def __call__(self, examples):
        texts = []
        images = []
        for example in examples:
            image = example["image"]
            # print(example["query"])
            question = example["query"]
            answer = example["answers"]
            messages = [
                {
                    "role": "user",
                    "content": [
                        {"type": "text", "text": "OCR the text in the image."},
                        {"type": "image"},
                        {"type": "text", "text": question}
                    ]
                },
                {
                    "role": "assistant",
                    "content": [
                        {"type": "text", "text": answer}
                    ]
                }
            ]
            text = processor.apply_chat_template(messages, add_generation_prompt=False)
            texts.append(text.strip())
            images.append([image])

        batch = processor(text=texts, images=images, return_tensors="pt", padding=True)

        labels = batch["input_ids"].clone()
        # labels[labels == processor.tokenizer.pad_token_id] = self.image_token_id
        batch["labels"] = labels

        return batch

data_collator = MyDataCollator(processor)

from transformers import TrainingArguments, Trainer, Seq2SeqTrainer, Seq2SeqTrainingArguments

training_args = Seq2SeqTrainingArguments(
    predict_with_generate=True,
    output_dir = "idefics2",
    learning_rate = 2e-4,
    fp16 = True,
    per_device_train_batch_size = 8,
    per_device_eval_batch_size = 8,
    gradient_accumulation_steps = 2,
    dataloader_pin_memory = False,
    save_total_limit = 3,
    eval_strategy ="steps",
    save_strategy = "steps",
    eval_steps = 200,
    save_steps = 10000,
    max_steps = 50000,
    logging_steps = 10,
    remove_unused_columns = False,
    push_to_hub=True,
    label_names = ["labels"],
    load_best_model_at_end = False,
    report_to = "wandb",
    optim = "paged_adamw_8bit",
    run_name=f"idefics2-vision-LoRA-{datetime.now().strftime('%Y-%m-%d-%H-%M-%s')}",
    hub_model_id="hadrakey/alphapen_idefics2_finetune_v1",
)

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

    labels_ids = pred.label_ids
    pred_ids = pred.predictions
    # print(pred_ids)
    # print(labels_ids)
    # max_length = max(pred_ids.shape[1], labels_ids.shape[1])
    # generated_texts = processor.batch_decode(generated_ids[:, inputs["input_ids"].size(1):], skip_special_tokens=True)
    pred_str = processor.batch_decode(pred_ids, skip_special_tokens=True)
    pred_str = [word.lower() for word in pred_str]
    # print(pred_str)
    # 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)
    label_str = [word.lower() for word in label_str]
    # print(label_str)
    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}


trainer = Seq2SeqTrainer(
    model = model,
    args = training_args,
    data_collator = data_collator,
    train_dataset = train_dataset,
    eval_dataset = eval_dataset,
    compute_metrics=compute_metrics,
)

trainer.train()