utils.py

from constants import *
from transformers import AutoTokenizer
import torch
import numpy as np
from PIL import Image
from torchvision import transforms


def get_tokenizer():
    tokenizer = AutoTokenizer.from_pretrained("gpt2")
    point_tokens = [f"coord_bin_{i}" for i in range(0, NUM_BINS)]
    new_tokens = [
        "<point_start>", "<point_end>", "<result_start>",
        "<result_end>", "<pointx_start>", "<pointx_end>",
        "<pointy_start>", "<pointy_end>", 
        *point_tokens
    ]
    tokenizer.add_tokens(new_tokens)
    # Ensure pad token is set (GPT2 usually doesn't have one by default)
    if tokenizer.pad_token is None:
        tokenizer.add_special_tokens({'pad_token': '[PAD]'}) # Or use eos_token if preferred
        # tokenizer.pad_token_id = tokenizer.eos_token_id # Alternative if we want padding to be EOS

    print(f"Tokenizer pad token: {tokenizer.pad_token}, ID: {tokenizer.pad_token_id}")
    print(f"Tokenizer EOS token: {tokenizer.eos_token}, ID: {tokenizer.eos_token_id}")

    # Check if pad token ID is valid
    if tokenizer.pad_token_id is None:
         raise ValueError("Tokenizer pad token ID is not set!")

    return tokenizer, len(tokenizer)

def image_to_tensor(image, image_size=IMAGE_SIZE):
    if image.mode != 'RGB':
        image = image.convert('RGB')
    # We avoid the hassle of calculating 
    # changed co-ordinates for rotation etc for now. Can be added later.
    transform = transforms.Compose([
        transforms.Resize((image_size, image_size)),
        transforms.ToTensor(),
        transforms.Normalize(mean=IMAGE_MEAN, std=IMAGE_STD)
    ])
    return transform(image)

def tensor_to_image(tensor):
    tensor = tensor.clone().detach()
    if tensor.is_cuda:
        tensor = tensor.cpu()
    mean = torch.tensor(IMAGE_MEAN).view(3, 1, 1)
    std = torch.tensor(IMAGE_STD).view(3, 1, 1)
    tensor = tensor * std + mean
    tensor = torch.clamp(tensor, 0, 1)
    image_np = tensor.numpy().transpose(1, 2, 0)
    image_np = (image_np * 255).astype(np.uint8)
    return Image.fromarray(image_np)

tokenizer, vocab_size = get_tokenizer() # Initialize tokenizer globally