modeling_persianocr.py

import torch
import torch.nn as nn
from transformers import PreTrainedModel, PretrainedConfig

def GN(c, groups=16):
    return nn.GroupNorm(min(groups, c), c)

class LightResNetCNN(nn.Module):
    def __init__(self, in_channels=1, adaptive_height=8):
        super().__init__()
        self.adaptive_height = adaptive_height
        self.layer1 = nn.Sequential(nn.Conv2d(in_channels, 32, 3, 1, 1), GN(32), nn.ReLU(), nn.MaxPool2d(2, 2))
        self.layer2 = nn.Sequential(nn.Conv2d(32, 64, 3, 1, 1), GN(64), nn.ReLU(), nn.MaxPool2d(2, 2))
        self.layer3 = nn.Sequential(nn.Conv2d(64, 128, 3, 1, 1), GN(128), nn.ReLU(), nn.MaxPool2d(2, 2))
        self.layer4 = nn.Sequential(nn.Conv2d(128, 256, 3, 1, 1), GN(256), nn.ReLU())
        self.layer5 = nn.Sequential(nn.Conv2d(256, 256, 3, 1, 1), GN(256), nn.ReLU())
        self.layer6 = nn.Sequential(nn.Conv2d(256, 128, 3, 1, 1), GN(128), nn.ReLU())
        self.adaptive_pool = nn.AdaptiveAvgPool2d((adaptive_height, None))

    def forward(self, x):
        for i in range(1, 7):
            x = getattr(self, f"layer{i}")(x)
        x = self.adaptive_pool(x)
        return x

class PositionalEncoding(nn.Module):
    def __init__(self, d_model, max_len=2000):
        super().__init__()
        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2) * (-torch.log(torch.tensor(10000.0)) / d_model))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        self.register_buffer("pe", pe.unsqueeze(0))

    def forward(self, x):
        return x + self.pe[:, :x.size(1), :]

class PersianOCRConfig(PretrainedConfig):
    model_type = "persianocr"

    def __init__(self, num_classes=100, d_model=1280, nhead=16, num_layers=8, dropout=0.2, adaptive_height=8, **kwargs):
        super().__init__(**kwargs)
        self.num_classes = num_classes
        self.d_model = d_model
        self.nhead = nhead
        self.num_layers = num_layers
        self.dropout = dropout
        self.adaptive_height = adaptive_height

class PersianOCRModel(PreTrainedModel):
    config_class = PersianOCRConfig

    def __init__(self, config):
        super().__init__(config)
        self.cnn = LightResNetCNN(in_channels=1, adaptive_height=config.adaptive_height)
        self.proj = nn.Linear(128 * config.adaptive_height, config.d_model)
        self.posenc = PositionalEncoding(config.d_model)
        encoder_layer = nn.TransformerEncoderLayer(config.d_model, config.nhead, batch_first=True, dropout=config.dropout)
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=config.num_layers)
        self.fc = nn.Linear(config.d_model, config.num_classes)
        self.post_init()

    def forward(self, x, labels=None):
        f = self.cnn(x)
        B, C, H, W = f.size()
        f = f.permute(0, 3, 1, 2).reshape(B, W, C * H)
        f = self.posenc(self.proj(f))
        out = self.transformer(f)
        logits = self.fc(out)
        return {"logits": logits}