Rename modeling_persianocr.py to model.py

modeling_persianocr.py → model.py

@@ -1,9 +1,10 @@
-import torch
 import torch.nn as nn
-from transformers import PreTrainedModel, PretrainedConfig
+import torch
 
-def GN(c, groups=16):
-    return nn.GroupNorm(min(groups, c), c)
+# -----------------------------
+# 3️⃣ Model definition
+# -----------------------------
+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):
@@ -15,8 +16,7 @@ class LightResNetCNN(nn.Module):
         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))
-
+        self.adaptive_pool = nn.AdaptiveAvgPool2d((self.adaptive_height, None))
     def forward(self, x):
         for i in range(1, 7):
             x = getattr(self, f"layer{i}")(x)
@@ -32,40 +32,23 @@ class PositionalEncoding(nn.Module):
         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):
+class CNN_Transformer_OCR(nn.Module):
+    def __init__(self, num_classes, d_model=1280, nhead=16, num_layers=8, dropout=0.2):
+        super().__init__()
+        self.cnn = LightResNetCNN(in_channels=1, adaptive_height=8)
+        self.proj = nn.Linear(128 * 8, d_model)
+        self.posenc = PositionalEncoding(d_model)
+        encoder_layer = nn.TransformerEncoderLayer(d_model, nhead, batch_first=True, dropout=dropout)
+        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
+        self.fc = nn.Linear(d_model, num_classes)
+    def forward(self, x):
         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}
+        out = self.fc(out)
+        return out.log_softmax(2)