Download checkpoint from HF hub in OcrReorderPipeline

app.py

@@ -34,6 +34,48 @@ tokenizer = AutoTokenizer.from_pretrained(
     repo, subfolder="preprocessor"
 )
 
+# Ensure decoder_start_token_id is set
+if t5_model.config.decoder_start_token_id is None:
+    # Fallback to bos_token_id if present
+    t5_model.config.decoder_start_token_id = tokenizer.bos_token_id
+
+# Projection head: load from checkpoint
+ckpt_file = hf_hub_download(repo_id=repo, filename="pytorch_model.bin")
+ckpt      = torch.load(ckpt_file, map_location="cpu")
+proj_state= ckpt["projection"]
+projection = torch.nn.Sequential(
+    torch.nn.Linear(768, t5_model.config.d_model),
+    torch.nn.LayerNorm(t5_model.config.d_model),
+    torch.nn.GELU()
+)
+projection.load_state_dict(proj_state)
+projection.eval()
+
+# Move models to CPU (Spaces are CPU-only)
+device = torch.device("cpu")
+layout_model.to(device)
+t5_model.to(device)
+projection.to(device)
+repo = "Uddipan107/ocr-layoutlmv3-base-t5-small"
+
+# Processor for LayoutLMv3
+processor = AutoProcessor.from_pretrained(
+    repo,
+    subfolder="preprocessor",
+    apply_ocr=False
+)
+
+# LayoutLMv3 encoder
+layout_model = LayoutLMv3Model.from_pretrained(repo)
+layout_model.eval()
+
+# T5 decoder & tokenizer
+t5_model = T5ForConditionalGeneration.from_pretrained(repo)
+t5_model.eval()
+tokenizer = AutoTokenizer.from_pretrained(
+    repo, subfolder="preprocessor"
+)
+
 # Projection head: load from checkpoint
 ckpt_file = hf_hub_download(repo_id=repo, filename="pytorch_model.bin")
 ckpt      = torch.load(ckpt_file, map_location="cpu")