model

app.py

@@ -3,11 +3,13 @@ import re
 import torch
 import torch.nn as nn
 import gradio as gr
+from huggingface_hub import hf_hub_download
 
 ########################################
 # Model definitions (same as in notebook)
 ########################################
 
+
 class EncoderGRU(nn.Module):
     def __init__(self, input_dim, emb_dim, hid_dim, num_layers=1, dropout=0.1, pad_idx=0):
         super().__init__()
@@ -72,12 +74,27 @@ class Seq2Seq(nn.Module):
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
+# 🔴 CHANGE THIS: your actual model repo id
+MODEL_REPO = "harishwar017/hindi-roman-gru"
+
+# Download files from HF Hub into the Space’s local cache
+src_json_path = hf_hub_download(repo_id=MODEL_REPO, filename="src_stoi.json")
+tgt_json_path = hf_hub_download(repo_id=MODEL_REPO, filename="tgt_stoi.json")
+model_path    = hf_hub_download(repo_id=MODEL_REPO, filename="best_hindi_roman_gru.pt")
+
 # Load vocabularies
-with open("src_stoi.json", "r", encoding="utf-8") as f:
+with open(src_json_path, "r", encoding="utf-8") as f:
     src_stoi = json.load(f)
 
-with open("tgt_stoi.json", "r", encoding="utf-8") as f:
+with open(tgt_json_path, "r", encoding="utf-8") as f:
     tgt_stoi = json.load(f)
+    
+# # Load vocabularies
+# with open("src_stoi.json", "r", encoding="utf-8") as f:
+#     src_stoi = json.load(f)
+
+# with open("tgt_stoi.json", "r", encoding="utf-8") as f:
+#     tgt_stoi = json.load(f)
 
 # Build inverse mapping for target
 tgt_itos = {int(v): k for k, v in tgt_stoi.items()}  # keys might be strings in JSON
@@ -127,13 +144,9 @@ model = Seq2Seq(
     device=device,
 ).to(device)
 
-state_dict = torch.load("best_hindi_roman_gru.pt", map_location=device)
-model.encoder.load_state_dict(
-    {k.replace("encoder.", ""): v for k, v in state_dict.items() if k.startswith("encoder.")}
-)
-model.decoder.load_state_dict(
-    {k.replace("decoder.", ""): v for k, v in state_dict.items() if k.startswith("decoder.")}
-)
+# Load weights that you saved from training: torch.save(model.state_dict(), "best_hindi_roman_gru.pt")
+state_dict = torch.load(model_path, map_location=device)
+model.load_state_dict(state_dict)
 model.eval()
 
 
@@ -234,10 +247,7 @@ demo = gr.Interface(
     inputs=gr.Textbox(lines=3, label="Hindi sentence"),
     outputs=gr.Textbox(lines=3, label="Romanized (Latin script)"),
     title="Hindi → Roman Transliteration (Char-level GRU)",
-    description=(
-        "Paste a Hindi sentence; the model splits it into words, "
-        "applies a character-level GRU transliteration model, and rejoins the output."
-    ),
+    description="Paste a Hindi sentence; the model splits it into words, transliterates each with a GRU, and rejoins the output.",
 )
 
 if __name__ == "__main__":

requirements.txt

@@ -1,2 +1,3 @@
 torch
 gradio
+huggingface_hub