Upload app.py with huggingface_hub

app.py

@@ -0,0 +1,206 @@
+"""
+Gradio Web Demo for Seq2Seq Document Generation
+================================================
+Interactive UI: paste a long source document, get a generated summary
+with greedy or beam search decoding, plus an attention heatmap.
+
+Run locally:    python app.py
+Deploy:         HuggingFace Spaces (Gradio SDK)
+"""
+
+import os
+import sys
+import io
+import base64
+import torch
+import gradio as gr
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "src"))
+
+from model import build_model
+from inference import generate_document, greedy_decode
+from preprocessing import Vocabulary, tokenize
+
+# ----------------------------------------------------------------------
+# Load model once at startup
+# ----------------------------------------------------------------------
+BASE = os.path.dirname(os.path.abspath(__file__))
+DEVICE = torch.device("cpu")
+
+print("Loading vocabularies...")
+SRC_VOCAB = Vocabulary.load(os.path.join(BASE, "data", "src_vocab.pkl"))
+TGT_VOCAB = Vocabulary.load(os.path.join(BASE, "data", "tgt_vocab.pkl"))
+
+print("Loading model...")
+MODEL = build_model(
+    src_vocab_size=len(SRC_VOCAB),
+    tgt_vocab_size=len(TGT_VOCAB),
+    embed_dim=256, hidden_dim=256, attention_dim=128,
+    n_layers=2, dropout=0.3,
+    pad_idx=Vocabulary.PAD_IDX, sos_idx=Vocabulary.SOS_IDX,
+    device=DEVICE,
+)
+ckpt = torch.load(os.path.join(BASE, "models", "best_model.pt"),
+                  map_location=DEVICE, weights_only=False)
+MODEL.load_state_dict(ckpt["model_state_dict"])
+MODEL.eval()
+print(f"Loaded checkpoint from epoch {ckpt.get('epoch')} "
+      f"(val_loss: {ckpt.get('val_loss'):.4f})")
+
+
+# ----------------------------------------------------------------------
+# Helpers
+# ----------------------------------------------------------------------
+def attention_heatmap(source_text, output_tokens, attn_matrix):
+    """Render attention weights as a matplotlib figure."""
+    src_tokens = tokenize(source_text)[:attn_matrix.shape[1] - 2]  # trim special
+    out_tokens = output_tokens[:attn_matrix.shape[0]]
+
+    # Trim attention matrix to match displayed tokens
+    attn = attn_matrix[:len(out_tokens), :len(src_tokens) + 2]
+
+    fig, ax = plt.subplots(figsize=(max(8, len(src_tokens) * 0.25),
+                                     max(4, len(out_tokens) * 0.35)))
+    im = ax.imshow(attn, cmap="viridis", aspect="auto")
+    ax.set_xticks(range(len(src_tokens) + 2))
+    ax.set_xticklabels(["<SOS>"] + src_tokens + ["<EOS>"],
+                       rotation=75, fontsize=8)
+    ax.set_yticks(range(len(out_tokens)))
+    ax.set_yticklabels(out_tokens, fontsize=9)
+    ax.set_xlabel("Source Tokens")
+    ax.set_ylabel("Generated Tokens")
+    ax.set_title("Bahdanau Attention Weights")
+    plt.colorbar(im, ax=ax, fraction=0.025)
+    plt.tight_layout()
+    return fig
+
+
+def generate(source_text, method, beam_width, max_len):
+    """Main inference function called by Gradio."""
+    if not source_text.strip():
+        return "Please enter source text.", None, ""
+
+    try:
+        if method == "Greedy":
+            text, meta = generate_document(
+                MODEL, source_text, SRC_VOCAB, TGT_VOCAB,
+                method="greedy", max_len=int(max_len), device=DEVICE
+            )
+            # Build attention heatmap
+            fig = None
+            attn = meta.get("attention")
+            if attn is not None and hasattr(attn, "shape"):
+                out_tokens = text.split()
+                try:
+                    fig = attention_heatmap(source_text, out_tokens,
+                                            attn.cpu().numpy() if torch.is_tensor(attn) else np.asarray(attn))
+                except Exception as e:
+                    print(f"Heatmap error: {e}")
+            info = f"Method: Greedy decode | Output length: {len(text.split())} tokens"
+            return text, fig, info
+
+        else:  # Beam
+            text, meta = generate_document(
+                MODEL, source_text, SRC_VOCAB, TGT_VOCAB,
+                method="beam", beam_width=int(beam_width),
+                max_len=int(max_len), device=DEVICE
+            )
+            info = (f"Method: Beam Search (width={int(beam_width)}) | "
+                    f"Score: {meta.get('score', 0):.4f} | "
+                    f"Output length: {len(text.split())} tokens")
+            return text, None, info
+
+    except Exception as e:
+        return f"Error: {e}", None, ""
+
+
+# ----------------------------------------------------------------------
+# Gradio UI
+# ----------------------------------------------------------------------
+EXAMPLES = [
+    ["The quarterly financial report for TechNova indicates revenue of $2500M, "
+     "representing a 15% increase year over year. Operating expenses increased "
+     "to $1200M. Net income was $450M. The board approved a dividend of $2.50 "
+     "per share. Management projects continued growth in the coming quarters "
+     "driven by AI integration.", "Greedy", 5, 60],
+
+    ["The ProMax X1 by CloudPeak features a 8-core processor, 6000mAh battery, "
+     "and AI-powered assistant. It is designed for professionals who need high "
+     "performance computing. Available in Black, Silver, and Blue, the device "
+     "weighs 195g and includes fast charging and biometric auth. Pricing starts "
+     "at $999.", "Beam", 5, 60],
+
+    ["This study examines the relationship between remote work frequency and "
+     "productivity using a dataset of 5000 observations from Fortune 500 "
+     "companies. We employ regression analysis to analyze temporal patterns. "
+     "Results indicate a strong positive correlation (p < 0.001). The findings "
+     "suggest that targeted interventions improve outcomes.", "Greedy", 5, 60],
+]
+
+DESCRIPTION = """
+# Seq2Seq Document Generation with Bahdanau Attention
+
+Encoder-decoder model that compresses long-form documents (financial reports,
+product specs, research abstracts) into concise summaries.
+
+- **Architecture:** Bidirectional GRU Encoder + Bahdanau (Additive) Attention + GRU Decoder
+- **Parameters:** 3.9M | **Framework:** PyTorch
+- **Training:** 15 epochs on 5,000 synthetic pairs | **Best Val PPL:** 9.08
+- **Decoding:** Greedy + Beam Search (width 5)
+
+Paste a document below, pick a decoding strategy, and see the model
+generate a summary. Greedy mode also renders the **attention heatmap**
+showing which source tokens the decoder focused on at each step.
+"""
+
+with gr.Blocks(title="Seq2Seq Doc Generation") as demo:
+    gr.Markdown(DESCRIPTION)
+
+    with gr.Row():
+        with gr.Column(scale=3):
+            src = gr.Textbox(
+                label="Source Document",
+                placeholder="Paste a long document here...",
+                lines=8,
+            )
+            with gr.Row():
+                method = gr.Radio(
+                    ["Greedy", "Beam"], value="Greedy",
+                    label="Decoding Method"
+                )
+                beam_width = gr.Slider(2, 10, value=5, step=1,
+                                       label="Beam Width")
+                max_len = gr.Slider(20, 120, value=60, step=5,
+                                    label="Max Output Tokens")
+            btn = gr.Button("Generate", variant="primary")
+
+        with gr.Column(scale=2):
+            output = gr.Textbox(label="Generated Summary", lines=4)
+            info = gr.Textbox(label="Decoding Info", lines=1)
+            heatmap = gr.Plot(label="Attention Heatmap (Greedy only)")
+
+    gr.Examples(
+        examples=EXAMPLES,
+        inputs=[src, method, beam_width, max_len],
+        outputs=[output, heatmap, info],
+        fn=generate,
+        cache_examples=False,
+    )
+
+    btn.click(generate,
+              inputs=[src, method, beam_width, max_len],
+              outputs=[output, heatmap, info])
+
+    gr.Markdown(
+        "---\n"
+        "**Repo:** [github.com/Reethika30/nlp-seq2seq-docgen]"
+        "(https://github.com/Reethika30/nlp-seq2seq-docgen)"
+    )
+
+
+if __name__ == "__main__":
+    demo.launch(server_name="0.0.0.0", share=False)