app.py

import streamlit as st
import numpy as np
import pickle
import tensorflow as tf
import keras
from tensorflow.keras.models import load_model
from tensorflow.keras.preprocessing.sequence import pad_sequences
from tensorflow.keras.layers import GRU, Bidirectional, Embedding, Dense, TimeDistributed
import pandas as pd

# ==========================================
# 🔧 HOTFIX: Keras 3 Compatibility Patch
# ==========================================
# We define a custom GRU that ignores 'time_major' and other old arguments
@keras.saving.register_keras_serializable(package="Complex")
class CompatibleGRU(GRU):
    def __init__(self, *args, **kwargs):
        # Remove arguments that Keras 3 no longer supports
        if 'time_major' in kwargs:
            kwargs.pop('time_major')
        if 'implementation' in kwargs:
            kwargs.pop('implementation')
        if 'reset_after' in kwargs: 
            # reset_after is supported but sometimes causes conflicts in restoration
            # We keep it unless it causes issues, but usually time_major is the culprit.
            pass 
        super().__init__(*args, **kwargs)

# We map the model's "GRU" layer to our safe "CompatibleGRU"
custom_objects = {
    'GRU': CompatibleGRU,
    'Bidirectional': Bidirectional,
    'Embedding': Embedding,
    'Dense': Dense,
    'TimeDistributed': TimeDistributed
}

# ==========================================
# 📥 Data Loading
# ==========================================
@st.cache_resource
def load_assets():
    # Attempt to load with our patched custom objects
    try:
        model = load_model("bigru_pos_tagger.h5", custom_objects=custom_objects, compile=False)
    except Exception as e:
        st.error(f"Standard load failed: {e}. Trying legacy method...")
        # Fallback to legacy loading if the direct load fails
        model = keras.src.legacy.saving.legacy_h5_format.load_model_from_hdf5(
            "bigru_pos_tagger.h5", custom_objects=custom_objects
        )

    with open("word2idx_bigru.pkl", "rb") as f:
        word2idx = pickle.load(f)

    with open("idx2tag_bigru.pkl", "rb") as f:
        idx2tag = pickle.load(f)

    return model, word2idx, idx2tag

# Initialize global variables
try:
    model, word2idx, idx2tag = load_assets()
except Exception as e:
    st.error(f"Critical Error Loading Model: {e}")
    st.stop()

MAX_LEN = 134

# ==========================================
# 🧠 Processing Functions
# ==========================================
def pos_tag_sentence(sentence):
    tokens = sentence.split()
    if not tokens:
        return []
    
    # word2idx is accessible here
    token_ids = [word2idx.get(token, 1) for token in tokens]
    token_ids_padded = pad_sequences([token_ids], maxlen=MAX_LEN, padding="post")
    
    # verbose=0 prevents console spam
    predictions = model.predict(token_ids_padded, verbose=0)[0]
    predicted_tags = [idx2tag[np.argmax(tag)] for tag in predictions][:len(tokens)]
    return list(zip(tokens, predicted_tags))

def format_tagged_line(tagged_list):
    return " ".join([f"{word}\\{tag}" for word, tag in tagged_list])

# ==========================================
# 🖥️ Streamlit UI
# ==========================================
st.set_page_config(page_title="Konkani POS Tagger", layout="wide")
st.title("Automatic Konkani POS Tagger")

tab1, tab2 = st.tabs(["Single Sentence", "Batch File Upload"])

# --- TAB 1: Single Sentence ---
with tab1:
    user_input = st.text_area("Enter a Konkani sentence:", height=100, key="single_input")
    if st.button("Tag Sentence"):
        if user_input.strip():
            tagged = pos_tag_sentence(user_input.strip())
            st.subheader("Tagged Result:")
            st.text(format_tagged_line(tagged))
            st.table(pd.DataFrame(tagged, columns=["Word", "Tag"]))
        else:
            st.warning("Please enter text first.")

# --- TAB 2: Batch File ---
with tab2:
    st.subheader("Upload Text File")
    uploaded_file = st.file_uploader("Upload a .txt file (One sentence per line)", type=["txt"])
    
    if uploaded_file is not None:
        if st.button("Process and Download"):
            # Read input file
            input_text = uploaded_file.getvalue().decode("utf-8")
            lines = input_text.splitlines()
            
            output_lines = []
            progress_bar = st.progress(0)
            status_text = st.empty()
            
            for i, line in enumerate(lines):
                if line.strip():
                    tagged = pos_tag_sentence(line.strip())
                    output_lines.append(format_tagged_line(tagged))
                else:
                    output_lines.append("") 
                
                # Update progress
                progress = (i + 1) / len(lines)
                progress_bar.progress(progress)
                status_text.text(f"Processing line {i+1} of {len(lines)}...")
            
            result_text = "\n".join(output_lines)
            status_text.text("Processing complete!")
            
            st.success(f"Successfully processed {len(lines)} lines!")
            st.download_button(
                label="📥 Download Tagged File",
                data=result_text,
                file_name="tagged_konkani_output.txt",
                mime="text/plain"
            )