Upload app.py

app.py

@@ -0,0 +1,65 @@
+# app.py
+import streamlit as st
+import numpy as np
+import matplotlib.pyplot as plt
+from model.inference import TTSInference
+
+# Page Config
+st.set_page_config(page_title="My Custom TTS Engine", layout="wide")
+
+st.title("🎙️ Custom Architecture TTS Playground")
+st.markdown("This project demonstrates a custom PyTorch implementation of a Transformer-based TTS.")
+
+# Sidebar for Model Controls
+with st.sidebar:
+    st.header("Model Settings")
+    checkpoint = st.selectbox("Select Checkpoint", ["checkpoints/checkpoint_epoch_50c.pth", "checkpoints/checkpoint_epoch_3c.pth", "checkpoints/checkpoint_epoch_8.pth"])
+    device = st.radio("Device", ["cpu", "cuda"])
+    st.info("Load a specific training checkpoint to compare progress.")
+
+# Initialize the Inference Engine
+# (In a real app, use @st.cache_resource to load this once)
+tts_engine = TTSInference(checkpoint_path=checkpoint, device=device)
+
+# Main Input Area
+text_input = st.text_area("Enter Text to Speak:", "Deep learning is fascinating.", height=100)
+
+col1, col2 = st.columns([1, 2])
+
+with col1:
+    if st.button("Generate Audio", type="primary"):
+        with st.spinner("Running Inference..."):
+            # Call your backend
+            audio_data, sample_rate, mel_spec = tts_engine.predict(text_input)
+            
+            # Play Audio
+            st.success("Generation Complete!")
+            st.audio(audio_data, sample_rate=sample_rate)
+
+            # --- VISUALIZATION (Crucial for Path 2) ---
+            # Showing the spectrogram proves you understand the data, not just the result.
+            st.subheader("Mel Spectrogram Analysis")
+            fig, ax = plt.subplots(figsize=(10, 3))
+            im = ax.imshow(mel_spec, aspect='auto', origin='lower', cmap='inferno')
+            plt.colorbar(im, ax=ax)
+            plt.title("Generated Mel Spectrogram")
+            plt.xlabel("Time Frames")
+            plt.ylabel("Mel Channels")
+            st.pyplot(fig)
+
+with col2:
+    st.subheader("Architecture Details")
+    st.code("""
+    
+    class TextToMel(nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.encoder = TransformerEncoder(...)
+            self.decoder = TransformerDecoder(...)
+            
+        def forward(self, text):
+            # 1. Embed text
+            # 2. Add Positional Encodings
+            # 3. Predict Mel Frames
+            return mel_output
+    """, language="python")