app.py

# ==========================================
# Mumbai Bird Call Identifier — FINAL FIX
# ==========================================

import streamlit as st
import torch
import torchaudio
import numpy as np
import io
import matplotlib.cm as cm
from torchvision import models, transforms
from PIL import Image

# ================== PAGE CONFIG ==================
st.set_page_config(
    page_title="Mumbai Bird Call Identifier",
    page_icon="🐦",
    layout="centered"
)

# ================== CONSTANTS ==================
SAMPLE_RATE = 22050
DURATION = 5
TARGET_SAMPLES = SAMPLE_RATE * DURATION

HIGH_CONF = 0.60
MEDIUM_CONF = 0.35

# ⚠️ MUST MATCH TRAINING
SPECTROGRAM_COLORMAP = "magma"   # ← change ONLY if training used something else

# ================== LOAD MODEL ==================
@st.cache_resource
def load_model():
    checkpoint = torch.load("multi_species_model.pth", map_location="cpu")

    label_map = checkpoint["label_map"]
    index_to_species = {v: k for k, v in label_map.items()}
    class_names = [index_to_species[i] for i in range(len(index_to_species))]

    model = models.mobilenet_v3_small(pretrained=False)
    model.classifier[3] = torch.nn.Linear(
        model.classifier[3].in_features,
        len(class_names)
    )

    model.load_state_dict(checkpoint["model_state_dict"])
    model.eval()

    return model, class_names


model, class_names = load_model()

# ================== AUDIO TRANSFORMS ==================
mel_transform = torchaudio.transforms.MelSpectrogram(
    sample_rate=SAMPLE_RATE,
    n_fft=512,
    win_length=512,
    hop_length=256,
    f_min=50,
    f_max=11000,
    n_mels=128,
    norm="slaney",
    mel_scale="slaney"
)

db_transform = torchaudio.transforms.AmplitudeToDB(
    stype="power",
    top_db=80
)

# ================== IMAGE TRANSFORM (MATCH TRAINING) ==================
val_transform = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406],
        std=[0.229, 0.224, 0.225]
    )
])

# ================== UI ==================
st.title("🐦 Mumbai Balcony Bird Call Identifier")
st.markdown(
    """
Identify **204 Indian bird species** from their calls.  
Model trained on **PNG spectrogram images** from Mumbai & Maharashtra.
"""
)

audio_file = st.file_uploader(
    "Upload bird call audio (WAV / MP3 / M4A / OGG)",
    type=["wav", "mp3", "m4a", "ogg"]
)

# ================== PROCESS ==================
if audio_file:
    st.audio(audio_file)

    with st.spinner("🧠 Analyzing bird call..."):
        # -------- LOAD AUDIO --------
        waveform, sr = torchaudio.load(io.BytesIO(audio_file.read()))

        if waveform.shape[0] > 1:
            waveform = waveform.mean(dim=0, keepdim=True)

        if sr != SAMPLE_RATE:
            waveform = torchaudio.functional.resample(
                waveform, sr, SAMPLE_RATE
            )

        # -------- ENERGY-BASED CROP --------
        energy = waveform.abs()
        peak = energy.argmax()
        start = max(0, peak - TARGET_SAMPLES // 2)
        end = start + TARGET_SAMPLES
        waveform = waveform[:, start:end]

        if waveform.shape[1] < TARGET_SAMPLES:
            waveform = torch.nn.functional.pad(
                waveform,
                (0, TARGET_SAMPLES - waveform.shape[1])
            )

        # -------- MEL SPECTROGRAM --------
        mel = mel_transform(waveform)
        mel = db_transform(mel)
        mel = mel.squeeze(0)

        # -------- MEL → COLORED PNG (CRITICAL FIX) --------
        mel_np = mel.numpy()
        mel_norm = (mel_np - mel_np.min()) / (mel_np.ptp() + 1e-8)

        cmap = cm.get_cmap(SPECTROGRAM_COLORMAP)
        colored = cmap(mel_norm)[:, :, :3]  # drop alpha
        mel_img = (colored * 255).astype(np.uint8)

        mel_pil = Image.fromarray(mel_img)

        # -------- MODEL INPUT --------
        model_input = val_transform(mel_pil).unsqueeze(0)

        # -------- INFERENCE --------
        with torch.no_grad():
            logits = model(model_input)
            probs = torch.softmax(logits[0], dim=0)

        top5_probs, top5_idx = torch.topk(probs, 5)

    # ================== RESULTS ==================
    st.markdown("---")

    top1_prob = top5_probs[0].item()
    top1_species = class_names[top5_idx[0]]

    if top1_prob >= HIGH_CONF:
        st.success("✅ High confidence identification")
    elif top1_prob >= MEDIUM_CONF:
        st.warning("⚠️ Medium confidence identification")
    else:
        st.error("❓ Low confidence – possibly unknown species")

    st.markdown(f"## 🐦 {top1_species}")
    st.metric("Confidence", f"{top1_prob*100:.1f}%")

    st.markdown("### 🔍 Other possible matches")
    for i in range(1, 5):
        st.markdown(
            f"- **{class_names[top5_idx[i]]}** — {top5_probs[i].item():.1%}"
        )

    st.markdown("---")
    st.subheader("📊 Spectrogram used by the model")
    st.image(mel_img, use_container_width=True)

else:
    st.info("👆 Upload a bird call audio file to begin")

# ================== FOOTER ==================
st.markdown("---")
st.caption(
    "⚠️ This model predicts among known species only. "
    "Low confidence may indicate an unseen species or noisy audio."
)