app.py

import gradio as gr
import torch
import torch.nn.functional as F
import librosa
import soundfile as sf
import numpy as np
import zipfile
from huggingface_hub import hf_hub_download

# ================= MODEL (unchanged) =================
try:
    from bs_roformer import BSRoformer
    from attend import Attend
except ImportError:
    pass

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

def safe_attend_forward(self, q, k, v, mask=None):
    return F.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0)

try:
    Attend.forward = safe_attend_forward
except Exception:
    pass

def load_model():
    ckpt = hf_hub_download(
        repo_id="Tachyeon/IAM-RoFormer-Model-Weights",
        filename="v11_consensus_epoch_30.pt"
    )
    model = BSRoformer(
        dim=512, depth=12, stereo=True, num_stems=4,
        time_transformer_depth=1, freq_transformer_depth=1,
        flash_attn=True
    ).to(DEVICE)

    state = torch.load(ckpt, map_location=DEVICE)
    model.load_state_dict(state["model"] if "model" in state else state)
    model.eval()
    return model

model = load_model()

# ================= SEPARATION + ZIP =================
def separate_audio(path):
    if not path:
        return [None]*5

    mix, sr = librosa.load(path, sr=44100, mono=False)
    if mix.ndim == 1:
        mix = np.stack([mix, mix])

    x = torch.tensor(mix).float().to(DEVICE)[None]
    L = x.shape[-1]

    out = torch.zeros(1,4,2,L, device=DEVICE)
    cnt = torch.zeros_like(out)

    chunk = 44100*10
    hop = chunk - 44100

    with torch.no_grad(), torch.autocast("cuda", enabled=torch.cuda.is_available()):
        for s in range(0, L, hop):
            e = min(s+chunk, L)
            part = x[:,:,s:e]
            if part.shape[-1] < chunk:
                part = F.pad(part,(0,chunk-part.shape[-1]))
            pred = model(part)
            out[:,:,:,s:e] += pred[:,:,:,:e-s]
            cnt[:,:,:,s:e] += 1

    stems = (out / cnt.clamp(min=1)).cpu().numpy()[0]

    names = [
        "lead_vocals.wav",
        "mridangam_percussion.wav",
        "tanpura_drone.wav",
        "violin_accompaniment.wav"
    ]

    for i, name in enumerate(names):
        sf.write(name, stems[i].T, sr)

    zip_path = "stems.zip"
    with zipfile.ZipFile(zip_path, "w", zipfile.ZIP_DEFLATED) as z:
        for name in names:
            z.write(name)

    # return stems + zip + signal to show download button
    return [names[0], names[1], names[2], names[3], zip_path, gr.update(visible=True)]

# ================= UI (VISUALLY UNCHANGED) =================
css = """
@import url('https://fonts.googleapis.com/css2?family=Anton&family=Poppins:wght@400;500;600&display=swap');
:root{
  --bg1:#2a141d;
  --bg2:#14080d;
  --ink:#f3ece6;
  --muted:#b6aeb0;
  --accent:#ff6f9f;
  --panel: rgba(255,255,255,0.03);
  --panel-2: rgba(255,255,255,0.02);
  --radius: 12px;
}
html, body, .gradio-container {
  height: 100%;
  margin: 0;
  background: linear-gradient(180deg, var(--bg1), var(--bg2)) !important;
  color: var(--ink);
  font-family: Poppins, sans-serif;
}
.app {
  max-width: 1160px;
  margin: 0 auto;
  padding: 48px 40px;
  display: grid;
  grid-template-rows: auto 1fr;
  gap: 36px;
}
.brand { display:flex; flex-direction:column; gap:6px; }
.logo { font-family: Anton, sans-serif; font-size:46px; }
.tagline { font-size:14px; color:var(--accent); opacity:0.9; }
.main {
  display:grid;
  grid-template-columns: 1fr 420px;
  gap: 48px;
  align-items: start;
}
.left h3 { margin: 0; font-size:18px; font-weight:600; }
.left p { margin:6px 0 18px; color:var(--muted); font-size:13px; }
.left .gradio-audio {
  background: var(--panel) !important;
  border-radius: var(--radius);
  min-height: 260px;
  display:flex;
  align-items:center;
  justify-content:center;
}
.button-primary {
  margin-top: 18px;
  height:46px;
  width:100%;
  font-size:15px !important;
  font-weight:600 !important;
  background: linear-gradient(90deg,#ff6f9f,#ffbf7a) !important;
  color: #14080d !important;
  border-radius: 10px !important;
}
.stems {
  display:grid;
  grid-template-columns: 1fr 1fr;
  gap: 22px;
}
.stem-surface {
  background: var(--panel-2);
  border-radius: 14px;
  padding: 12px;
  min-height: 140px;
  display:flex;
  flex-direction:column;
  justify-content:center;
  gap:6px;
}
.stem-label {
  font-size:13px;
  font-weight:600;
  color: var(--accent);
}
.stem-info {
  font-size:11px;
  color:var(--muted);
  opacity:0.85;
}
.stem-surface .gradio-audio label { display:none !important; }
.stem-surface audio { width:92%; max-height:36px; }
"""

with gr.Blocks() as demo:
    with gr.Column(elem_classes="app"):

        with gr.Row(elem_classes="brand"):
            gr.HTML('<div class="logo">SWARA STUDIO</div>')
            gr.HTML('<div class="tagline">Separating Music Into Its elements</div>')

        with gr.Row(elem_classes="main"):
            # LEFT
            with gr.Column(elem_classes="left"):
                gr.HTML("""
                    <h3>Select a track</h3>
                    <p>We’ll break it down into individual parts</p>
                """)
                input_audio = gr.Audio(type="filepath")
                run_btn = gr.Button("Separate", elem_classes="button-primary")

                ### CHANGED: clean download action
                download_btn = gr.Button("Download all stems", visible=False)
                zip_out = gr.File(visible=False)

            # RIGHT
            with gr.Column():
                with gr.Row(elem_classes="stems"):

                    with gr.Column(elem_classes="stem-surface"):
                        gr.HTML("""
                            <div class="stem-label">Lead Vocals</div>
                            <div class="stem-info">Primary melodic voice and lyrical content</div>
                        """)
                        out_vocals = gr.Audio(interactive=False)

                    with gr.Column(elem_classes="stem-surface"):
                        gr.HTML("""
                            <div class="stem-label">Mridangam / Percussion</div>
                            <div class="stem-info">Rhythmic transients and percussive articulation</div>
                        """)
                        out_drums = gr.Audio(interactive=False)

                    with gr.Column(elem_classes="stem-surface"):
                        gr.HTML("""
                            <div class="stem-label">Tanpura / Drone</div>
                            <div class="stem-info">Sustained harmonic bed and tonal reference</div>
                        """)
                        out_bass = gr.Audio(interactive=False)

                    with gr.Column(elem_classes="stem-surface"):
                        gr.HTML("""
                            <div class="stem-label">Violin / Accompaniment</div>
                            <div class="stem-info">Melodic support and expressive ornamentation</div>
                        """)
                        out_other = gr.Audio(interactive=False)

    run_btn.click(
        separate_audio,
        input_audio,
        [out_vocals, out_drums, out_bass, out_other, zip_out, download_btn]
    )

    download_btn.click(lambda z: z, zip_out, zip_out)

if __name__ == "__main__":
    demo.launch(css=css, theme=gr.themes.Base())