model_checkpoints/Primitive_Extractor/clotho/training_script.md

Primitive Extractor - Clotho Reconstruction

Overview

This checkpoint represents the Clotho Reconstruction fine-tuning of the WindOne Sound V3 Primitive Extractor.

Starting from the proven ESC-50 Baseline weights, this module was trained as an Autoencoder on the Clotho dataset. A temporary linear decoder (512 -> 1024 -> 768) was attached to the end of the Primitive Extractor.

Why Reconstruction Loss?

We deliberately chose L1 Reconstruction Loss instead of Contrastive Loss (like CLIP) for this phase.

• Acoustic Fidelity: The downstream VoiceHead requires exact acoustic textures to generate clean mel-spectrograms. Contrastive loss forces the 512-dim space to prioritize text semantics, which can distort fine-grained audio details.
• Separation of Concerns: Reconstruction loss guarantees the 512-dim primitives retain the exact acoustic information of the 768-dim BEATs features. Semantic and temporal alignment is handled later by the CausalGraphNetwork and COGENMemory.

Training Configuration

• Base Model: PrimitiveExtractor (ESC-50 Baseline) + Temporary Decoder
• Dataset: Clotho Development Set (3,631 audio files)
• Train/Val Split: 2,904 / 727
• Input: 768-dim BEATs features
• Output: 768-dim Reconstructed features (via 512-dim bottleneck)
• Epochs: 8
• Batch Size: 8
• Optimizer: AdamW (lr=1e-4, weight_decay=1e-2)
• Loss: L1 Reconstruction Loss
• Checkpoint Metric: Validation Reconstruction Loss (Lower is better)

Results Summary

• Best Val Loss: 0.1251 (Epoch 8)
• Best Val Similarity: 91.94% (Cosine Similarity between original and reconstructed 768-dim vectors)
• Verdict: The 512-dim primitives successfully retain ~92% of the acoustic information from the original BEATs features.

Training Script

import os
import sys
import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
from huggingface_hub import hf_hub_download, HfApi
from google.colab import drive
from tqdm.auto import tqdm

# ==========================================
# 1. SETUP & PATHS
# ==========================================
drive.mount('/content/drive', force_remount=False)

CLOTHO_BEATS_768_CACHE = '/content/clotho_beats_768_cache' 
repo_id = "Hitelcy/WindoneRepo2"

sys.path.append('/content/drive/MyDrive/WindOneSound_300M')
from model_v3.primitive_extractor import PrimitiveExtractor

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

# ==========================================
# 2. LOAD ESC-50 BASELINE WEIGHTS
# ==========================================
print("📥 Downloading ESC-50 Baseline Primitive Extractor from Hugging Face...")
esc50_ckpt_path = hf_hub_download(
    repo_id=repo_id,
    filename="model_checkpoints/Primitive_Extractor/esc50/primitive_extractor_esc50_best.pt",
    repo_type="dataset",
    local_dir="/tmp/hf_ckpt"
)

# ==========================================
# 3. DEFINE AUTOENCODER MODEL
# ==========================================
class PrimitiveAutoEncoder(nn.Module):
    def __init__(self):
        super().__init__()
        self.extractor = PrimitiveExtractor(input_dim=768, hidden_dim=1728, output_dim=512)
        self.decoder = nn.Sequential(
            nn.Linear(512, 1024),
            nn.GELU(),
            nn.Linear(1024, 768)
        )

    def forward(self, x):
        primitives = self.extractor(x)
        reconstructed = self.decoder(primitives)
        return reconstructed

model = PrimitiveAutoEncoder()

esc50_state = torch.load(esc50_ckpt_path, map_location='cpu', weights_only=False)
original_state_dict = esc50_state['model_state_dict']

# Strip 'extractor.' prefix to load into bare PrimitiveExtractor
extractor_state_dict = {}
for k, v in original_state_dict.items():
    if k.startswith('extractor.'):
        new_key = k.replace('extractor.', '')
        extractor_state_dict[new_key] = v
        
model.extractor.load_state_dict(extractor_state_dict)
print("✅ ESC-50 weights successfully loaded into Extractor.")

model = model.to(device)

extractor_params = sum(p.numel() for p in model.extractor.parameters() if p.requires_grad)
decoder_params = sum(p.numel() for p in model.decoder.parameters() if p.requires_grad)
print(f" Params -> Extractor: {extractor_params:,} | Decoder: {decoder_params:,} | Total: {extractor_params + decoder_params:,}\n")

# ==========================================
# 4. DATASET & DATALOADER
# ==========================================
class ClothoBeatsDataset(torch.utils.data.Dataset):
    def __init__(self, cache_dir, split='train', train_ratio=0.8):
        self.cache_files = sorted([f for f in os.listdir(cache_dir) if f.endswith('.pt')])
        split_idx = int(len(self.cache_files) * train_ratio)
        self.files = self.cache_files[:split_idx] if split == 'train' else self.cache_files[split_idx:]

    def __len__(self):
        return len(self.files)

    def __getitem__(self, idx):
        return torch.load(os.path.join(CLOTHO_BEATS_768_CACHE, self.files[idx]), weights_only=True)

train_dataset = ClothoBeatsDataset(CLOTHO_BEATS_768_CACHE, split='train')
val_dataset = ClothoBeatsDataset(CLOTHO_BEATS_768_CACHE, split='val')

train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=8, shuffle=True, num_workers=2)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=8, shuffle=False, num_workers=2)

print(f"✅ Train size: {len(train_dataset)}, Val size: {len(val_dataset)}")

# ==========================================
# 5. TRAINING LOOP
# ==========================================
criterion = nn.L1Loss() 
optimizer = optim.AdamW(model.parameters(), lr=1e-4, weight_decay=1e-2)

num_epochs = 8
best_val_loss = float('inf')
api = HfApi()

print("\n🚀 Starting Clotho Reconstruction Training (Max 8 Epochs)...")

for epoch in range(1, num_epochs + 1):
    model.train()
    train_losses = []
    
    for features in tqdm(train_loader, desc=f"Epoch {epoch}/8 Train"):
        features = features.to(device) 
        
        optimizer.zero_grad()
        reconstructed = model(features)
        loss = criterion(reconstructed, features)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)
        optimizer.step()
        
        train_losses.append(loss.item())

    model.eval()
    val_losses = []
    val_similarities = [] 
    
    with torch.no_grad():
        for features in tqdm(val_loader, desc=f"Epoch {epoch}/8 Val"):
            features = features.to(device)
            reconstructed = model(features)
            
            loss = criterion(reconstructed, features)
            val_losses.append(loss.item())
            
            B, T, D = features.shape
            flat_orig = features.view(B * T, D)
            flat_recon = reconstructed.view(B * T, D)
            
            sim = F.cosine_similarity(flat_orig, flat_recon, dim=1).mean().item()
            val_similarities.append(sim)

    avg_train_loss = sum(train_losses) / len(train_losses)
    avg_val_loss = sum(val_losses) / len(val_losses)
    avg_val_sim = (sum(val_similarities) / len(val_similarities)) * 100 
    
    print(f"\nEpoch {epoch}/{num_epochs} | Train Loss: {avg_train_loss:.4f} | Val Loss: {avg_val_loss:.4f} | Val Similarity: {avg_val_sim:.2f}%\n")

    if avg_val_loss < best_val_loss:
        best_val_loss = avg_val_loss
        print(f" New best Val Loss! Saving checkpoint to Hugging Face...")
        
        save_path = "model_checkpoints/Primitive_Extractor/clotho/primitive_extractor_clotho_recon_best.pt"
        
        torch.save({
            'epoch': epoch,
            'model_state_dict': model.state_dict(), 
            'val_recon_loss': avg_val_loss,
            'val_similarity_percent': avg_val_sim,
            'train_recon_loss': avg_train_loss
        }, '/tmp/primitive_clotho_best.pt')
        
        try:
            api.upload_file(
                path_or_fileobj='/tmp/primitive_clotho_best.pt',
                path_in_repo=save_path,
                repo_id=repo_id,
                repo_type="dataset"
            )
            print(f"✅ Successfully uploaded checkpoint to {repo_id}/{save_path}")
        except Exception as e:
            print(f"❌ HF Upload failed: {e}")

print("\n Clotho Reconstruction Training complete!")
print(f"Best Validation Loss: {best_val_loss:.4f}")