Upload config.yaml with huggingface_hub

config.yaml

@@ -0,0 +1,43 @@
+# Configuration file for SAM-Audio model training
+# Optimized for RTX 4090 (24GB VRAM)
+
+# Training parameters
+num_epochs: 50
+batch_size: 32                    # Increased for 4090's 24GB VRAM
+learning_rate: 3e-4               # Slightly higher LR works well with larger batches
+random_seed: 42
+weight_decay: 0.01
+max_grad_norm: 1.0                # Gradient clipping for stability
+gradient_accumulation_steps: 1    # Increase if OOM (effective batch = batch_size * this)
+warmup_epochs: 2                  # Learning rate warmup
+patience: 10                      # Early stopping patience
+
+# Data parameters
+mask_ratio: 0.15
+temperature: 0.07
+num_segments: 64
+
+# Loss weights
+mask_weight: 0.5
+contrastive_weight: 0.3
+
+# Encoder parameters
+encoder_dims: [64, 128, 256, 512]
+kernel_size: 3
+stride: 2
+dropout_rate: 0.1
+use_layer_norm: true              # More stable than BatchNorm for audio
+
+# RTX 4090 optimizations
+compile_model: true               # torch.compile() for faster execution
+num_workers: 4                    # DataLoader workers (adjust based on CPU cores)
+
+# Logging and saving
+wandb_project: "sam-carnatic"
+run_name: "sam-audio-rtx4090"
+best_model_path: "best_model.pth"
+final_model_path: "final_model.pth"
+
+# Model pushing
+push_to_hub: true
+model_name: "sarayusapa/sam-carnatic"