⚡ ULTRON: Sync config.py

config.py

@@ -0,0 +1,69 @@
+"""
+⚡ ULTRON Systems // OHLCV Transformer Configuration Matrix
+Central configuration for data streaming, feature extraction, model architecture, and Colab GPU training.
+"""
+from dataclasses import dataclass, field
+from typing import List, Optional
+import os
+import torch
+
+@dataclass
+class DatasetConfig:
+    repo_id: str = "mito0o852/OHLCV-1m"
+    # Choose date ranges to process. Each file is data/ohlcv_YYYY-MM.parquet
+    start_year: int = 2020
+    end_year: int = 2025
+    
+    # Target top liquid tickers to maximize signal and limit noise (leave empty for all tickers)
+    # Default liquid basket: Mega-cap tech, ETFs, high volume liquid equities
+    liquid_tickers: List[str] = field(default_factory=lambda: [
+        "AAPL", "MSFT", "NVDA", "AMZN", "GOOGL", "META", "TSLA",
+        "SPY", "QQQ", "IWM", "AMD", "NFLX", "INTC", "JPM", "BAC"
+    ])
+    
+    # Sequence length in minutes (e.g., 512 minutes ~ 8.5 trading hours)
+    seq_len: int = 512
+    
+    # Overlapping window stride for dataset generation (e.g. step by 16 minutes)
+    window_stride: int = 16
+    
+    # Future prediction horizons in minutes: [1m, 3m, 5m, 10m, 15m]
+    pred_horizons: List[int] = field(default_factory=lambda: [1, 3, 5, 10, 15])
+    
+    # Local ephemeral storage path for downloading 1 parquet file at a time
+    cache_dir: str = "/tmp/ohlcv_cache"
+
+@dataclass
+class ModelConfig:
+    in_features: int = 8             # Number of stationary features generated per bar
+    d_model: int = 128               # Transformer latent dimension
+    n_heads: int = 4                 # Multi-head attention heads
+    n_layers: int = 4                # Number of Conformer/Transformer layers
+    d_ff: int = 256                  # Feed-forward expansion dimension
+    conv_kernels: List[int] = field(default_factory=lambda: [3, 7, 15, 31]) # Multi-scale conv kernels
+    patch_stride: int = 4            # Stride of overlapping 1D conv tokenizer
+    conv_depthwise_kernel: int = 7   # Conformer depthwise conv kernel
+    dropout: float = 0.1
+    use_revin: bool = True           # Reversible Instance Normalization
+    num_horizons: int = 5            # Matching len(DatasetConfig.pred_horizons)
+
+@dataclass
+class TrainingConfig:
+    batch_size: int = 64             # Physical batch size on GPU
+    grad_accum_steps: int = 4        # Effective batch size = 64 * 4 = 256
+    learning_rate: float = 3e-4
+    min_lr: float = 1e-6
+    weight_decay: float = 1e-2
+    warmup_steps: int = 300
+    max_steps_per_shard: int = 250   # Steps per monthly shard to ensure rapid progression
+    max_epochs: int = 5
+    fp16: bool = True                # Mixed precision (FP16/BF16)
+    
+    # Checkpoint settings
+    save_every_steps: int = 200
+    checkpoint_dir: str = "/content/drive/MyDrive/ohlcv_checkpoints"
+    local_checkpoint_dir: str = "./checkpoints"
+    
+    # Telemetry
+    log_interval: int = 25
+    device: str = "cuda" if torch.cuda.is_available() else "cpu"