Spaces:
Paused
Paused
File size: 8,699 Bytes
c4ee290 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 | """
ShortSmith v2 - Trained Hype Scorer
Uses the MLP model trained on Mr. HiSum dataset to score segments.
Falls back to heuristic scoring if weights not available.
"""
import os
from pathlib import Path
from typing import Optional, List, Tuple
import numpy as np
import torch
import torch.nn as nn
from utils.logger import get_logger
logger = get_logger("scoring.trained_scorer")
class HypeScorerMLP(nn.Module):
"""
2-layer MLP for hype scoring.
Must match the architecture from training notebook.
"""
def __init__(
self,
visual_dim: int = 512,
audio_dim: int = 13,
hidden_dim: int = 256,
dropout: float = 0.3,
):
super().__init__()
self.visual_dim = visual_dim
self.audio_dim = audio_dim
input_dim = visual_dim + audio_dim
self.network = nn.Sequential(
# Layer 1
nn.Linear(input_dim, hidden_dim),
nn.BatchNorm1d(hidden_dim),
nn.ReLU(),
nn.Dropout(dropout),
# Layer 2
nn.Linear(hidden_dim, hidden_dim // 2),
nn.BatchNorm1d(hidden_dim // 2),
nn.ReLU(),
nn.Dropout(dropout),
# Output layer
nn.Linear(hidden_dim // 2, 1),
)
def forward(self, features: torch.Tensor) -> torch.Tensor:
"""Forward pass with concatenated features."""
return self.network(features)
class TrainedHypeScorer:
"""
Trained neural network hype scorer.
Uses MLP trained on Mr. HiSum "Most Replayed" data.
"""
# Default weights path relative to project root
DEFAULT_WEIGHTS_PATH = "weights/hype_scorer_weights.pt"
def __init__(
self,
weights_path: Optional[str] = None,
device: Optional[str] = None,
visual_dim: int = 512,
audio_dim: int = 13,
):
"""
Initialize trained scorer.
Args:
weights_path: Path to trained weights (.pt file)
device: Device to run on (cuda/cpu/mps)
visual_dim: Visual feature dimension
audio_dim: Audio feature dimension
"""
self.visual_dim = visual_dim
self.audio_dim = audio_dim
self.model = None
self.device = device or self._get_device()
# Find weights file
if weights_path is None:
# Look in common locations
candidates = [
self.DEFAULT_WEIGHTS_PATH,
"hype_scorer_weights.pt",
"weights/hype_scorer_weights.pt",
os.path.join(os.path.dirname(__file__), "..", "weights", "hype_scorer_weights.pt"),
]
for candidate in candidates:
if os.path.exists(candidate):
weights_path = candidate
break
if weights_path and os.path.exists(weights_path):
self._load_model(weights_path)
else:
logger.warning(
f"Trained weights not found. TrainedHypeScorer will use fallback scoring. "
f"To use trained model, place weights at: {self.DEFAULT_WEIGHTS_PATH}"
)
def _get_device(self) -> str:
"""Detect best available device."""
if torch.cuda.is_available():
return "cuda"
elif hasattr(torch.backends, "mps") and torch.backends.mps.is_available():
return "mps"
return "cpu"
def _load_model(self, weights_path: str) -> None:
"""Load trained model weights."""
try:
logger.info(f"Loading trained hype scorer from {weights_path}")
# Initialize model
self.model = HypeScorerMLP(
visual_dim=self.visual_dim,
audio_dim=self.audio_dim,
)
# Load weights
state_dict = torch.load(weights_path, map_location=self.device)
# Handle different save formats
if isinstance(state_dict, dict) and "model_state_dict" in state_dict:
state_dict = state_dict["model_state_dict"]
self.model.load_state_dict(state_dict)
self.model.to(self.device)
self.model.eval()
logger.info(f"✓ Trained hype scorer loaded successfully on {self.device}")
except Exception as e:
logger.error(f"Failed to load trained model: {e}")
self.model = None
@property
def is_available(self) -> bool:
"""Check if trained model is loaded."""
return self.model is not None
@torch.no_grad()
def score(
self,
visual_features: np.ndarray,
audio_features: np.ndarray,
) -> float:
"""
Score a single segment.
Args:
visual_features: Visual feature vector (visual_dim,)
audio_features: Audio feature vector (audio_dim,)
Returns:
Hype score (0-1)
"""
if not self.is_available:
return self._fallback_score(visual_features, audio_features)
# Prepare input
features = np.concatenate([visual_features, audio_features])
tensor = torch.tensor(features, dtype=torch.float32).unsqueeze(0).to(self.device)
# Forward pass
raw_score = self.model(tensor)
# Normalize to 0-1 with sigmoid
score = torch.sigmoid(raw_score).item()
return score
@torch.no_grad()
def score_batch(
self,
visual_features: np.ndarray,
audio_features: np.ndarray,
) -> np.ndarray:
"""
Score multiple segments in batch.
Args:
visual_features: Visual features (N, visual_dim)
audio_features: Audio features (N, audio_dim)
Returns:
Array of hype scores (N,)
"""
if not self.is_available:
return np.array([
self._fallback_score(visual_features[i], audio_features[i])
for i in range(len(visual_features))
])
# Prepare batch input
features = np.concatenate([visual_features, audio_features], axis=1)
tensor = torch.tensor(features, dtype=torch.float32).to(self.device)
# Forward pass
raw_scores = self.model(tensor)
# Normalize to 0-1
scores = torch.sigmoid(raw_scores).squeeze().cpu().numpy()
return scores
def _fallback_score(
self,
visual_features: np.ndarray,
audio_features: np.ndarray,
) -> float:
"""
Fallback heuristic scoring when model not available.
Uses similar logic to training data generation.
"""
# Visual contribution (mean of first 50 dims if available)
visual_len = min(50, len(visual_features))
visual_score = np.mean(visual_features[:visual_len]) * 0.5 + 0.5
visual_score = np.clip(visual_score, 0, 1)
# Audio contribution
if len(audio_features) >= 8:
audio_score = (
audio_features[0] * 0.4 + # RMS energy
audio_features[5] * 0.3 + # Spectral flux (if available)
audio_features[7] * 0.3 # Onset strength (if available)
) * 0.5 + 0.5
else:
audio_score = np.mean(audio_features) * 0.5 + 0.5
audio_score = np.clip(audio_score, 0, 1)
# Combined
return float(0.5 * visual_score + 0.5 * audio_score)
def compare_segments(
self,
visual_a: np.ndarray,
audio_a: np.ndarray,
visual_b: np.ndarray,
audio_b: np.ndarray,
) -> int:
"""
Compare two segments.
Returns:
1 if A is more engaging, -1 if B is more engaging, 0 if equal
"""
score_a = self.score(visual_a, audio_a)
score_b = self.score(visual_b, audio_b)
if score_a > score_b + 0.05:
return 1
elif score_b > score_a + 0.05:
return -1
return 0
# Singleton instance for easy access
_trained_scorer: Optional[TrainedHypeScorer] = None
def get_trained_scorer(
weights_path: Optional[str] = None,
force_reload: bool = False,
) -> TrainedHypeScorer:
"""
Get singleton trained scorer instance.
Args:
weights_path: Optional path to weights file
force_reload: Force reload even if already loaded
Returns:
TrainedHypeScorer instance
"""
global _trained_scorer
if _trained_scorer is None or force_reload:
_trained_scorer = TrainedHypeScorer(weights_path=weights_path)
return _trained_scorer
__all__ = ["TrainedHypeScorer", "HypeScorerMLP", "get_trained_scorer"]
|