weights.py

import json
from collections import defaultdict
import numpy as np

TRAIN_JSON_PATH = "./data/miko/train.json"
ACTION_STATS_PATH = "./data/action_statistics.json"

# Constants to stabilize training
MIN_TOTAL_LEN = 1.0     # Minimum allowed total duration
MAX_WEIGHT = 0.05       # Cap to prevent a single label from dominating

# Load train.json
with open(TRAIN_JSON_PATH, "r") as f:
    dataset = json.load(f)

# Collect durations for each act_cat
act_cat_stats = defaultdict(float)

for entry in dataset:
    labels = entry.get("frame_ann", {}).get("labels", [])
    for label in labels:
        start = label.get("start_t")
        end = label.get("end_t")
        act_cats = label.get("act_cat", [])
        if start is None or end is None:
            continue
        duration = max(0.01, end - start)
        for act in act_cats:
            act_cat_stats[act] += duration

# Recompute weights
cleaned_stats = {}
for act, total_len in act_cat_stats.items():
    total_len = max(MIN_TOTAL_LEN, total_len)
    if np.isfinite(total_len):
        raw_weight = 1.0 / total_len
        weight = min(raw_weight, MAX_WEIGHT)
        weight = max(0.0, weight)  # ✅ Clip negatives to 0
        cleaned_stats[act] = {
            "total_len": total_len,
            "total_weight": 1,
            "weight": weight
        }
    else:
        print(f"⚠️ Skipping act_cat '{act}' due to invalid total_len={total_len}")

# Normalize weights
total_weight_sum = sum(stats["weight"] for stats in cleaned_stats.values())
if total_weight_sum == 0:
    raise ValueError("❌ All weights are zero. Check act_cat labels or durations.")

for stats in cleaned_stats.values():
    stats["weight"] /= total_weight_sum

# Save updated statistics
with open(ACTION_STATS_PATH, "w") as f:
    json.dump(cleaned_stats, f, indent=2)

print(f"✅ Saved {len(cleaned_stats)} act_cat entries to {ACTION_STATS_PATH}")
print(f"🎯 Final normalized weight sum: {sum(stats['weight'] for stats in cleaned_stats.values()):.4f}")