TildAlice – Tech Blog for Developers
Python, AI, Deep Learning, Data Analysis, Computer Vision, Reinforcement Learning โ tips and tutorials
Latest Posts
-
MoE Router Collapse: Why 90% of Tokens Hit 2 Experts
87% of tokens routing to 2 experts? That's router collapse killing your MoE model. Here's the auxiliary loss fix and diagnostic code to catch it early.
-
TFLite Inference Fails on Android: 5 ONNX Mobile Fixes
Fix TFLite Android crashes with ONNX Runtime โ 5 solutions for dynamic shapes, quantization bugs, GPU failures, and memory leaks with real benchmarks.
-
ONNX INT8 vs FP16: 3x Latency Drop on Jetson Orin Nano
YOLOv8n INT8 cut latency from 47ms to 15ms on Jetson Orin Nano โ but small-object mAP dropped 5.7%. Real tradeoff numbers with power benchmarks.
-
Haar Cascades to Mediapipe Face Mesh: Migration Guide
Haar Cascades miss 23% of faces in real-world tests. Mediapipe Face Mesh hits 96% accuracy and gives you 468 landmarks for free โ here's how to migrate.
-
Gymnasium Custom Environment: 7 Patterns That Save Hours
Build Gymnasium custom environments faster with 7 proven patterns. Fix common pitfalls in reset(), step(), and observation spaces that waste hours.
-
SAM vs Grounded-SAM: Zero-Shot Segmentation mIoU Gap
SAM hits 46.5 mIoU on COCO but drops to 28-32 on custom objects. Grounded-SAM adds text prompts at 180ms latency. Real benchmark data inside.
-
yfinance to Polygon.io: 4 Breaking Changes in Migration
yfinance to Polygon.io migration breaks in 4 hidden ways: timezone handling, adjusted close formulas, rate limits, and missing bars. Real fixes.
-
Pre-commit Ruff Locks: Pin to CI or Accept 2-Week Drift
Fix pre-commit Ruff version conflicts by pinning to CI or accepting controlled drift. Compare both strategies and choose the right approach.