https://b1ackjack12.github.io/https://b1ackjack12.github.io/about/https://b1ackjack12.github.io/contact/https://b1ackjack12.github.io/posts/batch-512-textbook-fixes-measured/https://b1ackjack12.github.io/posts/batch-level-gpu-augmentation-accuracy/https://b1ackjack12.github.io/posts/channels-last-memory-format-measured/https://b1ackjack12.github.io/posts/cpu-inference-threading-measured/https://b1ackjack12.github.io/posts/cpu-vs-gpu-serving-crossover/https://b1ackjack12.github.io/posts/dataloader-num-workers-windows/https://b1ackjack12.github.io/posts/early-stopping-checkpoint-selection-measured/https://b1ackjack12.github.io/posts/ema-decay-dose-response/https://b1ackjack12.github.io/posts/ema-weights-measured/https://b1ackjack12.github.io/posts/forgotten-model-eval-measured/https://b1ackjack12.github.io/posts/golden-sample-test-measured/https://b1ackjack12.github.io/posts/gpu-image-augmentation-torchvision-v2/https://b1ackjack12.github.io/posts/gpu-inference-batching-measured/https://b1ackjack12.github.io/posts/har-augmentation-fifty-splits/https://b1ackjack12.github.io/posts/har-gyro-failure-insurance/https://b1ackjack12.github.io/posts/har-loso-error-analysis/https://b1ackjack12.github.io/posts/har-model-to-int8-onnx/https://b1ackjack12.github.io/posts/har-window-length-measured/https://b1ackjack12.github.io/posts/image-preprocessing-bugs-ranked/https://b1ackjack12.github.io/posts/imu-augmentation-measured/https://b1ackjack12.github.io/posts/imu-har-subject-vs-random-split/https://b1ackjack12.github.io/posts/imu-rotation-angle-sweep/https://b1ackjack12.github.io/posts/label-smoothing-calibration-measured/https://b1ackjack12.github.io/posts/label-smoothing-dose-response/https://b1ackjack12.github.io/posts/loss-item-logging-tax-measured/https://b1ackjack12.github.io/posts/lr-schedule-face-off/https://b1ackjack12.github.io/posts/mastering-ai-trends-via-primary-sources/https://b1ackjack12.github.io/posts/mixed-precision-autocast-in-practice/https://b1ackjack12.github.io/posts/optimizing-deep-learning-cuda-cudnn-setup/https://b1ackjack12.github.io/posts/optimizing-deep-learning-models-for-edge-via-quantization/https://b1ackjack12.github.io/posts/optimizing-llm-token-efficiency/https://b1ackjack12.github.io/posts/optimizing-pytorch-deep-learning-pipelines/https://b1ackjack12.github.io/posts/ort-graph-optimization-levels/https://b1ackjack12.github.io/posts/preparing-imu-sensor-data-for-deep-learning/https://b1ackjack12.github.io/posts/pytorch-core-architecture-tensors-autograd-modules/https://b1ackjack12.github.io/posts/pytorch-dataloader-num-workers-windows/https://b1ackjack12.github.io/posts/pytorch-same-seed-different-model/https://b1ackjack12.github.io/posts/pytorch-vs-jax-2026-framework-guide/https://b1ackjack12.github.io/posts/quantizing-onnx-models-in-practice/https://b1ackjack12.github.io/posts/seed-noise-vs-real-improvements/https://b1ackjack12.github.io/posts/sgd-vs-adam-fixed-budget/https://b1ackjack12.github.io/posts/stacking-the-months-wins/https://b1ackjack12.github.io/posts/temperature-scaling-measured/https://b1ackjack12.github.io/posts/test-time-augmentation-measured/https://b1ackjack12.github.io/posts/tf32-cudnn-benchmark-measured/https://b1ackjack12.github.io/posts/torch-compile-on-windows/https://b1ackjack12.github.io/posts/training-knob-effect-sizes/https://b1ackjack12.github.io/posts/training-past-gpu-memory-limits/https://b1ackjack12.github.io/posts/what-actually-breaks-exporting-pytorch-to-onnx/https://b1ackjack12.github.io/posts/why-perplexity-is-a-vanity-metric-for-llm-evaluation/https://b1ackjack12.github.io/privacy-policy/