mlp

Experiments Log Template

Use this file as a standard format for academic experiment tracking. Copy the template section for each run.

How to use

1. Duplicate the Experiment Entry block.
2. Fill all fields before and after running.
3. Keep one entry per experiment ID.
4. Commit this file with code changes when possible.

---

Experiment Entry

Metadata

• Experiment ID: EXP-YYYYMMDD-XX
• Date: YYYY-MM-DD
• Author:
• Branch/Commit:
• Objective:
• Hypothesis:

Environment

• OS:
• Compiler (g++/clang++ + version):
• CMake version:

• Backend: CPU

  OpenMP

  BLAS

  CUDA

  OpenMP+BLAS

  OpenMP+CUDA

• BLAS backend: disabled

  CBLAS/OpenBLAS

  Accelerate

• Build command:
• Run command:
• Hardware (CPU/GPU):
• Random seed:

Model and Training Setup

• Dataset:
• Train/Validation/Test split:
• Input features:
• Output targets:
• Architecture:
• Activation(s):
• Loss:
• Optimizer:
• Learning rate:
• Epochs:
• Batch size:
• Regularization:

Results

• Final train loss:
• Final validation loss:
• Final test loss:
• Metrics (accuracy/F1/etc.):
• Runtime:
• Benchmark speedup (if using mlp_dense_benchmark):
• Allocation profile (if working on issue #6):
• Sanitizer/Valgrind result (if relevant):
• Notes on convergence:

Analysis

• Did the hypothesis hold?
• Main failure modes:
• What changed vs previous experiment?
• Threats to validity:

Next Actions

---

Example (short)

• Experiment ID: EXP-20260311-01
• Objective: Improve XOR convergence speed with larger hidden layer.
• Changes: Dense(2,8) -> Dense(2,16), same optimizer and epochs.
• Result: Lower loss after 1000 epochs, similar final predictions.
• Next: test lower learning rates (0.1, 0.05).

This site is open source. Improve this page.