Эксперимент: Rank-Adaptive Tensor-Train MLP (#1408)

{ "worked": false, "confidence": 9, "verdict": "Built a TT-SVD implementation, TTLinear contraction, rank-cap/tolerance handling, rounding, and a reproducible CUDA mini-regression experiment. The uncapped TT-SVD satisfied the requested relative Frobenius bound (2.46e-15 error versus 0.08), and contraction matched dense reconstruction to 4.16e-6, while the capped rank-8 approximation compressed parameters from 4096 to 1280 but violated the tolerance with 0.775 relative error. In training, the compressed TT model used 448 parameters versus 4160 for dense, but reached MSE 0.0744 versus 2.72e-5 and was 1.9x slower, so the proposed practical win was not observed.", "metrics": { "baseline": "Dense: 4160 parameters, final MSE 2.7150e-05, step-50 MSE 0.01797, 0.712 s", "idea": "TT: 448 parameters, ranks [1,4,4,1], final MSE 0.07440, step-50 MSE 0.07721, 1.358 s; capped rank-8 SVD compression 1280 vs 4096 matrix parameters, relative error 0.7747; uncapped relative error 2.46e-15 and bound satisfied" }, "how_to_run": "/home/maxwelhelp/main/bin/python3 tt_mlp_experiment.py", "files": [ "tt_mlp_experiment.py" ], "limitations": "Only a small synthetic regression task and one 64x64 matrix factorization were tested; no CIFAR-10, matched-FLOP comparison, conventional low-rank baseline, peak-memory measurement, rank-growth study, or larger MLP evaluation was performed." }