LinearSolve::luc: Result for LinearSolve of badly conditioned matrix {{7.61902,0.0118411,0.00677937,0.10901,0.0815707,0.0166805,0.184858,0.0340636,0.0288654,0.0182311,<<246>>},<<9>>,<<246>>} may contain significant numerical errors.

[dbl001]

Thanks for creating these Mathematica packages. When I run NMF on a GPT-2 embedding matrix, I get warnings from LinearSolve. I tried the following RegularizationParameters: 0.01, 0.1, 1, 2. I also tried different PrecisionGoals: 8, 16, 32, 64. Any suggestions?

gpt2Model = NetModel["GPT2 Transformer Trained on WebText Data"]
embeddings = 
 NetExtract[gpt2Model, {"embedding", "embeddingtokens", "Weights"}]
{wMat, hMat} = 
  ResourceFunction["NonNegativeMatrixFactorization"][
   Normal[embeddings], 256, "RegularizationParameter" -> 2, 
   MaxSteps -> 300, PrecisionGoal -> 64];

[snailwowo]

(⊙o⊙)嗯 我已经收到你的来信啦