From edf26b631cda6459b237747645efab225a59a77b Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Thomas=20M=C3=BCller?= Date: Sat, 30 Oct 2021 10:49:35 +0200 Subject: [PATCH] Add documentation for new losses and activations --- DOCUMENTATION.md | 46 ++++++++++++++++++++++++++++++++++++++++++++-- 1 file changed, 44 insertions(+), 2 deletions(-) diff --git a/DOCUMENTATION.md b/DOCUMENTATION.md index bc727acb..55b4f286 100644 --- a/DOCUMENTATION.md +++ b/DOCUMENTATION.md @@ -14,10 +14,12 @@ Lightning fast implementation of small multi-layer perceptrons (MLPs). Restricte { "otype": "FullyFusedMLP", // Component type. "activation": "ReLU", // Activation of hidden layers. - // Can be "ReLU" or "Sigmoid". + // Can be "ReLU", "Sigmoid", + // "Squareplus" or "Softplus". "output_activation": "None", // Activation of the output layer. // Can be "None", "ReLU", "Sigmoid", - // or "Exponential". + // "Exponential", "Squareplus" or + // "Softplus". "n_neurons": 128, // Neurons in each hidden layer. // May only be 32, 64 or 128. "n_hidden_layers": 5, // Number of hidden layers. @@ -112,6 +114,46 @@ The encoding used in Neural Radiance Caching [Müller et al. 2021] (to appear). ## Losses +### L1 + +Standard L1 loss. + +```json5 +{ + "otype": "L1" // Component type. +} +``` + +### Relative L1 + +Relative L1 loss normalized by the network prediction. + +```json5 +{ + "otype": "RelativeL1" // Component type. +} +``` + +### MAPE + +Mean absolute percentage error (MAPE). The same as Relative L1, but normalized by the target. + +```json5 +{ + "otype": "MAPE" // Component type. +} +``` + +### SMAPE + +Symmetric mean absolute percentage error (SMAPE). The same as Relative L1, but normalized by the mean of the prediction and the target. + +```json5 +{ + "otype": "SMAPE" // Component type. +} +``` + ### L2 Standard L2 loss.