Ridge Regression Weights

This release includes the weights for the ridge regression models trained on top of the CNNs.

For the DHS out-of-country models...

• The keys are: ['angola_w', 'angola_b', 'benin_w', 'benin_b', 'burkina_faso_w', 'burkina_faso_b', 'cameroon_w', 'cameroon_b', 'cote_d_ivoire_w', 'cote_d_ivoire_b', 'democratic_republic_of_congo_w', 'democratic_republic_of_congo_b', 'ethiopia_w', 'ethiopia_b', 'ghana_w', 'ghana_b', 'guinea_w', 'guinea_b', 'kenya_w', 'kenya_b', 'lesotho_w', 'lesotho_b', 'malawi_w', 'malawi_b', 'mali_w', 'mali_b', 'mozambique_w', 'mozambique_b', 'nigeria_w', 'nigeria_b', 'rwanda_w', 'rwanda_b', 'senegal_w', 'senegal_b', 'sierra_leone_w', 'sierra_leone_b', 'tanzania_w', 'tanzania_b', 'togo_w', 'togo_b', 'uganda_w', 'uganda_b', 'zambia_w', 'zambia_b', 'zimbabwe_w', 'zimbabwe_b']
• The out-of-country ridge weights were trained via "leave-one-country-out" cross-validation. Thus, the 'angola_w'/'angola_b' weights should be used for the points from Angola, and similarly for the other countries.

For the DHS incountry models...

• The keys are: ['A_w', 'A_b', 'B_w', 'B_b', 'C_w', 'C_b', 'D_w', 'D_b', 'E_w', 'E_b']
• The incountry ridge weights were trained via "leave-one-fold-out" cross-validation. Thus, the 'A_w'/'A_b' weights should be used for the points that belonged in "A" fold, and similarly for the other folds.

Weight dimensions:

• For the ms and nl models, the ridge regression model has dimension d=512, plus 1 bias term.
• For the msnl_concat models, the ridge regression model has dimension d=1024, plus 1 bias term.
• • '*_w' is a d-dimensional np.ndarray corresponding to the d-dimensional linear coefficients, and '*_b' is a (1,)-shape np.ndarray corresponding to the bias term

LSMS model ridge regression weights will be added here at a later time.

LSMS Model Checkpoints

Changes from v1.0.1 to v1.0.2:

• Added TensorFlow model checkpoints for LSMS Delta and Index-of-Delta models
• See v1.0.1 for checkpoints of all DHS models

Checkpoints are named as LSMS{type}_Incountry_{fold}_{bands}_{init}_b{batch}_fc{reg}_conv{reg}_lr{lr} where

• {type}: either "Delta" or "IndexOfDelta"
  • "Delta" corresponds to "difference of indexes" in the Nature Comms. paper
  • "IndexOfDelta" correponds to "index of differences" in the Nature Comms. paper
• {fold}: the fold that the model was tested on
• {bands}: one of MS (multispectral), NL (nightlights), or MSNL (both)
• {init}: the weights initialization strategy used
• {batch}: batch size
• {reg}: the L2 regularization coefficient is 0.{reg}
• {lr}: the initial learning rate is 0.{lr}

More model checkpoints

TensorFlow Model Checkpoints for models trained on DHS data.

Model Category	Naming Scheme
out-of-country (OOC)	DHS_OOC_{fold}_{bands}_{init}_b{batch}_fc{reg}_conv{reg}_lr{lr}
in-country	DHS_Incountry_{fold}_{bands}_{init}_b{batch}_fc{reg}_conv{reg}_lr{lr}
transfer learning	transfer_nlcenter_{bands}_b{batch}_fc{reg}_conv{reg}_lr{lr}

• {fold}: the fold that the model was tested on
• {bands}: one of MS, NL, or RGB
• {init}: the weights initialization strategy used
• {batch}: batch size
• {reg}: the L2 regularization coefficient is 0.{reg} if {reg} does not include a period ., or {reg} otherwise
• {lr}: the initial learning rate is 0.{lr} if {lr} does not include a period ., or {lr} otherwise

Changes from v1.0 to v1.0.1:

• Model checkpoint zip files no longer have nested folders.
• Each zip file now includes a params.json file which includes the parameters used to train the model.
• Added checkpoints for DHS OOC NL and RGB models. Note: the model weights for the DHS OOC MS models are unchanged.
• Added checkpoints for DHS Incountry MS and NL models.
• Added checkpoints for transfer learning models.
• Added ImageNet pretrained weights (imagenet_resnet18_tensorpack.npz)

Model Checkpoints

July 29, 2020: DHS Out-of-Country Checkpoints

Released TensorFlow checkpoints for MS (multispectral), NL (nightlights), and RGB models.

Checkpoints are named as DHS_OOC_{fold}_{bands}_{init}_b{batch}_fc{reg}_conv{reg}_lr{lr} where

• {fold}: the fold that the model was tested on
• {bands}: one of MS, NL, or RGB
• {init}: the weights initialization strategy used
• {batch}: batch size
• {reg}: the L2 regularization coefficient is 0.{reg}
• {lr}: the initial learning rate is 0.{lr}

Checkpoints for the other models will be released later.