hparams.md

Datasets

The following table shows how the dataset parameter (which is passed to --dataset in train.py or --datasets in sweep.py) maps to the datasets described in the paper.

Parameter Value	Dataset	Synthetic Shift
eICU	eICU	Base
eICUCorrLabel	eICU	CorrLabel
eICUCorrNoise	eICU	CorrNoise
eICUSubsampleUnobs	eICU	BiasSampUnobs
eICUSubsampleObs	eICU	BiasSampObs
CXR	CXR (Multitask)	Base
CXRBinary	CXR (Binary)	Base
CXRSubsampleUnobs	CXR (Binary)	BiasSampUnobs
CXRSubsampleObs	CXR (Binary)	BiasSampObs
ColoredMNIST	Colored MNIST

Data Hyperparameters

The following table shows the list of parameters that can be passed as part of the JSON encoded dictionary to the --hparams argument in train.py or sweep.py. Hyperparameters not explicitly specified during training will default to (one of) their paper values.

hparam	Applicable Datasets	Possible Values	Description	Paper Value(s)
eicu_architecture	eICU*	{MLP, GRU}	Model architecture for eICU tasks	GRU
corr_label_train_corrupt_dist	eICUCorrLabel	[0, 0.5]	𝛿 from Section 4.2 of the paper	0.1
corr_label_train_corrupt_mean	eICUCorrLabel	[0, 1]	β from Section 4.2 of the paper	{0.1, 0.3, 0.5}
corr_label_val_corrupt	eICUCorrLabel	[0, 1]	p_{val} from Section 4.2 of the paper	0.5
corr_label_test_corrupt	eICUCorrLabel	[0, 1]	p_{test} from Section 4.2 of the paper	0.9
corr_noise_train_corrupt_dist	eICUCorrNoise	(-∞, +∞)	𝛿 from Section 4.3 of the paper	{0.1, 0.5}
corr_noise_train_corrupt_mean	eICUCorrNoise	(-∞, +∞)	β from Section 4.3 of the paper	{1.0, 2.0}
corr_noise_val_corrupt	eICUCorrNoise	(-∞, +∞)	λ_{val} from Section 4.3 of the paper	0.0
corr_noise_test_corrupt	eICUCorrNoise	(-∞, +∞)	λ_{test} from Section 4.3 of the paper	-1.0
corr_noise_std	eICUCorrNoise	[0, ∞)	σ from Section 4.3 of the paper	0.5
corr_noise_feature	eICUCorrNoise	column	Feature in eICU dataset to add noise to.	admissionweight
subsample_g1_mean	*Subsample*	[0, 1]	Average μ_M on the training environments	eICU: 0.7 CXR: 0.15
subsample_g2_mean	*Subsample*	[0, 1]	Average μ_F on the training environments	eICU: 0.1 CXR: 0.025
subsample_g1_dist	*Subsample*	[0, 0.5]	Distance of μ_M between each training environment	eICU: 0.1 CXR: 0.1
subsample_g2_dist	*Subsample*	[0, 0.5]	Distance of μ_F between each training environment	eICU: 0.05 CXR: 0.01
cxr_augment	CXR*	{0, 1}	Whether to use simple geometric image augmentations during training	1
use_cache	CXR*	{0, 1}	Whether to load images from pre-cached binaries, or directly from downloaded images	N/A
cmnist_eta	ColoredMNIST	[0, 1]	η from Appendix B.1 of the paper	[0, 0.5]
cmnist_beta	ColoredMNIST	[0, 1]	β from Appendix B.1 of the paper	[0.05, 0.5]
cmnist_delta	ColoredMNIST	[0, 1]	𝛿 from Appendix B.1 of the paper	[0, 0.3]