Mass-Conserving Perceptron (MCP) SingleNode Model Zoo

This repository presents a cleaned, refactored, documented, and reorganized version of the research code originally released on Zenodo, with assistance from OpenAI's ChatGPT (GPT-5.6 Sol, High reasoning mode).

The original code was developed for the experiments reported in Wang & Gupta (2024), while this repository restructures those implementations into a more usable, reproducible, and maintainable model zoo.

It contains single-node Mass-Conserving Perceptron (MCP) model variants, together with cleaned training and evaluation scripts, historical checkpoints, standardized model notation, and documented execution conventions for model reproduction, continuation, fine-tuning, and further development.

Readers are encouraged to consult the original paper for the formal model names, notation, and mathematical definitions associated with each MCP variant.

Full article: Wang & Gupta (2024)
DOI: 10.1029/2023WR036461

Wang, Y.-H. and Gupta, H.V. (2024). A mass-conserving-perceptron for machine-learning-based modeling of geoscientific systems. Water Resources Research, 60(4), e2023WR036461.

Conceptual overview of a single-node MCP

The figure below illustrates the functional components and modeling utilities that can be represented within a single-node MCP, including:

input bias correction
input bypass
gate functional complexity
output-gate context dependence
unobserved mass exchange with the environment
mass conservation.

Conceptual illustration of the functional components and modeling utilities represented in a single-node MCP.

1. Notation

The model notation used throughout the repository is:

Symbol	Meaning
$O$	Output gate
$L$	Loss gate
$R$	Remember gate
$\kappa$	Constant gate
$\sigma$	Sigmoid-based time-variable gate
$\mathrm{con}$	PET-constrained loss gate
$B_{Lx}$	Piecewise-linear precipitation bias-correction gate with dimension $x$
$B_{Qx}$	Piecewise-quadratic precipitation bias-correction gate with dimension $x$
$A_x$	Single hidden ANN layer with $x$ nodes
$+$	Multi-information (MI): the gate uses additional state information
MR	Mass-relaxation mechanism

2. Base single-node MCP models

The basic family contains constant ($\kappa$) and sigmoid-variable ($\sigma$) output/loss gates.

For variable gates, the output gate depends on the current storage state and the loss gate depends on current PET.

Model	Notation	Description	MCP_Zoo class
M1	$\mathrm{MC}\lbrace O_{\kappa}L_{\kappa}\rbrace$	Constant output + constant loss	`MCPBRNN_constant_OutLoss`
M2	$\mathrm{MC}\lbrace O_{\kappa}L_{\sigma}\rbrace$	Constant output + PET-variable loss	`MCPBRNN_Generic_constant_Out_variableLoss`
M3	$\mathrm{MC}\lbrace O_{\sigma}L_{\kappa}\rbrace$	Storage-variable output + constant loss	`MCPBRNN_Generic_variable_Out_constantLoss`
M4	$\mathrm{MC}\lbrace O_{\sigma}L_{\sigma}\rbrace$	Storage-variable output + PET-variable loss	`MCPBRNN_Generic_Scaling`
M5	$\mathrm{MC}\lbrace O_{\sigma}L_{\sigma}^{\mathrm{con}}\rbrace$	M4 + PET-constrained loss	`MCPBRNN_Generic_PETconstraint_Scaling`

Historical checkpoint references:

Model	Historical checkpoint
M1	`model_epoch42.pt`
M2	`model_epoch468.pt`
M3	`model_epoch4963.pt`
M4	`model_epoch48.pt`
M5	`model_epoch27.pt`

3. Precipitation bias-correction models

These models extend M5 with a precipitation-dependent bias-correction function.

For both model families below, M1–M5 indicate the dimensionality of the bias-correction function, with dimensions ranging from 1 to 5. They do not refer to the five base MCP models described in Section 2.

3.1 Piecewise-linear correction: `M-IBCorrPL`

The model family is denoted as

$MC\lbrace O_{\sigma}L_{\sigma}^{con}B_{Lx}\rbrace,\quad x=1,\ldots,5$

where $B_{Lx}$ represents a piecewise-linear precipitation bias-correction function with dimension $x$.

Accordingly:

M1-IBCorrPL corresponds to $B_{L1}$
M2-IBCorrPL corresponds to $B_{L2}$
M3-IBCorrPL corresponds to $B_{L3}$
M4-IBCorrPL corresponds to $B_{L4}$
M5-IBCorrPL corresponds to $B_{L5}$

3.2 Piecewise-quadratic correction: `M-IBCorrPQ`

The model family is denoted as

$MC\lbrace O_{\sigma}L_{\sigma}^{con}B_{Qx}\rbrace,\quad x=1,\ldots,5$

where $B_{Qx}$ represents a piecewise-quadratic precipitation bias-correction function with dimension $x$.

Accordingly:

M1-IBCorrPQ corresponds to $B_{Q1}$
M2-IBCorrPQ corresponds to $B_{Q2}$
M3-IBCorrPQ corresponds to $B_{Q3}$
M4-IBCorrPQ corresponds to $B_{Q4}$
M5-IBCorrPQ corresponds to $B_{Q5}$

Folders:

M-IBCorrPL/
├── M1-IBCorrPL/
├── M2-IBCorrPL/
├── M3-IBCorrPL/
├── M4-IBCorrPL/
└── M5-IBCorrPL/

M-IBCorrPQ/
├── M1-IBCorrPQ/
├── M2-IBCorrPQ/
├── M3-IBCorrPQ/
├── M4-IBCorrPQ/
└── M5-IBCorrPQ/

4. Mass-relaxation models

These models extend M5 by incorporating a mass-relaxation (MR) mechanism.

The four MR variants are:

Model	Notation	Description
MR-1	$MC\lbrace O_{\sigma}L_{\sigma}^{\mathrm{con}}M_{\sigma}^{R}\rbrace$	Regular
MR-2	$MC\lbrace O_{\sigma}L_{\sigma}^{\mathrm{con}}M_{I}^{R}\rbrace$	Independent
MR-3	$MC\lbrace O_{\sigma}L_{\sigma}^{\mathrm{con}}M_{\sigma r}^{R}\rbrace$	Regular-Relaxed
MR-4	$MC\lbrace O_{\sigma}L_{\sigma}^{\mathrm{con}}M_{Ir}^{R}\rbrace$	Independent-Relaxed

Folders:

MR-1/
MR-2/
MR-3/
MR-4/

5. Complex gate models: `M-ComplexGate`

$A_x$ denotes a one-hidden-layer ANN with $x=1,\ldots,5$ hidden nodes.

5.1 LossGateOnly

The model family is denoted as

$MC\lbrace O_{\sigma}L_{A_x}^{\mathrm{con}}\rbrace,\quad x=1,\ldots,5$

Folder structure:

M-ComplexGate/
└── LossGateOnly/
    ├── loss-dim-1/
    ├── loss-dim-2/
    ├── loss-dim-3/
    ├── loss-dim-4/
    └── loss-dim-5/

Class:

MCPBRNN_Generic_LossANNGate_PETconstraint

The output gate is storage-dependent sigmoid. The loss gate is a PET-driven ANN.

5.2 OutputGateOnly

The model family is denoted as

$MC\lbrace O_{A_x}L_{\sigma}^{\mathrm{con}}\rbrace,\quad x=1,\ldots,5$

Folder structure:

M-ComplexGate/
└── OutputGateOnly/
    ├── output-dim-1/
    ├── output-dim-2/
    ├── output-dim-3/
    ├── output-dim-4/
    └── output-dim-5/

Class:

MCPBRNN_Generic_OutputANNGate_PETconstraint

The output gate is a storage-driven ANN. The loss gate remains PET-dependent sigmoid.

5.3 BothLossOutputGate

The model family is denoted as

$MC\lbrace O_{A_x}L_{A_y}^{\mathrm{con}}\rbrace,\quad x,y=1,\ldots,5$

There are 25 combinations:

loss-dim-1-out-dim-1
...
loss-dim-5-out-dim-5

Here, $x$ and $y$ denote the dimensions of the output-gate and loss-gate ANN functions, respectively.

Class:

MCPBRNN_Generic_ANNGate_PETconstraint_Generic

The output and loss gates each use a one-hidden-layer ANN.

6. Multi-information complex gate models: `M-MI-ComplexGate`

MI means that a gate uses additional information beyond its standard current-time input.

For the current implementation:

the output gate can use current storage plus lagged storage ($t-1$);
the loss gate can additionally use current storage information;
the superscript $^{+}$ denotes the MI extension.

6.1 LossGateOnly

The model family is denoted as

$MC\lbrace O_{\sigma}L_{A_x^{+}}^{\mathrm{con}}\rbrace,\quad x=1,\ldots,5$

Class:

MCPBRNN_Generic_PETconstraint_MIloss

6.2 OutputGateOnly

The model family is denoted as

$MC\lbrace O_{A_x^{+}}L_{\sigma}^{\mathrm{con}}\rbrace,\quad x=1,\ldots,5$

Class:

MCPBRNN_Generic_PETconstraint_MIoutput

6.3 BothLossOutputGate

The model family is denoted as

$MC\lbrace O_{A_x^{+}}L_{A_y^{+}}^{\mathrm{con}}\rbrace,\quad x,y=1,\ldots,5$

There are 25 combinations.

Here, $x$ and $y$ denote the dimensions of the output-gate and loss-gate ANN functions, respectively.

Class:

MCPBRNN_Generic_PETconstraint_MIoutputloss

6.4 MI-Only sigmoid models

These cases use direct sigmoid gates instead of ANN layers.

Folder	Notation	Class
`loss-gate-only/`	$MC\lbrace O_{\sigma}L_{\sigma^{+}}^{\mathrm{con}}\rbrace$	`MCPBRNN_Generic_PETconstraint_MIloss_Sigmoid`
`output-gate-only/`	$MC\lbrace O_{\sigma^{+}}L_{\sigma}^{\mathrm{con}}\rbrace$	`MCPBRNN_Generic_PETconstraint_MIoutput_Sigmoid`
`loss-output-gate/`	$MC\lbrace O_{\sigma^{+}}L_{\sigma^{+}}^{\mathrm{con}}\rbrace$	`MCPBRNN_Generic_PETconstraint_MIoutputloss_Sigmoid`

7. Recommended repository structure

project/
├── 20220527-MDUPLEX-LeafRiver/
├── MCPBRNN_lib_tools/
│   ├── MCP_Zoo.py
│   ├── Eval_Metric.py
│   └── Loss_Function.py
│
├── M1/
├── M2/
├── M3/
├── M4/
├── M5/
│
├── M-IBCorrPL/
│   ├── M1-IBCorrPL/
│   └── ...
│
├── M-IBCorrPQ/
│   ├── M1-IBCorrPQ/
│   └── ...
│
├── M-MassRelaxation/
│   ├── MR1/
│   ├── MR2/
│   ├── MR3/
│   └── MR4/
│
├── M-ComplexGate/
│   ├── LossGateOnly/
│   ├── OutputGateOnly/
│   └── BothLossOutputGate/
│
└── M-MI-ComplexGate/
    ├── LossGateOnly/
    ├── OutputGateOnly/
    ├── BothLossOutputGate/
    └── MI-Only/
        ├── loss-gate-only/
        ├── output-gate-only/
        └── loss-output-gate/

Leaf River data: `20220527-MDUPLEX-LeafRiver`

The 20220527-MDUPLEX-LeafRiver folder contains the daily hydrometeorological data for the Leaf River catchment in Mississippi, USA.

The current training and evaluation scripts use two files:

LeafRiverDaily_43YR.txt — 43-year daily data containing three columns: precipitation ($P$), potential evapotranspiration ($PET$), and streamflow ($Q$).
LeafRiverDaily_43YR_Flag.txt — data-partition flags used to identify spinup, training, selection, and testing periods.

The partition flags are:

Flag	Period
`-99999`	Spinup
`-1`	Training
`0`	Selection
`1`	Testing

Other files in this directory are retained from earlier experiments and alternative data-partition schemes, but are not used by the current cleaned Model Zoo scripts.

All model classes should be consolidated in:

MCPBRNN_lib_tools/MCP_Zoo.py

8. Running the models

All model scripts can be executed from the repository root.

The scripts use argparse for optional runtime settings. Fixed-structure models can be run directly, whereas model families with multiple functional dimensions require the corresponding dimension argument to select the intended model case.

The included checkpoint is used by default unless --checkpoint is explicitly provided.

8.1 Base single-node MCP models

The five base models have fixed structures and can be run directly.

M1

Training:

python M1/mcpbrnn_Main_constantO_constantL_clean.py

Evaluation:

python M1/mcpbrnn_Main_constantO_constantL_EVAL_clean.py

M2

Training:

python M2/mcpbrnn_Main_constantO_variableL_clean.py

Evaluation:

python M2/mcpbrnn_Main_constantO_variantL_EVAL_clean.py

M3

Training:

python M3/mcpbrnn_Main_variableO_constantL_clean.py

Evaluation:

python M3/mcpbrnn_Main_variantO_constantL_EVAL_clean.py

M4

Training:

python M4/mcpbrnn_Main_M4_Scaling_clean.py

Evaluation:

python M4/mcpbrnn_Main_M4_Scaling_EVAL_clean.py

M5

Training:

python M5/mcpbrnn_Main_PETconstraint_clean.py

Evaluation:

python M5/evaluate_Main_PETconstraint_clean.py

8.2 Precipitation bias-correction models

The precipitation bias-correction families share one training script and one evaluation script across dimensions 1–5.

The desired bias-correction dimension is specified using:

--dim x

where x = 1,...,5.

Piecewise-linear correction: `M-IBCorrPL`

For example, to run dimension 3:

python M-IBCorrPL/mcpbrnn_Main_PETconstraint_IBcorrPL_clean.py --dim 3

Evaluation:

python M-IBCorrPL/mcpbrnn_Main_PETconstraint_IBcorrPL_EVAL_clean.py --dim 3

The script automatically associates --dim 3 with:

M-IBCorrPL/M3-IBCorrPL/

Piecewise-quadratic correction: `M-IBCorrPQ`

For example, to run dimension 3:

python M-IBCorrPQ/mcpbrnn_Main_PETconstraint_IBcorrPQ_clean.py --dim 3

Evaluation:

python M-IBCorrPQ/mcpbrnn_Main_PETconstraint_IBcorrPQ_EVAL_clean.py --dim 3

The script automatically associates --dim 3 with:

M-IBCorrPQ/M3-IBCorrPQ/

8.3 Mass-relaxation models

The four mass-relaxation cases have fixed structures and can be run directly.

MR-1

python MR-1/mcpbrnn_Main_MR1_Regular_clean.py
python MR-1/mcpbrnn_Main_MR1_Regular_EVAL_clean.py

MR-2

python MR-2/mcpbrnn_Main_MR2_Independent_clean.py
python MR-2/mcpbrnn_Main_MR2_Independent_EVAL_clean.py

MR-3

python MR-3/mcpbrnn_Main_MR3_Regular_Relaxed_clean.py
python MR-3/mcpbrnn_Main_MR3_Regular_Relaxed_EVAL_clean.py

MR-4

python MR-4/mcpbrnn_Main_MR4_Independent_Relaxed_clean.py
python MR-4/mcpbrnn_Main_MR4_Independent_Relaxed_EVAL_clean.py

8.4 High-dimensional complex gate models: `M-ComplexGate`

LossGateOnly

Specify the loss-gate ANN dimension using --dim.

For example:

python M-ComplexGate/LossGateOnly/mcpbrnn_Main_lossANNGate_clean.py --dim 3
python M-ComplexGate/LossGateOnly/mcpbrnn_Main_PETconstraint_lossANNGate_EVAL_clean.py --dim 3

This selects:

M-ComplexGate/LossGateOnly/loss-dim-3/

OutputGateOnly

Specify the output-gate ANN dimension using --dim.

For example:

python M-ComplexGate/OutputGateOnly/mcpbrnn_Main_outputANNGate_clean.py --dim 3
python M-ComplexGate/OutputGateOnly/mcpbrnn_Main_PETconstraint_outputANNGate_EVAL_clean.py --dim 3

This selects:

M-ComplexGate/OutputGateOnly/out-dim-3/

BothLossOutputGate

The loss- and output-gate ANN dimensions are specified independently using:

--loss_dim y --out_dim x

For example, for a loss-gate dimension of 3 and an output-gate dimension of 5:

python M-ComplexGate/BothLossOutputGate/mcpbrnn_Main_ANNGate_clean.py --loss_dim 3 --out_dim 5

Evaluation:

python M-ComplexGate/BothLossOutputGate/mcpbrnn_Main_ANNGate_PETconstraint_EVAL_clean.py --loss_dim 3 --out_dim 5

This selects:

M-ComplexGate/BothLossOutputGate/loss-dim-3-out-dim-5/

8.5 Multi-information complex gate models: `M-MI-ComplexGate`

LossGateOnly

Specify the loss-gate ANN dimension using --dim.

python M-MI-ComplexGate/LossGateOnly/mcpbrnn_Main_MI_Loss_PETconstraint_clean.py --dim 3
python M-MI-ComplexGate/LossGateOnly/mcpbrnn_Main_MI_Loss_PETconstraint_EVAL_clean.py --dim 3

OutputGateOnly

Specify the output-gate ANN dimension using --dim.

python M-MI-ComplexGate/OutputGateOnly/mcpbrnn_Main_MI_Output_PETconstraint_clean.py --dim 3
python M-MI-ComplexGate/OutputGateOnly/mcpbrnn_Main_MI_Output_PETconstraint_EVAL_clean.py --dim 3

BothLossOutputGate

Specify both ANN dimensions using --loss_dim and --out_dim.

For example:

python M-MI-ComplexGate/BothLossOutputGate/mcpbrnn_Main_MI_LossOutput_PETconstraint_clean.py --loss_dim 3 --out_dim 5

Evaluation:

python M-MI-ComplexGate/BothLossOutputGate/mcpbrnn_Main_MI_LossOutput_PETconstraint_EVAL_clean.py --loss_dim 3 --out_dim 5

8.6 MI-only sigmoid models

The MI-only sigmoid models have fixed structures and can be run directly.

Loss gate only

python M-MI-ComplexGate/MI-Only/loss-gate-only/mcpbrnn_Main_MI_Loss_PETconstraint_Sigmoid_clean.py
python M-MI-ComplexGate/MI-Only/loss-gate-only/mcpbrnn_Main_MI_Loss_PETconstraint_Sigmoid_EVAL_clean.py

Output gate only

python M-MI-ComplexGate/MI-Only/output-gate-only/mcpbrnn_Main_MI_Output_PETconstraint_Sigmoid_clean.py
python M-MI-ComplexGate/MI-Only/output-gate-only/mcpbrnn_Main_MI_Output_PETconstraint_Sigmoid_EVAL_clean.py

Loss and output gates

python M-MI-ComplexGate/MI-Only/loss-output-gate/mcpbrnn_Main_MI_LossOutput_PETconstraint_Sigmoid_clean.py
python M-MI-ComplexGate/MI-Only/loss-output-gate/mcpbrnn_Main_MI_LossOutput_PETconstraint_Sigmoid_EVAL_clean.py

8.7 Common command-line arguments

Training scripts support optional arguments such as:

--case_no
--epoch_no
--time_lag
--seed_no
--checkpoint
--data_dir
--output_dir
--device

Dimension-dependent model families additionally use:

--dim

or:

--loss_dim
--out_dim

The available arguments and defaults for any script can be inspected using:

python <script>.py --help

9. Checkpoint convention

For cleaned scripts, the preferred convention is:

<model-folder>/
├── <one checkpoint>.pt
├── training script
└── EVAL script

If --checkpoint is not supplied, the script searches the relevant model folder for exactly one .pt file.

Training workflow:

existing checkpoint
    ↓
strict load
    ↓
continue training
    ↓
evaluate every epoch
    ↓
select best epoch by selection KGE
    ↓
best_model_epochX.pt

EVAL workflow:

checkpoint
    ↓
strict load
    ↓
forward-only evaluation
    ↓
metrics + diagnostic time series

10. Common training settings

Unless a case explicitly differs:

time_lag = 0
learning_rate = 0.025
learning_rates = {300: 0.0125, 600: 0.0125}

Data split flags:

Flag	Split
`-99999`	spinup
`-1`	training
`0`	selection
`1`	testing

Reported metrics:

NSE
KGE
KGE-A
KGE-B
Corr
MSE / mse
KGEss

Training diagnostics also include KGE at lags 1, 2, and 3.

11. Time-series output convention

Cleaned output files use:

time, <result variable>, phase

with phases:

spinup1
spinup2
spinup3
simulation

The time axis contains:

three repetitions of WY1949 (1948-10-01 to 1949-09-30) for spinup;
formal simulation from 1948-10-01 through 1988-09-30;
total length: 15,705 rows.

Repeated calendar dates across spinup phases are intentional; phase distinguishes them.

12. Implementation conventions

Cleaned model classes follow these rules:

preserve historical checkpoint parameter names and tensor shapes;
preserve pre-update versus post-update hydrologic timing;
use remember gate terminology for gR;
use x.new_zeros(...) for device-safe output tensors;
keep the seq_len = 1 assumption explicit;
remove unused imports, dead prototype code, and obsolete initialization logic;
keep equations compact and readable;
do not rename state_dict parameter keys;
Python class names may be renamed only when necessary to avoid duplicate class names in MCP_Zoo.py.

For MR3 and MR4, only the Python class names were disambiguated; the checkpoint parameter keys remain unchanged.

13. Environment and dependencies

The repository was organized for the following Python environment.

`requirement.txt`

numpy==2.2.2
pandas==2.2.3
torch==2.9.0
torchvision==0.24.0
tqdm==4.67.1
scikit-learn==1.6.1

Install with:

pip install -r requirement.txt

`environment.yml`

name: MCP-SingleNode-ModelZoo
channels:
  - defaults
dependencies:
  - python=3.10
  - pip
  - notebook
  - pip:
      - numpy==2.2.2
      - pandas==2.2.3
      - torch==2.9.0
      - torchvision==0.24.0
      - tqdm==4.67.1
      - scikit-learn==1.6.1

Create and activate the Conda environment with:

conda env create -f environment.yml
conda activate MCP-SingleNode-ModelZoo

14. Conclusion

This repository provides a consolidated and reproducible collection of single-node MCP model variants, including cleaned training and evaluation scripts, checkpoints, and data conventions.

For questions or suggestions, please contact Yuan-Heng Wang, Ph.D. at yhwang0730@gmail.com or yhwang0730@arizona.edu.

Name		Name	Last commit message	Last commit date
20220527-MDUPLEX-LeafRiver		20220527-MDUPLEX-LeafRiver
M-ComplexGate		M-ComplexGate
M-IBCorrPL		M-IBCorrPL
M-IBCorrPQ		M-IBCorrPQ
M-MI-ComplexGate		M-MI-ComplexGate
M1		M1
M2		M2
M3		M3
M4		M4
M5		M5
MCPBRNN_lib_tools		MCPBRNN_lib_tools
MR-1		MR-1
MR-2		MR-2
MR-3		MR-3
MR-4		MR-4
.DS_Store		.DS_Store
.gitattributes		.gitattributes
ConceptualFigure.png		ConceptualFigure.png
LICENSE		LICENSE
README.md		README.md
environment.yml		environment.yml
requirements.txt		requirements.txt

Latest commit

History

Folders and files

Repository files navigation

Mass-Conserving Perceptron (MCP) SingleNode Model Zoo

Conceptual overview of a single-node MCP

1. Notation

2. Base single-node MCP models

3. Precipitation bias-correction models

3.1 Piecewise-linear correction: M-IBCorrPL

3.2 Piecewise-quadratic correction: M-IBCorrPQ

4. Mass-relaxation models

5. Complex gate models: M-ComplexGate

5.1 LossGateOnly

5.2 OutputGateOnly

5.3 BothLossOutputGate

6. Multi-information complex gate models: M-MI-ComplexGate

6.1 LossGateOnly

6.2 OutputGateOnly

6.3 BothLossOutputGate

6.4 MI-Only sigmoid models

7. Recommended repository structure

Leaf River data: 20220527-MDUPLEX-LeafRiver

8. Running the models

8.1 Base single-node MCP models

M1

M2

M3

M4

M5

8.2 Precipitation bias-correction models

Piecewise-linear correction: M-IBCorrPL

Piecewise-quadratic correction: M-IBCorrPQ

8.3 Mass-relaxation models

MR-1

MR-2

MR-3

MR-4

8.4 High-dimensional complex gate models: M-ComplexGate

LossGateOnly

OutputGateOnly

BothLossOutputGate

8.5 Multi-information complex gate models: M-MI-ComplexGate

LossGateOnly

OutputGateOnly

BothLossOutputGate

8.6 MI-only sigmoid models

Loss gate only

Output gate only

Loss and output gates

8.7 Common command-line arguments

9. Checkpoint convention

10. Common training settings

11. Time-series output convention

12. Implementation conventions

13. Environment and dependencies

requirement.txt

environment.yml

14. Conclusion

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages

3.1 Piecewise-linear correction: `M-IBCorrPL`

3.2 Piecewise-quadratic correction: `M-IBCorrPQ`

5. Complex gate models: `M-ComplexGate`

6. Multi-information complex gate models: `M-MI-ComplexGate`

Leaf River data: `20220527-MDUPLEX-LeafRiver`

Piecewise-linear correction: `M-IBCorrPL`

Piecewise-quadratic correction: `M-IBCorrPQ`

8.4 High-dimensional complex gate models: `M-ComplexGate`

8.5 Multi-information complex gate models: `M-MI-ComplexGate`

`requirement.txt`

`environment.yml`