gnnwr.models module

class gnnwr.models.GNNWR(train_dataset, valid_dataset, test_dataset, dense_layers=None, start_lr: float = 0.1, optimizer='Adagrad', drop_out=0.2, batch_norm=True, activate_func=PReLU(num_parameters=1), model_name='GNNWR_20231128-111456', model_save_path='../gnnwr_models', write_path='../gnnwr_runs/20231128-111456', use_gpu: bool = True, use_ols: bool = True, log_path='../gnnwr_logs/', log_file_name='gnnwr20231128-111456.log', log_level=20, optimizer_params=None)

Bases: object

GNNWR(Geographically neural network weighted regression) is a model to address spatial non-stationarity in various domains with complex geographical processes, which comes from the paper Geographically neural network weighted regression for the accurate estimation of spatial non-stationarity.

Parameters:

• train_dataset (baseDataset) – the dataset of training

• valid_dataset (baseDataset) – the dataset of validation

• test_dataset (baseDataset) – the dataset of testing

• dense_layers (list) –

  the dense layers of the model (default: None)

  Default structure is a geometric sequence of power of 2, the minimum is 2, and the maximum is the power of 2 closest to the number of neurons in the input layer.

  i.e. [2,4,8,16,32,64,128,256]

• start_lr (float) – the start learning rate of the model (default: 0.1)

• optimizer (str, optional) – the optimizer of the model (default: "Adagrad") choose from “SGD”,”Adam”,”RMSprop”,”Adagrad”,”Adadelta”

• drop_out (float) – the drop out rate of the model (default: 0.2)

• batch_norm (bool, optional) – whether use batch normalization (default: True)

• activate_func (torch.nn) – the activate function of the model (default: nn.PReLU(init=0.4))

• model_name (str) – the name of the model (default: "GNNWR_" + datetime.datetime.today().strftime("%Y%m%d-%H%M%S"))

• model_save_path (str) – the path of the model (default: "../gnnwr_models")

• write_path (str) – the path of the log (default: "../gnnwr_runs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))

• use_gpu (bool) – whether use gpu or not (default: True)

• use_ols (bool) – whether use ols or not (default: True)

• log_path (str) – the path of the log (default: "../gnnwr_logs/")

• log_file_name (str) – the name of the log (default: "gnnwr" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S") + ".log")

• log_level (int) – the level of the log (default: logging.INFO)

• optimizer_params (dict, optional) –

  the params of the optimizer and the scheduler (default: None)

  if optimizer is SGD, the params are:

  maxlr: float, the max learning rate (default: 0.1)

  minlr: float, the min learning rate (default: 0.01)

  upepoch: int, the epoch of learning rate up (default: 10000)

  decayepoch: int, the epoch of learning rate decay (default: 20000)

  decayrate: float, the rate of learning rate decay (default: 0.1)

  stop_change_epoch: int, the epoch of learning rate stop change (default: 30000)

  stop_lr: float, the learning rate when stop change (default: 0.001)

  if optimizer is Other, the params are:

  scheduler: str, the name of the scheduler (default: "CosineAnnealingWarmRestarts") in {"MultiStepLR","CosineAnnealingLR","CosineAnnealingWarmRestarts"}

  scheduler_milestones: list, the milestones of the scheduler MultiStepLR (default: [500,1000,2000,4000])

  scheduler_gamma: float, the gamma of the scheduler MultiStepLR (default: 0.5)

  scheduler_T_max: int, the T_max of the scheduler CosineAnnealingLR (default: 1000)

  scheduler_eta_min: float, the eta_min of the scheduler CosineAnnealingLR and CosineAnnealingWarmRestarts (default: 0.01)

  scheduler_T_0: int, the T_0 of the scheduler CosineAnnealingWarmRestarts (default: 100)

  scheduler_T_mult: int, the T_mult of the scheduler CosineAnnealingWarmRestarts (default: 3)

add_graph()

add the graph of the model to tensorboard

getLoss()

get network’s loss

Returns:

the list of the loss in training process and validation process

Return type:

list

getWeights()

get weight of each argument

Returns:

the Pandas dataframe of the weight of each argument in train_dataset

Return type:

dataframe

gpumodel_to_cpu(path, save_path, use_model=True)

convert gpu model to cpu model

Parameters:

• path (str) – the path of the model

• save_path (str) – the path of the new model

• use_model (bool) – whether use dict to load the model (default: True)

init_optimizer(optimizer, optimizer_params=None)

initialize the optimizer

Parameters:

• optimizer (str) – the optimizer of the model (default: "Adagrad") choose from “SGD”,”Adam”,”RMSprop”,”Adagrad”,”Adadelta”

• optimizer_params (dict, optional) –

  the params of the optimizer and the scheduler (default: None)

  if optimizer is SGD, the params are:

  maxlr: float, the max learning rate (default: 0.1)

  minlr: float, the min learning rate (default: 0.01)

  upepoch: int, the epoch of learning rate up (default: 10000)

  decayepoch: int, the epoch of learning rate decay (default: 20000)

  decayrate: float, the rate of learning rate decay (default: 0.1)

  stop_change_epoch: int, the epoch of learning rate stop change (default: 30000)

  stop_lr: float, the learning rate when stop change (default: 0.001)

  if optimizer is Other, the params are:

  scheduler: str, the name of the scheduler (default: "CosineAnnealingWarmRestarts") in {"MultiStepLR","CosineAnnealingLR","CosineAnnealingWarmRestarts"}

  scheduler_milestones: list, the milestones of the scheduler MultiStepLR (default: [500,1000,2000,4000])

  scheduler_gamma: float, the gamma of the scheduler MultiStepLR (default: 0.5)

  scheduler_T_max: int, the T_max of the scheduler CosineAnnealingLR (default: 1000)

  scheduler_eta_min: float, the eta_min of the scheduler CosineAnnealingLR and CosineAnnealingWarmRestarts (default: 0.01)

  scheduler_T_0: int, the T_0 of the scheduler CosineAnnealingWarmRestarts (default: 100)

  scheduler_T_mult: int, the T_mult of the scheduler CosineAnnealingWarmRestarts (default: 3)

load_model(path, use_dict=False, map_location=None)

To load a model that was automatically saved during the training process. Given that the saved file only contains the neural network components, the suggested procedure is as follows: first, instantiate an object corresponding to the model’s architectural design, and then subsequently invoke the load_model method to import the parameters and weights.

Parameters:

• path (str) – the path of the model

• use_dict (bool) – whether use dict to load the model (default: False)

• map_location (str) – the location of the model (default: None) the location can be "cpu" or "cuda"

predict(dataset)

predict the result of the dataset

Parameters:

dataset (baseDataset,predictDataset) – the dataset to be predicted

Returns:

the Pandas dataframe of the dataset with the predicted result

Return type:

dataframe

predict_weight(dataset)

predict the spatial weight of the dataset

Parameters:

dataset (baseDataset,predictDataset) – the dataset to be predicted

Returns:

the Pandas dataframe of the dataset with the predicted spatial weight

Return type:

dataframe

reg_result(filename=None, model_path=None, use_dict=False, only_return=False, map_location=None)

save the regression result of the model, including the weight of each argument, the bias, the predicted result

Parameters:

• filename (str) – the path of the result file (default: None) | if filename is None, the result will not be saved as file

• model_path (str) – the path of the model (default: None) | if model_path is None, the model will be loaded from self._modelSavePath + "/" + self._modelName + ".pkl"

• use_dict (bool) – whether use dict to load the model (default: False) | if use_dict is True, the model will be loaded from model_path as dict

• only_return (bool) – whether only return the result (default: False) | if only_return is True, the result will not be saved as file

• map_location (str) – the location of the model (default: None) the location can be "cpu" or "cuda"

Returns:

the Pandas dataframe of the result

Return type:

dataframe

result(path=None, use_dict=False, map_location=None)

print the result of the model, including the model structure, optimizer,the result of test dataset

Parameters:

• path (str) – the path of the model(default: None) | if path is None, the model will be loaded from self._modelSavePath + "/" + self._modelName + ".pkl"

• use_dict (bool) – whether use dict to load the model (default: False) | if use_dict is True, the model will be loaded from path as dict

• map_location (str) – the location of the model (default: None) the location can be "cpu" or "cuda"

run(max_epoch=1, early_stop=-1, print_frequency=50, show_detailed_info=True)

train the model and validate the model

Parameters:

• max_epoch (int) – the max epoch of the training (default: 1)

• early_stop (int) –

  if the model has not been updated for early_stop epochs, the training will stop (default: -1)

  if early_stop is -1, the training will not stop until the max epoch

• print_frequency (int) – the frequency of printing the information (default: 50)

class gnnwr.models.GTNNWR(train_dataset, valid_dataset, test_dataset, dense_layers=None, start_lr: float = 0.1, optimizer='Adam', drop_out=0.2, batch_norm=True, activate_func=PReLU(num_parameters=1), model_name='GTNNWR_20231128-111456', model_save_path='../gtnnwr_models', write_path='../gtnnwr_runs/20231128-111456', use_gpu: bool = True, use_ols: bool = True, log_path: str = '../gtnnwr_logs/', log_file_name: str = 'gtnnwr20231128-111456.log', log_level: int = 20, optimizer_params=None, STPNN_outsize=1, STNN_SPNN_params=None)

Bases: GNNWR

GTNNWR model is a model based on GNNWR and STPNN, which is a model that can be used to solve the problem of spatial-temporal non-stationarity.

Parameters:

• train_dataset (baseDataset) – the dataset for training

• valid_dataset (baseDataset) – the dataset for validation

• test_dataset (baseDataset) – the dataset for test

• dense_layers (list) – the dense layers of the model (default: None) | i.e. [[3],[128,64,32]] the first list in input is hidden layers of STPNN, the second one is hidden layers of SWNN.

• start_lr (float) – the start learning rate (default: 0.1)

• optimizer (str, optional) – the optimizer of the model (default: "Adagrad") choose from “SGD”,”Adam”,”RMSprop”,”Adagrad”,”Adadelta”

• drop_out (float) – the drop out rate of the model (default: 0.2)

• batch_norm (bool, optional) – whether use batch normalization (default: True)

• activate_func (torch.nn) – the activate function of the model (default: nn.PReLU(init=0.4))

• model_name (str) – the name of the model (default: "GNNWR_" + datetime.datetime.today().strftime("%Y%m%d-%H%M%S"))

• model_save_path (str) – the path of the model (default: "../gnnwr_models")

• write_path (str) – the path of the log (default: "../gnnwr_runs/" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S"))

• use_gpu (bool) – whether use gpu or not (default: True)

• use_ols (bool) – whether use ols or not (default: True)

• log_path (str) – the path of the log (default: "../gnnwr_logs/")

• log_file_name (str) – the name of the log (default: "gnnwr" + datetime.datetime.now().strftime("%Y%m%d-%H%M%S") + ".log")

• log_level (int) – the level of the log (default: logging.INFO)

• optimizer_params (dict, optional) –

  the params of the optimizer and the scheduler (default: None)

  if optimizer is SGD, the params are:

  maxlr: float, the max learning rate (default: 0.1)

  minlr: float, the min learning rate (default: 0.01)

  upepoch: int, the epoch of learning rate up (default: 10000)

  decayepoch: int, the epoch of learning rate decay (default: 20000)

  decayrate: float, the rate of learning rate decay (default: 0.1)

  stop_change_epoch: int, the epoch of learning rate stop change (default: 30000)

  stop_lr: float, the learning rate when stop change (default: 0.001)

  if optimizer is Other, the params are:

  scheduler: str, the name of the scheduler (default: "CosineAnnealingWarmRestarts") in {"MultiStepLR","CosineAnnealingLR","CosineAnnealingWarmRestarts"}

  scheduler_milestones: list, the milestones of the scheduler MultiStepLR (default: [500,1000,2000,4000])

  scheduler_gamma: float, the gamma of the scheduler MultiStepLR (default: 0.5)

  scheduler_T_max: int, the T_max of the scheduler CosineAnnealingLR (default: 1000)

  scheduler_eta_min: float, the eta_min of the scheduler CosineAnnealingLR and CosineAnnealingWarmRestarts (default: 0.01)

  scheduler_T_0: int, the T_0 of the scheduler CosineAnnealingWarmRestarts (default: 100)

  scheduler_T_mult: int, the T_mult of the scheduler CosineAnnealingWarmRestarts (default: 3)

• STPNN_outsize (int) – the output size of STPNN(default:1)

• STNN_SPNN_params (dict) –

  the params of STNN and SPNN(default:None)

  STPNN_batch_norm:bool

  whether use batchnorm in STNN and SPNN or not (Default:True)