www.flow.help.3_1_3.html Maven / Gradle / Ivy

Go to download


Some model types, such as DRF, GBM, and Deep Learning, support checkpointing. A checkpoint resumes model training so that you can iterate your model. The dataset must be the same. The following  model parameters must be the same when restarting a model from a checkpoint:



Must be the same as in checkpoint model





drop_na20_cols
response_column
activation


use_all_factor_levels
adaptive_rate
autoencoder


rho
epsilon
sparse


sparsity_beta
col_major
rate


rate_annealing
rate_decay
momentum_start


momentum_ramp
momentum_stable
nesterov_accelerated_gradient


ignore_const_cols
max_categorical_features
nfolds


distribution
tweedie_power



The following parameters can be modified when restarting a model from a checkpoint: 



Can be modified





seed
checkpoint
epochs


score_interval
train_samples_per_iteration
target_ratio_comm_to_comp


score_duty_cycle
score_training_samples
score_validation_samples


score_validation_sampling
classification_stop
regression_stop


quiet_mode
max_confusion_matrix_size
mini_batch_size


diagnostics
variable_importances
initial_weight_distribution


initial_weight_scale
force_load_balance
replicate_training_data


shuffle_training_data
single_node_mode
fast_mode


l1
l2
max_w2


input_dropout_ratio
hidden_dropout_ratios
loss


overwrite_with_best_model
missing_values_handling
average_activation


reproducible
export_weights_and_biases
elastic_averaging


elastic_averaging_moving_rate
elastic_averaging_regularization




After building your model, copy the model_id. To view the model_id, click the Model menu then click List All Models. 
Select the model type from the drop-down Model menu. 
Note: The model type must be the same as the checkpointed model. 


Paste the copied model_id in the checkpoint entry field. 
Click the Build Model button. The model will resume training.

Must be the same as in checkpoint model
`drop_na20_cols`	`response_column`	`activation`
`use_all_factor_levels`	`adaptive_rate`	`autoencoder`
`rho`	`epsilon`	`sparse`
`sparsity_beta`	`col_major`	`rate`
`rate_annealing`	`rate_decay`	`momentum_start`
`momentum_ramp`	`momentum_stable`	`nesterov_accelerated_gradient`
`ignore_const_cols`	`max_categorical_features`	`nfolds`
`distribution`	`tweedie_power`

Can be modified
`seed`	`checkpoint`	`epochs`
`score_interval`	`train_samples_per_iteration`	`target_ratio_comm_to_comp`
`score_duty_cycle`	`score_training_samples`	`score_validation_samples`
`score_validation_sampling`	`classification_stop`	`regression_stop`
`quiet_mode`	`max_confusion_matrix_size`	`mini_batch_size`
`diagnostics`	`variable_importances`	`initial_weight_distribution`
`initial_weight_scale`	`force_load_balance`	`replicate_training_data`
`shuffle_training_data`	`single_node_mode`	`fast_mode`
`l1`	`l2`	`max_w2`
`input_dropout_ratio`	`hidden_dropout_ratios`	`loss`
`overwrite_with_best_model`	`missing_values_handling`	`average_activation`
`reproducible`	`export_weights_and_biases`	`elastic_averaging`
`elastic_averaging_moving_rate`	`elastic_averaging_regularization`