49  Hyperparameter Tuning with spotpython and PyTorch Lightning Using a CondNet Model

• We use the Diabetes dataset to illustrate the hyperparameter tuning process of a CondNet model using the spotpython package.
• The CondNet model is a conditional neural network that can be used to model conditional distributions [LINK].

from spotpython.data.diabetes import Diabetes
from spotpython.hyperdict.light_hyper_dict import LightHyperDict
from spotpython.fun.hyperlight import HyperLight
from spotpython.utils.init import (fun_control_init, surrogate_control_init, design_control_init)
from spotpython.utils.eda import print_exp_table
from spotpython.spot import Spot
from spotpython.utils.file import get_experiment_filename
from math import inf
from spotpython.hyperparameters.values import set_hyperparameter

PREFIX="CondNet_01"

data_set = Diabetes()
input_dim = 10
output_dim = 1
cond_dim = 2

fun_control = fun_control_init(
    PREFIX=PREFIX,
    fun_evals=inf,
    max_time=1,
    data_set = data_set,
    core_model_name="light.regression.NNCondNetRegressor",
    hyperdict=LightHyperDict,
    _L_in=input_dim - cond_dim,
    _L_out=1,
    _L_cond=cond_dim,)

fun = HyperLight().fun


set_hyperparameter(fun_control, "optimizer", [ "Adadelta", "Adam", "Adamax"])
set_hyperparameter(fun_control, "l1", [3,4])
set_hyperparameter(fun_control, "epochs", [3,7])
set_hyperparameter(fun_control, "batch_size", [4,5])
set_hyperparameter(fun_control, "dropout_prob", [0.0, 0.025])
set_hyperparameter(fun_control, "patience", [2,3])
set_hyperparameter(fun_control, "lr_mult", [0.1, 20.0])

design_control = design_control_init(init_size=10)

print_exp_table(fun_control)

module_name: light
submodule_name: regression
model_name: NNCondNetRegressor
| name           | type   | default   |   lower |   upper | transform             |
|----------------|--------|-----------|---------|---------|-----------------------|
| l1             | int    | 3         |     3   |   4     | transform_power_2_int |
| epochs         | int    | 4         |     3   |   7     | transform_power_2_int |
| batch_size     | int    | 4         |     4   |   5     | transform_power_2_int |
| act_fn         | factor | ReLU      |     0   |   5     | None                  |
| optimizer      | factor | SGD       |     0   |   2     | None                  |
| dropout_prob   | float  | 0.01      |     0   |   0.025 | None                  |
| lr_mult        | float  | 1.0       |     0.1 |  20     | None                  |
| patience       | int    | 2         |     2   |   3     | transform_power_2_int |
| batch_norm     | factor | 0         |     0   |   1     | None                  |
| initialization | factor | Default   |     0   |   4     | None                  |

spot_tuner = Spot(fun=fun,fun_control=fun_control, design_control=design_control)
res = spot_tuner.run()

train_model result: {'val_loss': 24159.23828125, 'hp_metric': 24159.23828125}

train_model result: {'val_loss': 23455.20703125, 'hp_metric': 23455.20703125}

train_model result: {'val_loss': 43761.7265625, 'hp_metric': 43761.7265625}

train_model result: {'val_loss': 24069.580078125, 'hp_metric': 24069.580078125}

train_model result: {'val_loss': 22378.6015625, 'hp_metric': 22378.6015625}

train_model result: {'val_loss': 23910.111328125, 'hp_metric': 23910.111328125}

train_model result: {'val_loss': 23796.3828125, 'hp_metric': 23796.3828125}

train_model result: {'val_loss': 3585.37548828125, 'hp_metric': 3585.37548828125}

train_model result: {'val_loss': 22366.2421875, 'hp_metric': 22366.2421875}
train_model result: {'val_loss': 22662.296875, 'hp_metric': 22662.296875}

train_model result: {'val_loss': 4395.52978515625, 'hp_metric': 4395.52978515625}
spotpython tuning: 3585.37548828125 [#---------] 8.73% 

train_model result: {'val_loss': 7720.5400390625, 'hp_metric': 7720.5400390625}
spotpython tuning: 3585.37548828125 [#---------] 12.52% 

train_model result: {'val_loss': 3914.1572265625, 'hp_metric': 3914.1572265625}
spotpython tuning: 3585.37548828125 [##--------] 18.34% 

train_model result: {'val_loss': 2921.1494140625, 'hp_metric': 2921.1494140625}
spotpython tuning: 2921.1494140625 [##--------] 24.73% 

train_model result: {'val_loss': 3689.0537109375, 'hp_metric': 3689.0537109375}
spotpython tuning: 2921.1494140625 [###-------] 31.34% 

train_model result: {'val_loss': 20301.73828125, 'hp_metric': 20301.73828125}
spotpython tuning: 2921.1494140625 [######----] 61.34% 

train_model result: {'val_loss': 20522.70703125, 'hp_metric': 20522.70703125}
spotpython tuning: 2921.1494140625 [#########-] 90.31% 

train_model result: {'val_loss': 8519.857421875, 'hp_metric': 8519.857421875}
spotpython tuning: 2921.1494140625 [##########] 96.69% 

train_model result: {'val_loss': 4276.14111328125, 'hp_metric': 4276.14111328125}
spotpython tuning: 2921.1494140625 [##########] 100.00% Done...

Experiment saved to CondNet_01_res.pkl

49.1 Looking at the Results

49.1.1 Tuning Progress

After the hyperparameter tuning run is finished, the progress of the hyperparameter tuning can be visualized with spotpython’s method plot_progress. The black points represent the performace values (score or metric) of hyperparameter configurations from the initial design, whereas the red points represents the hyperparameter configurations found by the surrogate model based optimization.

spot_tuner.plot_progress()

49.1.2 Tuned Hyperparameters and Their Importance

Results can be printed in tabular form.

from spotpython.utils.eda import print_res_table
print_res_table(spot_tuner)

| name           | type   | default   |   lower |   upper | tuned                | transform             |   importance | stars   |
|----------------|--------|-----------|---------|---------|----------------------|-----------------------|--------------|---------|
| l1             | int    | 3         |     3.0 |     4.0 | 3.0                  | transform_power_2_int |         0.01 |         |
| epochs         | int    | 4         |     3.0 |     7.0 | 7.0                  | transform_power_2_int |         0.27 | .       |
| batch_size     | int    | 4         |     4.0 |     5.0 | 4.0                  | transform_power_2_int |         0.01 |         |
| act_fn         | factor | ReLU      |     0.0 |     5.0 | Swish                | None                  |         0.14 | .       |
| optimizer      | factor | SGD       |     0.0 |     2.0 | Adadelta             | None                  |         0.01 |         |
| dropout_prob   | float  | 0.01      |     0.0 |   0.025 | 0.006930970877686917 | None                  |         0.01 |         |
| lr_mult        | float  | 1.0       |     0.1 |    20.0 | 4.520416098796976    | None                  |         0.01 |         |
| patience       | int    | 2         |     2.0 |     3.0 | 2.0                  | transform_power_2_int |         0.01 |         |
| batch_norm     | factor | 0         |     0.0 |     1.0 | 1                    | None                  |         0.01 |         |
| initialization | factor | Default   |     0.0 |     4.0 | kaiming_uniform      | None                  |       100.00 | ***     |

A histogram can be used to visualize the most important hyperparameters.

spot_tuner.plot_importance(threshold=1.0)

spot_tuner.plot_important_hyperparameter_contour(max_imp=3)

l1:  0.005150865785137997
epochs:  0.2663700173362081
batch_size:  0.005150865785137997
act_fn:  0.142064953118143
optimizer:  0.005150865785137997
dropout_prob:  0.005150865785137997
lr_mult:  0.005150865785137997
patience:  0.005150865785137997
batch_norm:  0.005150865785137997
initialization:  100.0

49.1.3 Get the Tuned Architecture

import pprint
from spotpython.hyperparameters.values import get_tuned_architecture
config = get_tuned_architecture(spot_tuner)
pprint.pprint(config)

{'act_fn': Swish(),
 'batch_norm': True,
 'batch_size': 16,
 'dropout_prob': 0.006930970877686917,
 'epochs': 128,
 'initialization': 'kaiming_uniform',
 'l1': 8,
 'lr_mult': 4.520416098796976,
 'optimizer': 'Adadelta',
 'patience': 4}