38  Saving and Loading

This tutorial shows how to save and load objects in spotpython. It is split into the following parts:

38.1 spotpython: Saving and Loading Optimization Experiments

In this section, we will show how results from spotpython can be saved and reloaded. Here, spotpython can be used as an optimizer. If spotpython is used as an optimizer, no dictionary of hyperparameters has be specified. The fun_control dictionary is sufficient.

import os
import pprint
from spotpython.utils.file import load_experiment
from spotpython.utils.file import get_experiment_filename
import numpy as np
from math import inf
from spotpython.spot import spot
from spotpython.utils.init import (
    fun_control_init,
    design_control_init,
    surrogate_control_init,
    optimizer_control_init)
from spotpython.fun.objectivefunctions import Analytical
fun = Analytical().fun_branin
fun_control = fun_control_init(
            PREFIX="branin",            
            lower = np.array([0, 0]),
            upper = np.array([10, 10]),
            fun_evals=8,
            fun_repeats=1,
            max_time=inf,
            noise=False,
            tolerance_x=0,
            ocba_delta=0,
            var_type=["num", "num"],
            infill_criterion="ei",
            n_points=1,
            seed=123,
            log_level=20,
            show_models=False,
            show_progress=True)
design_control = design_control_init(
            init_size=5,
            repeats=1)
surrogate_control = surrogate_control_init(
            model_fun_evals=10000,
            min_theta=-3,
            max_theta=3,
            n_theta=2,
            theta_init_zero=True,
            n_p=1,
            optim_p=False,
            var_type=["num", "num"],
            seed=124)
optimizer_control = optimizer_control_init(
            max_iter=1000,
            seed=125)
spot_tuner = spot.Spot(fun=fun,
            fun_control=fun_control,
            design_control=design_control,
            surrogate_control=surrogate_control,
            optimizer_control=optimizer_control)
spot_tuner.run()
PREFIX = fun_control["PREFIX"]
filename = get_experiment_filename(PREFIX)
spot_tuner.save_experiment(filename=filename)
print(f"filename: {filename}")
(spot_tuner_1, fun_control_1, design_control_1,
    surrogate_control_1, optimizer_control_1) = load_experiment(filename)

The progress of the original experiment is shown in Figure 38.1 and the reloaded experiment in Figure 38.2.

spot_tuner.plot_progress(log_y=True)
Figure 38.1 
spot_tuner_1.plot_progress(log_y=True)
Figure 38.2

The results from the original experiment are shown in Table 38.1 and the reloaded experiment in Table 38.2.

Table 38.1 
spot_tuner.print_results()
Table 38.2 
spot_tuner_1.print_results()

38.1.1 Getting the Tuned Hyperparameters

The tuned hyperparameters can be obtained as a dictionary with the following code. Since spotpython is used as an optimizer, the numerical levels of the hyperparameters are identical to the optimized values of the underlying optimization problem, here: the Branin function.

from spotpython.hyperparameters.values import get_tuned_hyperparameters
get_tuned_hyperparameters(spot_tuner=spot_tuner)
Summary: Saving and Loading Optimization Experiments
  • If spotpython is used as an optimizer (without an hyperparameter dictionary), experiments can be saved and reloaded with the save_experiment and load_experiment functions.
  • The tuned hyperparameters can be obtained with the get_tuned_hyperparameters function.

38.2 spotpython as a Hyperparameter Tuner

If spotpython is used as a hyperparameter tuner, in addition to the fun_control dictionary a core_model dictionary has to be specified. Furthermore, a data set has to be selected and added to the fun_control dictionary. Here, we will use the Diabetes data set.

38.2.1 The Diabetes Data Set

The hyperparameter tuning of a PyTorch Lightning network on the Diabetes data set is used as an example. The Diabetes data set is a PyTorch Dataset for regression, which originates from the scikit-learn package, see https://scikit-learn.org/stable/modules/generated/sklearn.datasets.load_diabetes.html#sklearn.datasets.load_diabetes.

Ten baseline variables, age, sex, body mass index, average blood pressure, and six blood serum measurements were obtained for each of n = 442 diabetes patients, as well as the response of interest, a quantitative measure of disease progression one year after baseline. The Diabetes data set is has the following properties:

  • Samples total: 442
  • Dimensionality: 10
  • Features: real, \(-.2 < x < .2\)
  • Targets: integer \(25 - 346\)
from spotpython.data.diabetes import Diabetes
data_set = 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 gen_design_table
from spotpython.spot import spot
from spotpython.utils.file import get_experiment_filename

PREFIX="604"
fun_control = fun_control_init(
    save_experiment=True,
    PREFIX=PREFIX,
    fun_evals=inf,
    max_time=1,
    data_set = data_set,
    core_model_name="light.regression.NNLinearRegressor",
    hyperdict=LightHyperDict,
    _L_in=10,
    _L_out=1)

fun = HyperLight().fun

from spotpython.hyperparameters.values import set_hyperparameter
set_hyperparameter(fun_control, "optimizer", [ "Adadelta", "Adam", "Adamax"])
set_hyperparameter(fun_control, "l1", [3,4])
set_hyperparameter(fun_control, "epochs", [3,5])
set_hyperparameter(fun_control, "batch_size", [4,11])
set_hyperparameter(fun_control, "dropout_prob", [0.0, 0.025])
set_hyperparameter(fun_control, "patience", [2,3])

design_control = design_control_init(init_size=10)

print(gen_design_table(fun_control))

In contrast to the default setttin, where sava_experiment is set to False, here the fun_control dictionary is initialized save_experiment=True. Alternatively, an existing fun_control dictionary can be updated with {"save_experiment": True} as shown in the following code.

fun_control.update({"save_experiment": True})

If save_experiment is set to True, the results of the hyperparameter tuning experiment are stored in a pickle file with the name PREFIX after the tuning is finished in the current directory.

Alternatively, the spot object and the corresponding dictionaries can be saved with the save_experiment method, which is part of the spot object. Therefore, the spot object has to be created as shown in the following code.

spot_tuner = spot.Spot(fun=fun,fun_control=fun_control, design_control=design_control)
spot_tuner.save_experiment(path="userExperiment", overwrite=False)

Here, we have added a path argument to specify the directory where the experiment is saved. The resulting pickle file can be copied to another directory or computer and reloaded with the load_experiment function. It can also be used for performing the tuning run. Here, we will execute the tuning run on the local machine, which can be done with the following code.

res = spot_tuner.run()

After the tuning run is finished, a pickle file with the name spot_604_experiment.pickle is stored in the local directory. This is a result of setting the save_experiment argument to True in the fun_control dictionary. We can load the experiment with the following code. Here, we have specified the PREFIX as an argument to the load_experiment function. Alternatively, the filename (PKL_NAME) can be used as an argument.

from spotpython.utils.file import load_experiment
(spot_tuner_1, fun_control_1, design_control_1,
    surrogate_control_1, optimizer_control_1) = load_experiment(PREFIX=PREFIX)

For comparison, the tuned hyperparameters of the original experiment are shown first:

get_tuned_hyperparameters(spot_tuner, fun_control)

Second, the tuned hyperparameters of the reloaded experiment are shown:

get_tuned_hyperparameters(spot_tuner_1, fun_control_1)

Note: The numerical levels of the hyperparameters are used as keys in the dictionary. If the fun_control dictionary is used, the names of the hyperparameters are used as keys in the dictionary.

get_tuned_hyperparameters(spot_tuner_1, fun_control_1)

Plot the progress of the original experiment are identical to the reloaded experiment.

spot_tuner.plot_progress()
Figure 38.3 
spot_tuner_1.plot_progress()
Figure 38.4
Summary: Saving and Loading Hyperparameter-Tuning Experiments
  • If spotpython is used as an hyperparameter tuner (with an hyperparameter dictionary), experiments can be saved and reloaded with the save_experiment and load_experiment functions.
  • The tuned hyperparameters can be obtained with the get_tuned_hyperparameters function.

38.3 Saving and Loading PyTorch Lightning Models

Section 38.1 and Section 38.2 explained how to save and load optimization and hyperparameter tuning experiments and how to get the tuned hyperparameters as a dictionary. This section shows how to save and load PyTorch Lightning models.

38.3.1 Get the Tuned Architecture

In contrast to the function get_tuned_hyperparameters, the function get_tuned_architecture returns the tuned architecture of the model as a dictionary. Here, the transformations are already applied to the numerical levels of the hyperparameters and the encoding (and types) are the original types of the hyperparameters used by the model. Important: The config dictionary from get_tuned_architecture can be passed to the model without any modifications.

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

After getting the tuned architecture, the model can be created and tested with the following code.

from spotpython.light.testmodel import test_model
test_model(config, fun_control)

38.3.2 Load a Model from Checkpoint

The method load_light_from_checkpoint loads a model from a checkpoint file. Important: The model has to be trained before the checkpoint is loaded. As shown here, loading a model with trained weights is possible, but requires two steps:

  1. The model weights have to be learned using test_model. The test_model method writes a checkpoint file.
  2. The model has to be loaded from the checkpoint file.

38.3.2.1 Details About the load_light_from_checkpoint Method

  • The test_model method saves the last checkpoint to a file using the following code:
ModelCheckpoint(
    dirpath=os.path.join(fun_control["CHECKPOINT_PATH"], config_id), save_last=True
),

The filename of the last checkpoint has a specific structure:

  • A config_id is generated from the config dictionary. It does not use a timestamp. This differs from the config id generated in cvmodel.py and trainmodel.py, which provide time information for the TensorBoard logging.
  • Furthermore, the postfix _TEST is added to the config_id to indicate that the model is tested.
  • For example: runs/saved_models/16_16_64_LeakyReLU_Adadelta_0.0014_8.5895_8_False_kaiming_uniform_TEST/last.ckpt
from spotpython.light.loadmodel import load_light_from_checkpoint
model_loaded = load_light_from_checkpoint(config, fun_control)
vars(model_loaded)
import torch
torch.save(model_loaded, "model.pt")
mymodel = torch.load("model.pt")
# show all attributes of the model
vars(mymodel)

38.4 Converting a Lightning Model to a Plain Torch Model

38.4.1 The Function get_removed_attributes_and_base_net

spotpython provides a function to covert a PyTorch Lightning model to a plain PyTorch model. The function get_removed_attributes_and_base_net returns a tuple with the removed attributes and the base net. The base net is a plain PyTorch model. The removed attributes are the attributes of the PyTorch Lightning model that are not part of the base net.

This conversion can be reverted.

import numpy as np
import torch
from spotpython.utils.device import getDevice
from torch.utils.data import random_split
from spotpython.utils.classes import get_removed_attributes_and_base_net
from spotpython.hyperparameters.optimizer import optimizer_handler
removed_attributes, torch_net = get_removed_attributes_and_base_net(net=mymodel)
print(removed_attributes)
print(torch_net)

38.4.2 An Example how to use the Plain Torch Net

import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader, TensorDataset
from sklearn.datasets import load_diabetes
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
import matplotlib.pyplot as plt

# Load the Diabetes dataset from sklearn
diabetes = load_diabetes()
X = diabetes.data
y = diabetes.target

# Split the dataset into training and testing sets
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Scale the features
scaler = StandardScaler()
X_train = scaler.fit_transform(X_train)
X_test = scaler.transform(X_test)

# Convert the data to PyTorch tensors
X_train_tensor = torch.tensor(X_train, dtype=torch.float32)
y_train_tensor = torch.tensor(y_train, dtype=torch.float32)
X_test_tensor = torch.tensor(X_test, dtype=torch.float32)
y_test_tensor = torch.tensor(y_test, dtype=torch.float32)

# Create a PyTorch dataset
train_dataset = TensorDataset(X_train_tensor, y_train_tensor)
test_dataset = TensorDataset(X_test_tensor, y_test_tensor)

# Create a PyTorch dataloader
batch_size = 32
train_dataloader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_dataloader = DataLoader(test_dataset, batch_size=batch_size)

torch_net.to(getDevice("cpu"))

# train the net
criterion = nn.MSELoss()
optimizer = optim.Adam(torch_net.parameters(), lr=0.01)
n_epochs = 100
losses = []
for epoch in range(n_epochs):
    for inputs, targets in train_dataloader:
        targets = targets.view(-1, 1)
        optimizer.zero_grad()
        outputs = torch_net(inputs)
        loss = criterion(outputs, targets)
        losses.append(loss.item())
        loss.backward()
        optimizer.step()
# visualize the network training
plt.plot(losses)
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.show()