Train and evaluate your model
One of the most distinctive aspects of IGNNITION is that the high-level definition of the architecture of the GNN. This implies that when designing our GNN, we are not concerned about many aspects that affect the training process (e.g., optimizers, losses, training duration or simply the location of the datasets to be used). Nevertheless, all this information must be specified, for which we design a specific configuration file containing all this information. We shall now review first the following aspects.
Create the model
By making the call from below, IGNNITION is creating a model with all the information of the model_description.yaml file as well as the different training options specified. For this, the user must specify the path to the model directory where we can find the model_description.yaml file, and the training_options.yaml file.
model = ignnition.create_model(model_directory = './MY_EXAMPLE')
Create the computational graph
One of the main problems that any Machine Learning Engineer has faced is to properly debug the models. Machine learning often performs very well, but we are rarely capable of having proper insight into its inner working. This makes it very challenging and time-consuming to fix / troubleshoot an ML model.
For this, IGNNITION incorporates a debugging system that is based on producing a simplified computational graph of the model. For this, the user should make the following call over the previously created model
An important consideration is that this functionality requires a valid definition of the training dataset or of the predicted dataset. Thus, the user should specify at least one of them in the train_options.yaml, just as we show below:
# PATHS train_dataset: <PATH TO YOUR DATA>
This will create a directory named “computational_graph”, in the corresponding path indicated in the “train_options.yaml” file. We further extend on how to visualize or interpret the output of this operation at debugging assistant.
Train and validation
The main goal of IGNNITION is precisely to be able to train a GNN model easily. For this, the user must only make the following call:
This call proceeds to train the GNN specified in the model_description.yaml file. Additionally, a process of validation will be performed throughout the training phase, so as to provide better insight into the model performance. The behavior during the training phase is controlled by the train_options.yaml file, specified in section 2. Configuration file.
Furthermore, this operation will create (if it doesn’t exist already) a directory called “CheckPoints” in the specified path. In this directory, a new directory will be created corresponding to the experiment run (indexed by date). In it, the corresponding checkpoints will be stored, as well as the “events.out.tfevents.XXX” file, which contains the tensorboard information of the training metrics specified. Similarly, as before, the user can visualize this information by running
tensorboard --logdir <PATH TO THE "events.out.tfevents.XXX" FILE>
Then, by visiting localhost:6006, the user can analyze the different statistics produced during the training phase to evaluate the model (e.g., loss, mean absolute error, mean relative error, R2…)
We also incorporated a very useful function that allows the user to evaluate by obtaining performance metrics of a previously trained model. More specifically, the Evaluate functionality takes as input the specified validation dataset –or the array of samples passe. Then, after loading the indicated model, it will make the corresponding predictions for each of the samples, and then will compute the performance metrics of these predictions with respect to the true label found in the dataset.
For this, clearly, the user will have to indicate the path where the trained model can be found, as well as the metrics that we want IGNNITION to compute. All this information will be encoded in the train_options.yaml in the following fields:
validation_dataset: <PATH> load_model_path: <PATH>
More information regarding these fields can be found in the section 2. Configuration file. Moreover, to run this functionality, the user must only make the following call, which will return the result of the aforementioned metrics.
IGNNITION also allows making predictions over a previously pre-trained GNN. To do so, we provide two different alternative procedures which adapt to most scenarios: #### Feeding a dataset, First of all, IGNNITION provides the possibility of making predictions over a prediction dataset. For this, the user must define 2 special fields in the “train_options.yaml” file, which are:
predict_dataset: <PATH> load_model_path: <PATH>
In these fields, we can specify the dataset that we aim to predict, and the location of the checkpoint of the model that we need to restore, to later be used for the predicting phase. See more details on how to fill these fields in 2. Configuration file. Then, IGNNITION will compute the corresponding prediction of each of the samples of the prediction dataset. Moreover, to run this functionality, the user must only make the following call, which will return all the computed predictions.
In some cases, it may be useful to limit IGNNITION to compute the predictions over the first n samples only. To do so, simply pass this information during the Python call, as follows:
model.predict(num_predictions = n)
Feeding an array of samples
The second alternative is based on the idea of passing an array of samples during the Python call to the predicted functionality. This can be very useful, for instance, in Reinforcement Learning applications. To do so, the user must still provide a valid path to the checkpoint where the model has been stored –unless this operation is preceded by a train operator.
Afterward, simply make the following Python call:
In this case, my_samples is a simple Python array containing all the samples that we want to obtain predictions of. The syntax of these samples should be the same as in the dataset (see Generate your dataset for more details). A similar procedure as the one presented before can be followed to limit the predictions to the first n samples:
model.predict(prediction_samples= my_samples, num_predictions = n)
In this section, we review in-depth the content of the train_options.yaml file, which will contain all the configuration parameters that ultimately define the behavior of the specific functionality executed by the user. We must note that this file must be written in YAML format, which allows a very intuitive definition of all the possible fields in the form of KEY: VALUE.
Definition of the paths
At this point, we must provide the different paths which IGNNITION will use to locate the information and store the results of its execution. For this, the user must fill the following fields in the train_options.ini file. Let us note that all these paths can either be absolute paths or relative paths from the directory of this file.
Path to the training dataset
Indicate the path pointing to the training dataset, used by the train and validation functionality.
Path to the validation dataset
Indicate the path pointing to the validation/evaluation dataset, that will be used by the train and validation functionality, as well as the evaluation functionality.
Path to the prediction dataset
Defines the path to the prediction dataset, used by the predict functionality. Notice that this field needs only to be specified in the case that a predict functionality is executed. Otherwise, it will be ignored.
Load model path
Sometimes we might wish to use a previous checkpoint as a starting point for our training process (e.g., for evaluation functionality or for predicting). For this, the user can specify the path to such checkpoints, and IGNNITION will use it automatically.
Path where the Checkpoint and logs directory will be created when executing the train and validate functionality.
Additional file path
Path to an additional python file that may contain implementation of specific functions, such as the implementation of a certain metric or of a certain loss function
Path to the model description file
In this case, IGNNITION assumes that the definition of the GNN –through the model description file– is present in the very same directory as the train_options.yaml file itself. Hence, there is no need to specify anything at all regarding this file.
Model training parameters
Name of the loss function to use for the training of the model. It can be a name from tf.keras.losses library or a custom function that can be defined by the user.
Definition of the optimizer which follows the same syntax as the tf.keras.optimizers library. Thus the user must use the exact name used in this library to reference it.
optimizer: type: Adam
Following the documentation of the tf.keras.optimizers library, one can for instance define different attributes that model the inner working of the desired optimizer, in this case, ADAM. To do so, we simply include, just like the type attribute, any other parameter included in the aforementioned library that accepts this optimizer. Note that if no parameters are defined, IGNNITION will use the default values defined in the TensorFlow library. For illustrative purposes, however, let us suppose we want to change the beta_1 value to 0.9 and the beta_2 value to 0.9 also. This can be done as follows:
optimizer: type: Adam beta_1: 0.9 beta_2: 0.999
Use of schedules
Finally, we consider the case in which we want to define a schedule to be used with our optimizer. For this, again, we follow the syntax of the library tf.keras.optimizers.schedules. Thus, we must only create a parameter learning_rate, just as we would with the TensorFlow library, and pass to it the definition of the scheduler. This definition follows the same principle as before. Reference the schedule type using a valid name from the previously mentioned library, and also include any other desired parameter supported in such library (otherwise, IGNNITION will use the default values). Below we show a simple example defining an exponential decay schedule:
optimizer: type: Adam beta_1: 0.9 beta_2: 0.999 learning_rate: # defines the schedule here type: ExponentialDecay initial_learning_rate: 0.001 decay_steps: 80000 decay_rate: 0.6
Metrics define the list of metric criteria that we want to use to evaluate our GNN model. These metrics will be plotted during the training and validation phase.
For this definition, the user may specify in this list any name supported by the library tf.keras.metrics. Additionally, the user can define a custom metric by indicating its name, and then define a function with this very same name in the addition file.
IGNNITION also allows the user to customize basically any training option. For this, in this last part of the file, the user must specify the following fields. We recommend to copy-past the provided examples, and change only the desired fields -if any-.
Specify the batch_size in order to internally execute the mini-batch algorithm.
Number of epochs
Specify the number of epochs that the algorithm must run before termination.
This is an optional parameter that defines the number of elements that form each of the epochs (using a natural number). Note that if this is not specified, IGNNITION will consider the whole dataset as a single epoch. This option is useful if the dataset is very big, as we must recall that validation is only carried out after each of the epochs.
True / False to indicate if the training dataset should be shuffled.
True / False to indicate if the evaluation dataset should be shuffled.
Specify the number of evaluation samples to be used for the evaluation of our GNN.
Number of epochs after between validations.
Natural number indicating the number of checkpoints that we want to keep. Note that the system will automatically keep the best \(k\) checkpoints in terms of the loss.
When defining a model, we can either not use any normalization at all, define a normalization function that will be applied to all the datasets, or use batch normalization. This batch normalization will apply the same normalization function for all the elements of a single batch respectively. So far, IGNNITION supports the use of mean and max normalization.
When using an existing checkpoint as starting point of our GNN, it might be desirable to adapt also the initial epoch number. This is due to the fact that such value has implications on the learning rate (which normally gets smaller as the training advances). To do so, the user can (optionally) define the initial epoch to start the training by indicating its number (e.g., 100), which by default will take the value 0.