ipex-llm/docs/readthedocs/source/doc/DLlib/Overview/getting-started.md

DLLib Getting Start Guide

1. Creating dev environment

Scala project (maven & sbt)

• Maven

To use BigDL DLLib to build your own deep learning application, you can use maven to create your project and add bigdl-dllib to your dependency. Please add below code to your pom.xml to add BigDL DLLib as your dependency:

<dependency>
    <groupId>com.intel.analytics.bigdl</groupId>
    <artifactId>bigdl-dllib-spark_2.4.6</artifactId>
    <version>0.14.0</version>
</dependency>

• SBT

libraryDependencies += "com.intel.analytics.bigdl" % "bigdl-dllib-spark_2.4.6" % "0.14.0"

For more information about how to add BigDL dependency, please refer https://bigdl.readthedocs.io/en/latest/doc/UserGuide/scala.html#build-a-scala-project

IDE (Intelij)

Open up IntelliJ and click File => Open

Navigate to your project. If you have add BigDL DLLib as dependency in your pom.xml. The IDE will automatically download it from maven and you are able to run your application.

For more details about how to setup IDE for BigDL project, please refer https://bigdl-project.github.io/master/#ScalaUserGuide/install-build-src/#setup-ide

2. Code initialization

NNContext is the main entry for provisioning the dllib program on the underlying cluster (such as K8s or Hadoop cluster), or just on a single laptop.

It is recommended to initialize NNContext at the beginning of your program:

import com.intel.analytics.bigdl.dllib.NNContext
import com.intel.analytics.bigdl.dllib.keras.Model
import com.intel.analytics.bigdl.dllib.keras.models.Models
import com.intel.analytics.bigdl.dllib.keras.optimizers.Adam
import com.intel.analytics.bigdl.dllib.nn.ClassNLLCriterion
import com.intel.analytics.bigdl.dllib.utils.Shape
import com.intel.analytics.bigdl.dllib.keras.layers._
import com.intel.analytics.bigdl.numeric.NumericFloat
import org.apache.spark.ml.feature.VectorAssembler
import org.apache.spark.sql.SQLContext
import org.apache.spark.sql.functions._
import org.apache.spark.sql.types.DoubleType

val sc = NNContext.initNNContext("dllib_demo")

For more information about NNContext, please refer to NNContext

3. Distributed Data Loading

Using Spark Dataframe APIs

DLlib supports Spark Dataframes as the input to the distributed training, and as the input/output of the distributed inference. Consequently, the user can easily process large-scale dataset using Apache Spark, and directly apply AI models on the distributed (and possibly in-memory) Dataframes without data conversion or serialization

We used Pima Indians onset of diabetes as dataset for the demo. It's a standard machine learning dataset from the UCI Machine Learning repository. It describes patient medical record data for Pima Indians and whether they had an onset of diabetes within five years. The dataset can be download with:

wget https://raw.githubusercontent.com/jbrownlee/Datasets/master/pima-indians-diabetes.data.csv

We create Spark session so we can use Spark API to load and process the data

val spark = new SQLContext(sc)

Load the data into Spark DataFrame

val path = "pima-indians-diabetes.data.csv"
val df = spark.read.options(Map("inferSchema"->"true","delimiter"->",")).csv(path)
      .toDF("num_times_pregrant", "plasma_glucose", "blood_pressure", "skin_fold_thickness", "2-hour_insulin", "body_mass_index", "diabetes_pedigree_function", "age", "class")

4. Model Definition

Using Keras-like APIs

To define a model, you can use the Keras Style API.

val x1 = Input(Shape(8))
val dense1 = Dense(12, activation="relu").inputs(x1)
val dense2 = Dense(8, activation="relu").inputs(dense1)
val dense3 = Dense(2).inputs(dense2)
val dmodel = Model(x1, dense3)

After creating the model, you will have to decide which loss function to use in training.

Now you can use compile function of the model to set the loss function, optimization method.

dmodel.compile(optimizer = new Adam(),
  loss = ClassNLLCriterion())

Now the model is built and ready to train.

5. Distributed Model Training

Now you can use 'fit' begin the training, please set the feature columns and label columns. Model Evaluation can be performed periodically during a training. If the model accepts single input(eg. column feature1) and single output(eg. column label), please set the feature columns of the model as :

model.fit(x=dataframe, batchSize=4, nbEpoch = 2,
  featureCols = Array("feature1"), labelCols = Array("label"))

If the feature column for the model is a Spark ML Vector. Please assemble related columns into a Vector and pass it to the model. eg.

val assembler = new VectorAssembler()
  .setInputCols(Array("num_times_pregrant", "plasma_glucose", "blood_pressure", "skin_fold_thickness", "2-hour_insulin", "body_mass_index", "diabetes_pedigree_function", "age"))
  .setOutputCol("features")
val assembleredDF = assembler.transform(df)
val df2 = assembleredDF.withColumn("label",col("class").cast(DoubleType) + lit(1))

If your model accepts multiple inputs(eg. column f1, f2, f3), please set the features as below:

model.fit(x=dataframe, batchSize=4, nbEpoch = 2,
  featureCols = Array("f1", "f2", "f3"), labelCols = Array("label"))

If one of the inputs is a Spark ML Vector, please assemble it before pass the data to the model.

Similarly, if the model accepts multiple outputs(eg. column label1, label2), please set the label columns as below:

model.fit(x=dataframe, batchSize=4, nbEpoch = 2,
  featureCols = Array("f1", "f2", "f3"), labelCols = Array("label1", "label2"))

Then split it into traing part and validation part

val Array(trainDF, valDF) = df2.randomSplit(Array(0.8, 0.2))

The model is ready to train.

dmodel.fit(x=trainDF, batchSize=4, nbEpoch = 2,
  featureCols = Array("features"), labelCols = Array("label"), valX = valDF
)

6. Model saving and loading

When training is finished, you may need to save the final model for later use.

BigDL allows you to save your BigDL model on local filesystem, HDFS, or Amazon s3.

• save

val modelPath = "/tmp/demo/keras.model"
dmodel.saveModel(modelPath)

• load

val loadModel = Models.loadModel(modelPath)

val preDF2 = loadModel.predict(valDF, featureCols = Array("features"), predictionCol = "predict")

You may want to refer Save/Load

7. Distributed evaluation and inference

After training finishes, you can then use the trained model for prediction or evaluation.

• inference

dmodel.predict(trainDF, featureCols = Array("features"), predictionCol = "predict")

• evaluation

dmodel.evaluate(trainDF, batchSize = 4, featureCols = Array("features"),
  labelCols = Array("label"))

8. Checkpointing and resuming training

You can configure periodically taking snapshots of the model.

val cpPath = "/tmp/demo/cp"
dmodel.setCheckpoint(cpPath, overWrite=false)

You can also set overWrite to true to enable overwriting any existing snapshot files

After training stops, you can resume from any saved point. Choose one of the model snapshots to resume (saved in checkpoint path, details see Checkpointing). Use Models.loadModel to load the model snapshot into an model object.

val loadModel = Models.loadModel(path)

9. Monitor your training

• Tensorboard

BigDL provides a convenient way to monitor/visualize your training progress. It writes the statistics collected during training/validation. Saved summary can be viewed via TensorBoard.

In order to take effect, it needs to be called before fit.

dmodel.setTensorBoard("./", "dllib_demo")

For more details, please refer visulization

10. Transfer learning and finetuning

• freeze and trainable BigDL DLLib supports exclude some layers of model from training.

dmodel.freeze(layer_names)

Layers that match the given names will be freezed. If a layer is freezed, its parameters(weight/bias, if exists) are not changed in training process.

BigDL DLLib also support unFreeze operations. The parameters for the layers that match the given names will be trained(updated) in training process

dmodel.unFreeze(layer_names)

For more information, you may refer freeze

11. Hyperparameter tuning

• optimizer

DLLib supports a list of optimization methods. For more details, please refer optimization

• learning rate scheduler

DLLib supports a list of learning rate scheduler. For more details, please refer lr_scheduler

• batch size

DLLib supports set batch size during training and prediction. We can adjust the batch size to tune the model's accuracy.

• regularizer

DLLib supports a list of regularizers. For more details, please refer regularizer

• clipping

DLLib supports gradient clipping operations. For more details, please refer gradient_clip

12. Running program

You can run a bigdl-dllib program as a standard Spark program (running on either a local machine or a distributed cluster) as follows:

# Spark local mode
${BIGDL_HOME}/bin/spark-submit-with-dllib.sh \
  --master local[2] \
  --class class_name \
  jar_path

# Spark standalone mode
## ${SPARK_HOME}/sbin/start-master.sh
## check master URL from http://localhost:8080
${BIGDL_HOME}/bin/spark-submit-with-dllib.sh \
  --master spark://... \
  --executor-cores cores_per_executor \
  --total-executor-cores total_cores_for_the_job \
  --class class_name \
  jar_path

# Spark yarn client mode
${BIGDL_HOME}/bin/spark-submit-with-dllib.sh \
 --master yarn \
 --deploy-mode client \
 --executor-cores cores_per_executor \
 --num-executors executors_number \
 --class class_name \
 jar_path

# Spark yarn cluster mode
${BIGDL_HOME}/bin/spark-submit-with-dllib.sh \
 --master yarn \
 --deploy-mode cluster \
 --executor-cores cores_per_executor \
 --num-executors executors_number \
 --class class_name
 jar_path

For more detail about how to run BigDL scala application, please refer https://bigdl.readthedocs.io/en/latest/doc/UserGuide/scala.html