Deep Java Library
Since Camel 3.3
Only producer is supported
Deep Java Library
The Deep Java Library component is used to infer Deep Learning models from message exchanges data. This component uses Deep Java Library as underlying library.
In order to use the DJL component, Maven users will need to add the
following dependency to their pom.xml
:
pom.xml
<dependency>
<groupId>org.apache.camel</groupId>
<artifactId>camel-djl</artifactId>
<version>x.x.x</version>
<!-- use the same version as your Camel core version -->
</dependency>
The Deep Java Library component supports 2 options, which are listed below.
Name | Description | Default | Type |
---|---|---|---|
lazyStartProducer (producer) |
Whether the producer should be started lazy (on the first message). By starting lazy you can use this to allow CamelContext and routes to startup in situations where a producer may otherwise fail during starting and cause the route to fail being started. By deferring this startup to be lazy then the startup failure can be handled during routing messages via Camel’s routing error handlers. Beware that when the first message is processed then creating and starting the producer may take a little time and prolong the total processing time of the processing. |
false |
boolean |
basicPropertyBinding (advanced) |
Whether the component should use basic property binding (Camel 2.x) or the newer property binding with additional capabilities |
false |
boolean |
The DJL component only supports producer endpoints.
The Deep Java Library endpoint is configured using URI syntax:
djl:application
with the following path and query parameters:
Path Parameters (1 parameters):
Name | Description | Default | Type |
---|---|---|---|
application |
Required Application name |
String |
Query Parameters (6 parameters):
Name | Description | Default | Type |
---|---|---|---|
artifactId (producer) |
Model Artifact |
String |
|
lazyStartProducer (producer) |
Whether the producer should be started lazy (on the first message). By starting lazy you can use this to allow CamelContext and routes to startup in situations where a producer may otherwise fail during starting and cause the route to fail being started. By deferring this startup to be lazy then the startup failure can be handled during routing messages via Camel’s routing error handlers. Beware that when the first message is processed then creating and starting the producer may take a little time and prolong the total processing time of the processing. |
false |
boolean |
model (producer) |
Model |
String |
|
translator (producer) |
Translator |
String |
|
basicPropertyBinding (advanced) |
Whether the endpoint should use basic property binding (Camel 2.x) or the newer property binding with additional capabilities |
false |
boolean |
synchronous (advanced) |
Sets whether synchronous processing should be strictly used, or Camel is allowed to use asynchronous processing (if supported). |
false |
boolean |
Model Zoo
The following table contains supported models in the model zoo:
CV | Image Classification | Resnet image classification | cv/image_classification |
ai.djl.zoo:resnet:0.0.1 |
{layers=50, flavor=v1, dataset=cifar10} |
---|---|---|---|---|---|
CV |
Image Classification |
MLP image classification |
|
|
{dataset=mnist} |
CV |
Image Classification |
MLP image classification |
|
|
{dataset=mnist} |
CV |
Image Classification |
Resnet image classification |
|
|
{layers=18, flavor=v1, dataset=imagenet} |
CV |
Image Classification |
Resnet image classification |
|
|
{layers=50, flavor=v2, dataset=imagenet} |
CV |
Image Classification |
Resnet image classification |
|
|
{layers=152, flavor=v1d, dataset=imagenet} |
CV |
Image Classification |
Resnet image classification |
|
|
{layers=50, flavor=v1, dataset=cifar10} |
CV |
Image Classification |
Resnext image classification |
|
|
{layers=101, flavor=64x4d, dataset=imagenet} |
CV |
Image Classification |
Senet image classification |
|
|
{layers=154, dataset=imagenet} |
CV |
Image Classification |
Senet and Resnext image classification |
|
|
{layers=101, flavor=32x4d, dataset=imagenet} |
CV |
Image Classification |
Senet and Resnext image classification |
|
|
{layers=101, flavor=64x4d, dataset=imagenet} |
CV |
Image Classification |
Squeezenet image classification |
|
|
{flavor=1.0, dataset=imagenet} |
CV |
Object Detection |
Single Shot Detection for Object Detection |
|
|
{flavor=tiny, dataset=pikachu} |
CV |
Object Detection |
Single-shot object detection |
|
|
{size=512, backbone=resnet50, flavor=v1, dataset=voc} |
CV |
Object Detection |
Single-shot object detection |
|
|
{size=512, backbone=vgg16, flavor=atrous, dataset=coco} |
CV |
Object Detection |
Single-shot object detection |
|
|
{size=512, backbone=mobilenet1.0, dataset=voc} |
CV |
Object Detection |
Single-shot object detection |
|
|
{size=300, backbone=vgg16, flavor=atrous, dataset=voc} |
DJL Engine implementation
Because DJL is deep learning framework agnostic, you don’t have to make a choice between frameworks when creating your projects. You can switch frameworks at any point. To ensure the best performance, DJL also provides automatic CPU/GPU choice based on hardware configuration.
MxNet engine
You can pull the MXNet engine from the central Maven repository by including the following dependency:
<dependency>
<groupId>ai.djl.mxnet</groupId>
<artifactId>mxnet-engine</artifactId>
<version>0.4.0</version>
<scope>runtime</scope>
</dependency>
DJL offers an automatic option that will download the jars the first time you run DJL. It will automatically determine the appropriate jars for your system based on the platform and GPU support.
<dependency>
<groupId>ai.djl.mxnet</groupId>
<artifactId>mxnet-native-auto</artifactId>
<version>1.6.0</version>
<scope>runtime</scope>
</dependency>
More information about MxNet engine installation
PyTorch engine
You can pull the PyTorch engine from the central Maven repository by including the following dependency:
<dependency>
<groupId>ai.djl.mxnet</groupId>
<artifactId>pytorch-engine</artifactId>
<version>0.4.0</version>
<scope>runtime</scope>
</dependency>
DJL offers an automatic option that will download the jars the first time you run DJL. It will automatically determine the appropriate jars for your system based on the platform and GPU support.
<dependency>
<groupId>ai.djl.mxnet</groupId>
<artifactId>pytorch-native-auto</artifactId>
<version>1.4.0</version>
<scope>runtime</scope>
</dependency>
More information about PyTorch engine installation
Examples
MNIST image classification from file
from("file:/data/mnist/0/10.png")
.to("djl:cv/image_classification?artifactId=ai.djl.mxnet:mlp:0.0.1");
Object detection
from("file:/data/mnist/0/10.png")
.to("djl:cv/image_classification?artifactId=ai.djl.mxnet:mlp:0.0.1");
Custom deep learning model
// create deep learning model
Model model = Model.newInstance();
model.setBlock(new Mlp(28 * 28, 10, new int[]{128, 64}));
model.load(Paths.get(MODEL_DIR), MODEL_NAME);
// create translator for pre-processing and postprocessing
ImageClassificationTranslator.Builder builder = ImageClassificationTranslator.builder();
builder.setSynsetArtifactName("synset.txt");
builder.setPipeline(new Pipeline(new ToTensor()));
builder.optApplySoftmax(true);
ImageClassificationTranslator translator = new ImageClassificationTranslator(builder);
// Bind model and translator beans
context.getRegistry().bind("MyModel", model);
context.getRegistry().bind("MyTranslator", translator);
from("file:/data/mnist/0/10.png")
.to("djl:cv/image_classification?model=MyModel&translator=MyTranslator");