Although the TensorFlow core is C/C++, the most famous API is Python (as you can see in Siraj or sintdex tutorials), but recently TensorFlow released a Java version for the API.
TensorFlow has a small tutorial about the Java API and a single more complex example to classify image. On this small tutorial the link to the tensorflow JAR is broken. I found a working JAR in this URL. The upstream tutorial was already corrected as you can see in github.
There's also the labeled image example, which is very simple as well, but a few additional steps must be taken:
- Download the content from this URL and unzip it. It is the labeled dataset
- Download the LabelImage class from github. When saving it locally remember to remove the package declaration.
- Having the tensorflow JAR you downloaded previously and the native library, you can compile and run the class passing the directory where you unzipped the labeled dataset and a test image.
In the example above you can see it found that the image contained a "Zebra". There are 1001 labeled images on the downloaded dataset and it usually labels the image correctly.
Maven
A few days ago the tensorflow artifacts was added to maven central repository. To use it you just need the dependency org.tensorflow:tensorflow:1.1.0-rc1. Seriously, no other configuration is required, you just need to add this dependency and everything should work. A test sample maven application can be found at my github.
You will find a simple JavaFX Application that allows you to select a file that will be analyzed by the same code used in the LabelImage official example. See some screenshot:
To run this application locally you must have maven and Java. Then download the dataset and the labels and modify the PATH_TO_DATA_DIR in App class to the path were you extracted it. Finally, just run:
mvn -q compile exec:java
And it should open the application. All my tests were made on a linux machine, hence I don't know if it would work on Windows.
Conclusion
TensorFlow Java API is very easy to "install" and use. I would say that it is easier than python installation!
Nenhum comentário:
Postar um comentário