Nessa palestra falei sobre a evolução do tensorflow mobile e como utilizei para detectar fotos borradas em um app android

1. Gustavo Monteiro
Mobile Developer
Utilizando TensorFlow Mobile
e seus desafios

2. TensorFlow For Poets
https://codelabs.developers.google.com/
codelabs/tensorflow-for-poets/
https://github.com/googlecodelabs/
tensorflow-for-poets-2

3. • Inception
• MobileNet
• Tamanho de imagens: 128px, 160px,
192px, 224px
• Tamanho Relativo do modelo: 1.0, 0.75,
0.50, 0.25
TensorFlow For Poets

4. TensorFlow For Poets
ARCHITECTURE=“mobilenet_0.50_224”
python -m scripts.retrain
--bottleneck_dir=bottlenecks
--how_many_training_steps=500
--model_dir=models/
--summaries_dir=training_summaries/"${ARCHITECTURE}"
--output_graph=retrained_graph.pb
--output_labels=retrained_labels.txt
--architecture="${ARCHITECTURE}"
--image_dir=flower_photos

5. TensorFlow For Poets
python -m scripts.label_image
--graph=retrained_graph.pb
--image=flower_photos/daisy/21652746_cc379e0eea_m.jpg

6. Modelo: fotos borradas
• "ahhh, mas dá pra fazer com OpenCV"
• Pack de 11mil fotos (uso liberado)
• Script python para borrar as fotos
• 6h treinando
• Labels: Blurred, Common

8. Evolução TensorFlow
Lib TensorFlow:
• Tamanho Modelo: 90mb
• Tempo de Inferência: 9s

9. Pete Warden
https://petewarden.com/2016/09/27/
tensorflow-for-mobile-poets/
Lead of TensorFlow Mobile
Otimizações, tutoriais e artigos

10. Evolução TensorFlow
Lib TensorFlow Lite:
• Tamanho Modelo: 3mb
• Tempo de Inferência: 150ms
Lib TensorFlow:
• Tamanho Modelo: 90mb
• Tempo de Inferência: 9s

12. Android
• https://github.com/tensorflow/tensorflow/
tree/master/tensorflow/examples/android
• Lib .jar (gradle) (0.1)
• Lib .so

13. Carregando as imagens
• Tamanho total : 1s ~
Bitmap image =
BitmapFactory.decodeFile(photo);

14. Stream photos = imageList.stream().parallel();
if (photos.isParallel()) {
photos.forEach((Consumer<String>) photo ->
calculatePhoto(photo, loadedImages));
}
Carregando as imagens

16. • Tamanho total : 1s ~
• SampleSize 2x: 750ms ~
• Calculated SampleSize: 300ms ~
Carregando as imagens

17. BitmapFactory.Options options = new
BitmapFactory.Options();
options.inJustDecodeBounds = true;
BitmapFactory.decodeFile(photo, options);
options.inSampleSize =
calculateInSampleSize(options, INPUT_SIZE,
INPUT_SIZE);
options.inJustDecodeBounds = false;
Bitmap image =
BitmapFactory.decodeFile(photo, options);

18. public static int calculateInSampleSize(
BitmapFactory.Options options, int reqWidth, int reqHeight) {
final int height = options.outHeight;
final int width = options.outWidth;
int inSampleSize = 1;
if (height > reqHeight || width > reqWidth) {
while ((height / inSampleSize) >= reqHeight
&& (width / inSampleSize) >= reqWidth) {
inSampleSize *= 2;
}
}
return inSampleSize;
}

19. for (Bitmap photo : imagesPath) {
Bitmap scaledBitmap =
Bitmap.createScaledBitmap(photo,
INPUT_SIZE, INPUT_SIZE, false);
List<String> results =
classifier.classifyFrame(scaledBitmap);
if (results.get(1).contains("common")) {
imagesToRemove.add(photo);
} else {
resultList.add(results);
}
}

21. Measured time: 386ms :: Decoded Bitmap Single
Measured time: 502ms :: Decoded Bitmap Single
Measured time: 465ms :: Decoded Bitmap Single
Measured time: 1269ms :: Decoded Bitmap Single
Medindo Desempenho

22. Measured time: 2ms :: Scaled Bitmap Single
Measured time: 109ms :: Recognized Image Time Single
Measured time: 1ms :: Scaled Bitmap Single
Measured time: 110ms :: Recognized Image Time Single
Measured time: 2ms :: Scaled Bitmap Single
Measured time: 110ms :: Recognized Image Time Single
Medindo Desempenho

23. Measured time: 74130ms :: Total time:
Decode, scale and recognize images.
Total images: 429
Timecost to put values into ByteBuffer: 13ms
Timecost to run model inference: 96ms
Medindo Desempenho

24. Obrigado
Twitter: @OGustavoDev
GitHub: Pteroto
Android Dev BR:
Gustavo Monteiro