Day #296 - TensorflowLite

TensorflowLite models run in gmail, google photos, google assistant, google photos etc..

Advantages

• Low Latency
• No Data connection required
• On device sensors

Key Points

• On Device ML on many platforms
• Tensorflow model saved in graph format
• Converted to Lite format

TFLite

• Model compression
• Quantization
• Optimized SIMD Kernels
• Converter to generate model. Interpreter to run models

Benefits

• Cross-Platform deployment
• Inference speed increases
• Binary size reduction
• Hardware acceleration roadmap

#Build and save keras model
model = build_your_model()
tf.keras.experimental.export_saved_model(model,saved_model_dir)

#convert keras to tensorflow lite model
converter = tf.lite.TFLiteConverter.from_saved_models(saved_model_dir)
#To experiment new feature
converter.experimental_new_converter = True
tflite_model = converter.convert()

Link1, Link2

Improve Performance of models

• Reduce precision of weights (16 bit instead of 32 bit precision) - Quantization
• Pruning - Remove connections during training
• Op Kernels - ARM
• Delegates - GPU (Run on Specialized hardware)

Tensorflow Lite on Micro-controllers is an impressive move. More than mobiles, cross platforms this is a very impressive step.

Happy Learning!!!