• Python 3.5
• Tesnorflow 1.4

• quantize the model: https://www.tensorflow.org/performance/quantization
• can you serve the model faster using different compilation techniques?

Need some random sample data to train the model as well

https://www.kaggle.com/c/imagenet-object-detection-challenge/data

https://cloud.google.com/compute/docs/load-balancing/network/example

Tests are based on importing tensorflow and loading the weights for imagenet with different hyper parameters

import tensorflow as tf
mobilenet = tf.keras.applications.mobilenet

• Loading mobilenet with alpha @ 0.25

  • wts = mobilenet.MobileNet(weights='imagenet', alpha = 0.25)
  • Memory 138.4MB

• Loading mobilenet with alpha @ 1.0

  • wts = mobilenet.MobileNet(weights='imagenet', alpha = 1.0)
  • Memory 157.7MB

• Loading mobilenet with alpha @ 0.25, no top and image size = 128 by 128

  • wts = mobilenet.MobileNet(weights='imagenet', include_top = False, alpha = 0.25, input_shape = (128, 128, 3))
  • Memory: 137.6MB

nohup ~/miniconda/envs/trxsfr-learn-web/bin/python ~/trxsfr-learning-web-app/app.py &