Day # - 163 - Faster ML Development with TensorFlow

Key Summary

• ML Models represented as Data Structure or Program
• Data Structure - Graph (Deferred Execution - Define and Run)
• Model - Program - Python Code (Eager Execution - Define by Run)
• Tensorflow is both Data Structure and Model
• Has both CPU and GPU Support
• Model is Data Structure - Easy to Serialize and De-Serialize it and Deploy on Devices (Mobile, TPU, XLA)
• Because model is data structure it is not tied down to language
• Distributed Training to train on large amount of data
• Model Structures - Convolution, AvgPool, MaxPool, Concat, Dropout, FullyConnected, Softmax

Model Structures - Static Vs Dynamic

• Traditional RNN vs Dynamic Models
• Model whose structure cannot be easily defined by graph
• Straightforward with Eager using Native Python control flow

	#Example 1 - Symolic
	import tensorflow as tf
	x = tf.constant(10.0)
	w = tf.constant(4.0)
	b = tf.constant(2.0)
	y = tf.multiply(x,w)
	print(y)
	z = tf.add(y,b)
	#you need to create session to perform actual computation
	sess = tf.Session()
	print(sess.run(z))
	#Model as Data Structure
	#During Session Computes and returns result
	#Example 2 - Eager
	#Restart your Spyder kernel to try this setting
	import tensorflow as tf
	import tensorflow.contrib.eager as tfe
	tfe.enable_eager_execution()
	x = tf.constant(10.0)
	w = tf.constant(4.0)
	b = tf.constant(2.0)
	y = tf.multiply(x,w)
	print(y)
	z = tf.add(y,b)
	print(z)
	#Example 3 - Debugging
	import tensorflow as tf
	from tensorflow.python import debug as tfdebug
	x = tf.constant(10.0)
	w = tf.constant(4.0)
	b = tf.constant(2.0)
	y = tf.multiply(x,w)
	print(y)
	z = tf.add(y,b)
	#you need to create session to perform actual computation
	sess = tf.Session()
	sess = tfdebug.LocalCLIDebugWrapperSession(sess)
	sess.run(tf.global_variables_initializer())
	sess.run(y)

view raw tf.py hosted with ❤ by GitHub

Happy Mastering DL!!!