Consider a system which figure out by themselves what features to pay attention and mark for feature vectors is termed as “Representation” ML-based systems. Deep learning systems are one type of representation systems. In other words, Deep Learning Algorithms learn what features(characteristics/attributes) requires more attention and each features are represented by data points (Vector). Neural Networks is most common and de facto class of deep learning algorithms. Neural networks help find unknown patterns in massive data sets. "Neurons" are most basic building blocks that actually “learn”. Below image shows high level overview where does deep learning algorithm fits.
Deep Learning(Layers of inter connected neural networks) based Binary Classifier(Source: PluralSight) |
Deep learning is subfield/subset of Machine learning. Machine learning blossoms Artificial intelligence with bang. However, Deep learning giving birth of limitless growth of AI. The very way to visualise how Deep learning influenced AI development in big way - Google Deep learning program ApphaGo beat GO Master Lee Sedol.
What is TensorFlow and TensorBoard - Deep learning de facto library and dashboard
Official doc of TensorFlow states it is an open source software library for numerical computation using data flow graphs. Everything in TensoFlow is a graph - nodes of graph represents operations and data flowing through graph and transformed along the way is termed as "Tensors".
Tensors : Central Unit of data in TensorFlow. A tensor consists of a set of primitive values shaped into an array of any number of dimensions. i.e : [1,2,3] , [[1,2],[2,3]], etc. Scalers(2, 3.5, etc) are 0-D tensor, vector [2,4,6] are 1-D tensors, matrix are 3-D tensors.
Rank, shape and data types are 3 important characteristics which define a Tensor. The number of dimensions in a tensor is Rank of tensor. and the number of elements in each dimension Shape of tensor.
Tensors | Rank |
---|---|
4 | 0 |
[1,2,3] | 1 |
[[1,3],[4,6]] | 2 |
[[[1],[2],[4],[5]]] | 4 |
Tensors | Shape |
---|---|
4 | [] |
[1,2,3] | [3] |
[[1,4,3],[6,8,9]] | [3,2] |
[[[1],[2],[4],[5]]] | [2,2,1] |
- TensorFlow provides a foundation for development of new ML algorithms and it supports large scale distributed models from training to production.
- It provides stable, efficient and performant python API.
- TensorFlow goes hand in hand with TensorBoard which gives graphical view(in the form of dashboard) of graph/activity with minute details for analysis.
Refer TensorFlow official guide to Install and setup TensorFlow environment. Once tensorFlow is setup prompt is changed and open python prompt, execute some commands(follow below lines of commands) to test installation.
➜ ~ source /Users/n0r0082/TensorFlowWork/bin/activate (TensorFlowWork) ➜ ~ python Python 2.7.10 (default, Jul 15 2017, 17:16:57) [GCC 4.2.1 Compatible Apple LLVM 9.0.0 (clang-900.0.31)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import tensorflow as tf >>> hello = tf.constant('Hello, TensorFlow') >>> sess = tf.Session() 2018-01-16 23:26:11.178955: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA >>> print(sess.run(hello)) Hello, TensorFlow
In order to explore usefulness of TensorBoard, we will perform some mathematical operations(square ,power, square root and addition) using python tensorFlow API's and construct graphs out of these operation & visualise the same in tensorBoard.
import tensorflow as tf #create constants a = tf.constant(6.5,name='constant_a') b = tf.constant(3.4,name='constant_b') c = tf.constant(3.0,name='constant_c') d = tf.constant(100.2,name='constant_d') #compute square, power and Sqrt square = tf.square(a,name="square_a") power = tf.pow(b,c,name="power_b_c") sqrt =tf.sqrt(d,name="sqrt_d") #Add all above computed results final_sum = tf.add_n([square,power,sqrt], "total_sum") #create session sess = tf.Session() #Execute above operations using session object print "square is ", sess.run(square) print "power is ", sess.run(power) print "sqrt is ", sess.run(sqrt) print "total sum is ", sess.run(final_sum) # Create writer object for tensorBoard and save operations as grapgh at ./m2_eample writer = tf.summary.FileWriter('./tensorBoardDir',sess.graph) #close resources used. writer.close() sess.close()
Execute above sample and find results
(TensorFlowWork) ➜ samples python sample-2.py 2018-01-17 15:08:16.781469: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.2 AVX AVX2 FMA square is 42.25 power is 39.304005 sqrt is 10.0099945 total sum is 91.563995
Using following command we can start tensorBoard which launches dashboard On some URL.
(TensorFlowWork) ➜ ~ tensorboard --logdir="tensorBoardDir" TensorBoard 0.4.0rc3 at http://m-C02T75XBG8WN:6006 (Press CTRL+C to quit)
Launch dashboard in browser at "http://0.0.0.0:6006". Below screenshot shows each operation as node and data flow from one operation to another as Tensor.
Each node represents operation and output of these three operations are tensors- feed into another operations final_sum.
Image recognition and processing with TensorFlow: Image is collection of pixel and each pixel contains some value. Coloured image is represented in the form of RGB values(3 channels) for each pixel and greyscale image each pixel stores value of intensity between 0.0 - 1.0 (1 channel). Here we will read an image with TensorFlow library and store image as tensors followed by rotate image - find transpose of image.
import tensorflow as tf import matplotlib.image as mp_img import matplotlib.pyplot as plot import os #input file filename = "./input.png" #get image object from input file image = mp_img.imread(filename) print "Image shape is: ", image.shape print "Image array is: ", image #plot.imshow(image) #plot.show() #create a variablefor holding image data x = tf.Variable(image, name='x') #inititlize all global variables init = tf.global_variables_initializer() with tf.Session() as sess: sess.run(init) #compute transpose of variable x (image matrix data) transpose = tf.image.transpose_image(x) #Run transpose data. result = sess.run(transpose) # Display shape of image - first two values shoudld be swapped. print "Transposed image shape: ", result.shape plot.imshow(result) plot.show()
Sample output: -
(TensorFlowWork) ➜ samples python sample-5.py Image shape is: (816, 866, 4) Image array is: [[[0.30980393 0.10980392 0.27058825 1. ] [0.30588236 0.10196079 0.27058825 1. ] [0.31764707 0.10980392 0.28235295 1. ] ... [0.18431373 0.05490196 0.17254902 1. ] [0.18039216 0.04705882 0.16470589 1. ] [0.16862746 0.03921569 0.15686275 1. ]] [[0.29803923 0.10196079 0.2627451 1. ] [0.30588236 0.10588235 0.27058825 1. ] [0.32156864 0.11372549 0.28627452 1. ] ... [0.18039216 0.05098039 0.16862746 1. ] [0.17254902 0.03921569 0.15686275 1. ] [0.1882353 0.05882353 0.1764706 1. ]] ... [0.07843138 0.02745098 0.09411765 1. ] [0.07843138 0.02745098 0.09411765 1. ] [0.07843138 0.02745098 0.09803922 1. ]]] Transposed image shape: (866, 816, 4)
Image orientation comparison :- First image is original image(un-comment line #15 and 16 to generate it) and second image is transposed image.
Original image with shape (816,866,4) and channel value is 4.
Original image displayed via tensorFlow |
Transposed image displayed via tensorFlow |
Note:- Using TensorFlow library we can re-size images, flip images too.
===== ****** =====
0 comments:
Post a Comment