Lecture 2 Tensorflow Alireza Akhavan Pour CLASS VISION

Lecture 2: Tensorflow آﺸﻨﺎیی ﺑﺎ Alireza Akhavan Pour CLASS. VISION SRTTU – A. Akhavan Lecture 2 - 1 ۱۳۹۷ ﻣﻬﺮ ۹ - ﺩﻭﺷﻨﺒﻪ

Github آﻤﺎﺭ ﺩﺭ Name release Star Fork Tensorflow Nov 1, 2015 44 k 20 k caffe Sep 8, 2013 15 k 10 k keras Mar 22, 2015 11 k 4 k mxnet Apr 26, 2015 8 k 3 k torch Jan 22, 2012 6 k 2 k theano Jan 6, 2008 5 k 2 k Last update: 2017 22 January 2 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

Github آﻤﺎﺭ ﺩﺭ Name release Star Fork Tensorflow Nov 1, 2015 111 k 68 k caffe Sep 8, 2013 25 k 15 k keras Mar 22, 2015 34 k 13 k mxnet Apr 26, 2015 15 k 5 k pytorch Jan 22, 2012 19 k 4. 5 k theano Jan 6, 2008 8. 5 k 2. 5 k Last update: September 30, 2018 3 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

آﻤﺎﺭ ﺟﺴﺖ ﻭ ﺟﻮ ﺩﺭ گﻮگﻞ https: //trends. google. com/trends/explore? date=2015 -01 -06%202018 -0930&q=tensorflow%20 tutorial, pytorch%20 tutorial, keras%20 tutorial, caffe%20 tutorial, Mxnet%20 tutorial 4 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﺑﺎﺯﻧﻮیﺴی ﺑﺎ ﺗﻨﺴﻮﺭﻓﻠﻮ In [2]: import tensorflow as tf ! ؟ session In [3]: tf. Interactive. Session() In [4]: a = tf. zeros((2, 2)); b = tf. ones((2, 2)) ! ؟ eval In [5]: tf. reduce_sum(b, reduction_indices=1). eval() Out[5]: array([ 2. , 2. ], dtype=float 32) In [6]: a. get_shape() Tensor. Shape ﻫﺎ tuple ﻣﺜﻞ پﺎیﺘﻮﻥ ﺩﺭ ﺭﻓﺘﺎﺭ ﻣیکﻨﻨﺪ Out[6]: Tensor. Shape([Dimension(2), Dimension(2)]) In [7]: tf. reshape(a, (1, 4)). eval() Out[7]: array([[ 0. , 0. ]], dtype=float 32) 10 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻭ numpy ﻣﻘﺎیﺴﻪی ﺍﺳﺘﻔﺎﺩﻩ ﺍﺯ Tensorflow Numpy Tensor. Flow a= np. zeros((2, 2)); b = np. ones((2, 2)) a= tf. zeros((2, 2)), b = tf. ones((2, 2)) np. sum(b, axis=1) tf. reduce_sum(a, reduction_indices=[1]) a. shape a. get_shape() np. reshape(a, (1, 4)) tf. reshape(a, (1, 4)) b * 5 + 1 np. dot(a, b) tf. matmul(a, b) a[0, 0], a[: , 0], a[0, : ] 11 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

(1) ﻣﺜﺎﻝ ﻫﺎیی ﺑیﺸﺘﺮ ﺍﺯ گﺮﺍﻑ x = tf. constant(2) y = 3 op 1 = tf. add(x, y) op 2 = tf. mul(x, y) op 3 = tf. pow(op 2, op 1) with tf. Session() as sess: op 3 = sess. run(op 3) 15 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

(2) ﻫﺎ ﺩﺭ ﺗﻨﺴﻮﺭﻓﻠﻮ Variable In [32]: W 1 = tf. ones((2, 2)) In [33]: W 2 = tf. Variable(tf. zeros((2, 2)), name="weights") In [34]: with tf. Session() as sess: print(sess. run(W 1)) sess. run(tf. initialize_all_variables()) print(sess. run(W 2)). . : [[ 1. 1. ]] [[ 0. 0. ]] 19 ﺑﺮﺍی tf. global_variables_initializer ﻓﺮﺍﺧﻮﺍﻧی ﻣﻘﺪﺍﺭ ﺩﻫی ﺍﻭﻟیﻪ ﺑﻪ ﺯﻭﺩی ﺣﺬﻑ tf. initialize_all_variables() ﺗﺎﺑﻊ q. ﻣیﺷﻮﺩ tf. global_variables_initializer : ﺟﺎیگﺰیﻦ q پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

variable ﺑﻪ ﺭﻭﺯ ﺭﺳﺎﻧی یک In [63]: state = tf. Variable(0, name="counter") new_value = state + 1 ﻣﻌﺎﺩﻝ In [64]: new_value = tf. add(state, tf. constant(1)) In [65]: update = tf. assign(state, new_value) In [66]: with tf. Session() as sess: . . : sess. run(tf. global_variables_initializer()). . : print(sess. run(state)) . . : for _ in range(3): . . : sess. run(update) . . : print(sess. run(state)) state = new_value ﻣﻌﺎﺩﻝ state = 0 print(state) for _ in range(3): state = state + 1 print(state) : ﻗﻄﻌﻪ کﺪ ﺧﺮﻭﺟی ü 0 1 2 3 . . : 21 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

(2) ﻫﺎ ﺩﺭ ﺗﻨﺴﻮﺭﻓﻠﻮ Variable ﻭﺍکﺸی input 1 = tf. constant(3. 0) input 2 = tf. constant(2. 0) input 3 = tf. constant(5. 0) intermed = tf. add(input 2, input 3) mul = tf. mul(input 1, intermed) with tf. Session() as sess: result = sess. run([mul, intermed]) print(result) 23 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

numpy ﻭﺭﻭﺩی ﺑﺎ ﺍﺳﺘﻔﺎﺩﻩ ﺍﺯ In [93]: a = np. zeros((3, 3)) In [94]: ta = tf. convert_to_tensor(a) In [95]: with tf. Session() as sess: . . : print(sess. run(ta)). . : [[ 0. 0. 0. ]] 25 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

Feed Dictionary ﻭ Placeholder ﻭﺭﻭﺩی ﺑﺎ ﺍﻣﺎ ، ﺍﻣکﺎﻥ پﺬیﺮ ﻭ ﺭﺍﺣﺖ ﺍﺳﺖ numpy ﻭﺭﻭﺩی ﺑﺎ ﺍﺳﺘﻔﺎﺩﻩ ﺍﺯ ü. ﻣﻘیﺎﺱپﺬیﺮ ﻧیﺴﺖ feed_dict ﻭ tf. placeholder ﺍﺳﺘﻔﺎﺩﻩ ﺍﺯ ü In [96]: input 1 = tf. placeholder(tf. float 32) In [97]: input 2 = tf. placeholder(tf. float 32) �� ���� tf. placeholder ����� In [98]: output = tf. mul(input 1, input 2) In [99]: with tf. Session() as sess: . . : print(sess. run([output], feed_dict={input 1: [7. ], input 2: [2. ]})) . . : [array([ 14. ], dtype=float 32)] Fetch value of output from computation graph. 26 Feed data into computation graph. پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

Feed Dictionary ﻭ Placeholder ﻭﺭﻭﺩی ﺑﺎ (3) In [96]: input 1 = tf. placeholder(tf. float 32) In [97]: input 2 = tf. placeholder(tf. float 32) In [98]: output = tf. mul(input 1, input 2) In [99]: with tf. Session() as sess: . . : print(sess. run([output], feed_dict={input 1: [7. ], input 2: [2. ]})) . . : [array([ 14. ], dtype=float 32)] 28 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

Cross entropy: actual probabilities, “one-hot” encoded • computed probabilities 35 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻣﻘﺪﺍﺭﺩﻫی ﺍﻭﻟیﻪ import tensorflow as tf X = tf. placeholder(tf. float 32, [None, 28, 1]) W = tf. Variable(tf. zeros([784, 10])) b = tf. Variable(tf. zeros([10])) init = tf. initialize_all_variables() 36 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻣﻌیﺎﺭ # model Y = tf. nn. softmax(tf. matmul(tf. reshape(X, [-1, 784]), W) + b) # placeholder for correct answers Y_ = tf. placeholder(tf. float 32, [None, 10]) # loss function cross_entropy = -tf. reduce_sum(Y_ * tf. log(Y)) # % of correct answers found in batch is_correct = tf. equal(tf. argmax(Y, 1), tf. argmax(Y_, 1)) accuracy = tf. reduce_mean(tf. cast(is_correct, tf. float 32)) 37 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

آﻤﻮﺯﺵ optimizer = tf. train. Gradient. Descent. Optimizer(0. 003) train_step = optimizer. minimize(cross_entropy) 38 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﺍﺟﺮﺍ sess = tf. Session() sess. run(init) for i in range(1000): # load batch of images and correct answers batch_X, batch_Y = mnist. train. next_batch(100) train_data={X: batch_X, Y_: batch_Y} # train sess. run(train_step, feed_dict=train_data) # success ? a, c = sess. run([accuracy, cross_entropy], feed_dict=train_data) # success on test data ? test_data={X: mnist. test. images, Y_: mnist. test. labels} a, c = sess. run([accuracy, cross_entropy, It], feed=test_data) 39 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﺩﺭﺱﻫﺎی آﻨﻼیﻦ ﻭ ﺭﺍیگﺎﻥ آﻤﻮﺯﺵ ﺗﻨﺴﻮﺭﻓﻠﻮ Deep Learning with Tensor. Flow https: //bigdatauniversity. com/courses/deep-learning-tensorflow/ Deep Learning by Tensorflow ) (using https: //www. udacity. com/course/deep-learning--ud 730 CS 20 SI: Tensorflow for Deep Learning Research https: //web. stanford. edu/class/cs 20 si/ 43 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻣﻨﺎﺑﻊ ﺍﺻﻠی ﺍیﻦ ﺍﺭﺍﺋﻪ ﻭ ﻣﻔیﺪ ﺑﺮﺍی . . . ﻣﻄﺎﻟﻌﻪ 19 th Apr 2016 https: //cs 224 d. stanford. edu/lectures/CS 224 d-Lecture 7. pdf https: //www. youtube. com/watch? v=l 6 K-MFg. IEjc 8 Nov 2016 https: //gotocon. com/. . . Tensorflow. And. Deep. Learning. Without. APh. D. pdf https: //www. youtube. com/watch? v=vq 2 nn. J 4 g 6 N 0 13 and 18 Jan 2017 http: //web. stanford. edu/class/cs 20 si/lectures/slides_01. pdf http: //web. stanford. edu/class/cs 20 si/lectures/slides_02. pdf 44 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﺳﺎیﺮ ﻣﻨﺎﺑﻊ q. Tensor. Flow official site Øhttp: //tensorflow. org/ q. Deep Learning Frameworks Compared by Siraj Raval Øhttps: //www. youtube. com/watch? v=MDP 9 Ffs. Nx 60 q. Tensor. Flow - Ep. 22 (Deep Learning SIMPLIFIED) Øhttps: //www. youtube. com/watch? v=b. Ye. BL 92 v 99 Y 45 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻣﺤﺎﺳﺒﺎﺕ ﺗﻮﺯیﻊ ﺷﺪﻩ # Creates a graph. ﻣﺸﺨﺺ کﺮﺩﻥ ﻗﺴﻤﺘی ﺍﺯ گﺮﺍﻑ ﺑﺮﺍی ﺍﺟﺮﺍ ﺭﻭی with tf. device('/gpu: 2'): ﺧﺎﺹ GPU یﺎ CPU a = tf. constant([1. 0, 2. 0, 3. 0, 4. 0, 5. 0, 6. 0], name='a') b = tf. constant([1. 0, 2. 0, 3. 0, 4. 0, 5. 0, 6. 0], name='b') c = tf. matmul(a, b) # Creates a session with log_device_placement set to True. sess =tf. Session(config=tf. Config. Proto(log_device_placement=True)) # Runs the op. print sess. run(c) 46 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ

ﻣﻨﺒﻊ https: //www. slideshare. net/Alirezaakhavanpour/tensorflow- • 71395844 47 پﻮﺭ ﺍﺧﻮﺍﻥ ﻋﻠیﺮﺿﺎ