1 00:00:00,000 --> 00:00:03,090 Before we can start building deep 2 00:00:03,090 --> 00:00:05,040 learning networks, we will spend some 3 00:00:05,040 --> 00:00:06,569 time learning about the different deep 4 00:00:06,569 --> 00:00:08,460 learning libraries and frameworks that 5 00:00:08,460 --> 00:00:10,920 are out there. In this video, I will 6 00:00:10,920 --> 00:00:12,750 briefly cover the libraries that we'll 7 00:00:12,750 --> 00:00:15,929 be teaching in this specialization. The 8 00:00:15,929 --> 00:00:18,090 most popular library is in descending 9 00:00:18,090 --> 00:00:22,109 order are TensorFlow, Keras, and PyTorch. 10 00:00:22,109 --> 00:00:25,769 There is also Theano, a library developed 11 00:00:25,769 --> 00:00:27,990 by the Montreal Institute for Learning 12 00:00:27,990 --> 00:00:30,869 Algorithms, and was the major library for 13 00:00:30,869 --> 00:00:32,610 deep learning development even before 14 00:00:32,610 --> 00:00:34,469 TensorFlow and PyTorch. 15 00:00:34,469 --> 00:00:36,809 However, the founders can't afford to 16 00:00:36,809 --> 00:00:38,700 continuously support it and maintain it, 17 00:00:38,700 --> 00:00:41,100 and therefore, the library lost its 18 00:00:41,100 --> 00:00:44,280 popularity. Because of that, in this 19 00:00:44,280 --> 00:00:46,620 specialization, we will focus on the 20 00:00:46,620 --> 00:00:51,539 three other popular libraries. Among the 21 00:00:51,539 --> 00:00:54,539 three libraries. TensorFlow is the most 22 00:00:54,539 --> 00:00:57,059 popular one. It is the library that is 23 00:00:57,059 --> 00:00:59,039 mostly used in production of deep 24 00:00:59,039 --> 00:01:01,170 learning models. It has a very large 25 00:01:01,170 --> 00:01:03,449 community. Just a quick look at the 26 00:01:03,449 --> 00:01:05,670 number of forks on the library's Github 27 00:01:05,670 --> 00:01:07,770 repository as well as the number of 28 00:01:07,770 --> 00:01:10,229 commits and pull requests should suffice 29 00:01:10,229 --> 00:01:12,689 in giving you an idea of how popular the 30 00:01:12,689 --> 00:01:16,710 library is. Tensorflow was developed by 31 00:01:16,710 --> 00:01:18,840 Google and released to the public in 32 00:01:18,840 --> 00:01:22,110 2015, and is still being actively used at 33 00:01:22,110 --> 00:01:23,909 Google for both research and production 34 00:01:23,909 --> 00:01:27,840 needs. PyTorch on the other hand, is the 35 00:01:27,840 --> 00:01:30,060 cousin of the Torch framework, which is 36 00:01:30,060 --> 00:01:32,430 in Lua, and supports machine learning 37 00:01:32,430 --> 00:01:35,930 algorithms running on GPUs in particular. 38 00:01:35,930 --> 00:01:38,939 However being derived from the Torch 39 00:01:38,939 --> 00:01:41,400 framework, PyTorch isn't just a set of 40 00:01:41,400 --> 00:01:43,950 wrappers to support a popular language 41 00:01:43,950 --> 00:01:46,530 like Python. Tt was actually rewritten 42 00:01:46,530 --> 00:01:49,590 and tailored to be fast and feel native. 43 00:01:49,590 --> 00:01:53,430 PyTorch was released in 2016 and has 44 00:01:53,430 --> 00:01:56,159 gained immense interest lately and is 45 00:01:56,159 --> 00:01:57,960 becoming the preferred language over 46 00:01:57,960 --> 00:01:58,649 TensorFlow, 47 00:01:58,649 --> 00:02:01,079 especially in academic research settings 48 00:02:01,079 --> 00:02:02,969 and applications of deep learning 49 00:02:02,969 --> 00:02:05,549 requiring optimizing custom expressions. 50 00:02:05,549 --> 00:02:07,860 PyTorch is supported and being actively 51 00:02:07,860 --> 00:02:12,120 used at Facebook. However, despite their 52 00:02:12,120 --> 00:02:13,800 popularity, both 53 00:02:13,800 --> 00:02:16,170 PyTorch and TensorFlow are not easy to 54 00:02:16,170 --> 00:02:19,110 use, and have a steep learning curve. So 55 00:02:19,110 --> 00:02:20,550 for people who are just starting to 56 00:02:20,550 --> 00:02:22,590 learn deep learning, there is no better 57 00:02:22,590 --> 00:02:24,780 library to use other than the Keras 58 00:02:24,780 --> 00:02:25,440 library. 59 00:02:25,440 --> 00:02:28,440 Keras is a high level API for building 60 00:02:28,440 --> 00:02:30,870 deep learning models. It has gained favor 61 00:02:30,870 --> 00:02:33,060 for its ease of use and syntactic 62 00:02:33,060 --> 00:02:35,880 simplicity facilitating fast development. 63 00:02:35,880 --> 00:02:37,860 As you'll see in the next couple of 64 00:02:37,860 --> 00:02:40,260 videos, building a very complex deep 65 00:02:40,260 --> 00:02:42,420 learning network can be achieved with 66 00:02:42,420 --> 00:02:44,520 Keras with only few lines of code. 67 00:02:44,520 --> 00:02:46,980 Keras normally runs on top of a 68 00:02:46,980 --> 00:02:49,290 low-level library such as TensorFlow. 69 00:02:49,290 --> 00:02:51,810 This means that to be able to use the 70 00:02:51,810 --> 00:02:53,880 Kares library, you will have to install 71 00:02:53,880 --> 00:02:56,280 TensorFlow first, and when you import 72 00:02:56,280 --> 00:02:58,740 Keras, it will be explicitly displayed 73 00:02:58,740 --> 00:03:01,080 what backend was used to install the 74 00:03:01,080 --> 00:03:03,870 Keras library. Keras is also supported 75 00:03:03,870 --> 00:03:06,390 by Google. I won't go into more details 76 00:03:06,390 --> 00:03:08,000 about the different libraries, 77 00:03:08,000 --> 00:03:09,980 but the take home message here is if 78 00:03:09,980 --> 00:03:11,740 you're interested in building something 79 00:03:11,740 --> 00:03:14,240 quickly go with the Keras library; you 80 00:03:14,240 --> 00:03:16,640 won't be disappointed. However, if you 81 00:03:16,640 --> 00:03:18,420 want to have more control over the 82 00:03:18,420 --> 00:03:20,120 different nodes and layers in the 83 00:03:20,120 --> 00:03:22,280 network, and want to watch closely what 84 00:03:22,280 --> 00:03:24,720 happens with the network over time, then 85 00:03:24,720 --> 00:03:26,900 PyTorch or TensorFlow would be the 86 00:03:26,900 --> 00:03:29,340 right library. It will really boil 87 00:03:29,340 --> 00:03:33,460 down to your personal preference. 88 00:03:33,460 --> 00:03:35,460 With that, in the next videos, we will start 89 00:03:35,460 --> 00:03:37,680 learning how to use the Keras library 90 00:03:37,680 --> 00:03:39,600 to build models for regression and 91 00:03:39,600 --> 00:03:41,300 classification problems.