1
00:00:00,600 --> 00:00:04,260
To start this course, we shall go from the very basics.

2
00:00:04,260 --> 00:00:11,550
So start by understanding tenses and variables and how to use them with TensorFlow.

3
00:00:11,580 --> 00:00:20,700
You could see here would go from the very basics like initialization and casting right up to more advanced

4
00:00:20,700 --> 00:00:24,870
tensor types like the sparse rect and string tenses.

5
00:00:24,900 --> 00:00:30,090
Then we'll also look at the variables and how to load and safe models in TensorFlow.

6
00:00:30,360 --> 00:00:36,630
Now from here we shall dive into this car prediction project and this competition project.

7
00:00:36,630 --> 00:00:42,660
Actually, you will learn how to build a simple linear regression model.

8
00:00:42,660 --> 00:00:50,610
So you first of all, start by preparing your data, then building out this model, creating some last

9
00:00:50,610 --> 00:00:54,910
function, and then train your own model.

10
00:00:54,930 --> 00:00:59,070
Then once the model is trained, you dive into performance measurement.

11
00:00:59,260 --> 00:01:04,290
We'll then see how well our model performs with the validation and testing.

12
00:01:04,290 --> 00:01:08,280
And then based on this performance, we'll carry out some corrective measures.

13
00:01:08,730 --> 00:01:15,510
Now we've looked at how to build a simple linear regression model for the next project, which I'll

14
00:01:15,510 --> 00:01:16,230
be working on.

15
00:01:16,230 --> 00:01:18,830
That is on malaria diagnosis.

16
00:01:18,840 --> 00:01:24,180
We shall understand how the convolutional neural networks work.

17
00:01:24,180 --> 00:01:31,980
So from your going from this fully connected neural networks, we'll now move to the convolutional neural

18
00:01:31,980 --> 00:01:36,180
networks, which are greatly used in the domain of computer vision.

19
00:01:36,180 --> 00:01:42,030
So just as before prepared the data, then we build the model.

20
00:01:42,030 --> 00:01:46,380
With TensorFlow, we have the similar training process.

21
00:01:46,380 --> 00:01:53,640
We evaluate the model, and then we load and save, or we learn how to save and then load our model

22
00:01:54,210 --> 00:01:55,770
to Google Drive.

23
00:01:56,070 --> 00:02:03,330
Now, from this step, that is from training with a convolutional neural network, we will then learn

24
00:02:03,330 --> 00:02:07,260
how to build even more advanced models with TensorFlow.

25
00:02:07,260 --> 00:02:16,080
So your will build this convolutional neural network with the sequential API, and then you will use

26
00:02:16,080 --> 00:02:22,140
more advanced methods like the functional models and even models of class.

27
00:02:22,140 --> 00:02:26,730
And then we'll take a deep dive at the model evaluation.

28
00:02:26,730 --> 00:02:33,990
That is, we'll look at even more better methods of evaluating our model, like the precision recall

29
00:02:33,990 --> 00:02:44,850
accuracy coefficient metrics And the RC curve from here would improve on our model performance by seeing

30
00:02:44,850 --> 00:02:51,600
how to leverage certain functionalities which come with TensorFlow, like the callbacks, which itself

31
00:02:51,600 --> 00:03:00,810
also has other specific callbacks like learning rate scheduling, model, check pointing, and even

32
00:03:00,810 --> 00:03:01,860
early stopping.

33
00:03:02,160 --> 00:03:07,520
Then from here we shall see how to tackle the problems of overfitting and undefeated.

34
00:03:07,530 --> 00:03:12,810
First of all, we'll start by understanding what those problems are and then we'll see how to solve

35
00:03:12,810 --> 00:03:14,380
or tackle these problems.

36
00:03:14,400 --> 00:03:22,230
Now, talking about overfitting, one very common means we generally use to tackle this problem is data

37
00:03:22,230 --> 00:03:23,310
augmentation.

38
00:03:23,310 --> 00:03:31,200
And so in the section we shall dive into different data augmentation methods we could use with TensorFlow.

39
00:03:31,200 --> 00:03:37,590
And that is going from the very basic methods which are already available in TensorFlow image package

40
00:03:37,590 --> 00:03:43,410
right up to the mix up and current mix augmentation strategies.

41
00:03:43,440 --> 00:03:52,620
Now, given that albumin stations is a data augmentation forecast library, we shall see how to make

42
00:03:52,620 --> 00:03:56,520
use of it and how to integrate this with TensorFlow.

43
00:03:56,700 --> 00:04:05,760
Now from here, we'll look at other advanced topics like how to create custom losses and metrics, how

44
00:04:05,760 --> 00:04:12,360
to switch between the eager and graph modes and understanding how this work.

45
00:04:12,360 --> 00:04:19,290
And from here, also creating our own custom training loops, which are very important when we want

46
00:04:19,290 --> 00:04:22,680
to build some customized systems.

47
00:04:23,340 --> 00:04:27,710
From here we'll move on to the tensor bot integration.

48
00:04:27,720 --> 00:04:37,800
So Tensor Bot is this tool which permits us not only log our data but also view model graphs carry out

49
00:04:37,800 --> 00:04:39,210
hyper parameter tuning.

50
00:04:39,210 --> 00:04:40,200
Thus peak out.

51
00:04:40,200 --> 00:04:50,070
Help us pick out the best hyper parameters for training setup and also carry out profiling and certain

52
00:04:50,070 --> 00:04:55,590
visualizations which are important in helping us take certain strategic decisions.

53
00:04:55,920 --> 00:04:59,910
Then from here we shall focus on weights and.

54
00:05:00,040 --> 00:05:07,300
I asked this and we're symbiosis is this machine learning operations tool which helps us with experiment

55
00:05:07,300 --> 00:05:13,810
tracking hyper parameter training to dataset versioning and even model version.

56
00:05:13,810 --> 00:05:20,290
And so with this tool, we are not only going to train our models, but we are also going to be able

57
00:05:20,290 --> 00:05:24,190
to keep track of our model training process.

58
00:05:24,640 --> 00:05:32,140
Now, at this point, we've looked at all the different topics, but with this project that is our malaria

59
00:05:32,140 --> 00:05:33,190
diagnosis.

60
00:05:33,190 --> 00:05:41,980
So we will then move on to human emotion recognition and we'll then build our model fully train this

61
00:05:41,980 --> 00:05:49,980
model, implemented augmentation and also see how to integrate TensorFlow records.

62
00:05:49,990 --> 00:05:56,230
But given that we've been working with very basic convolutional neural networks so far, we are then

63
00:05:56,230 --> 00:06:02,530
going to look at more advanced and modern convolutional neural networks like the nets, big nets, rest

64
00:06:02,560 --> 00:06:06,220
nets, mobile nets and efficient nets.

65
00:06:07,090 --> 00:06:13,750
And from here we will see how these models outperform the very basic model we had been working with

66
00:06:13,750 --> 00:06:14,590
so far.

67
00:06:14,770 --> 00:06:23,890
Then we'll also dive into this very interesting topic of trans fat learning where instead of training

68
00:06:23,890 --> 00:06:33,640
our models from scratch will make use of already pre trained convolutional neural networks to not only

69
00:06:33,640 --> 00:06:38,020
speed up our own training process, but also get much better results.

70
00:06:39,040 --> 00:06:45,700
Then after getting this better results is going to be important for us to understand what goes on in

71
00:06:45,700 --> 00:06:48,160
the black box.

72
00:06:48,160 --> 00:06:52,840
By black box here we simply mean the convolutional neural network.

73
00:06:52,840 --> 00:07:00,250
So we'll look at two methods, thus the visualization of the intermediate layers and also the grad cam

74
00:07:00,250 --> 00:07:00,760
method.

75
00:07:01,360 --> 00:07:05,470
Then we'll look at class imbalance and assembling.

76
00:07:06,370 --> 00:07:15,490
This is actually tackling class imbalance and how to bring together several models to produce one model,

77
00:07:16,120 --> 00:07:18,270
and then that will be it for the calf nets.

78
00:07:18,280 --> 00:07:23,260
We will now look at the even more recent models.

79
00:07:23,260 --> 00:07:25,660
That's the transformers in vision.

80
00:07:25,660 --> 00:07:32,380
So so far before this time, the transformers were mainly used in natural language processing, but

81
00:07:32,380 --> 00:07:36,370
now we have transformers which are also used in computer vision.

82
00:07:36,370 --> 00:07:39,460
And so in the section we'll see how the whites work.

83
00:07:39,460 --> 00:07:42,280
We're going to build a bit from scratch.

84
00:07:42,280 --> 00:07:50,050
We're also going to fine tune a bit from Haugen phase, and then after evaluating the model, we'll

85
00:07:50,050 --> 00:07:57,100
see how it outperforms even the best components We had already worked with our seen right here.

86
00:07:57,280 --> 00:07:59,530
So that will be it for the whites.

87
00:07:59,530 --> 00:08:04,720
We are going to look at more data efficient transformers and the screen transformers.

88
00:08:05,020 --> 00:08:07,990
Then from here we'll see how to deploy our model.

89
00:08:07,990 --> 00:08:11,290
So we create end to end systems which actually work.

90
00:08:11,290 --> 00:08:15,250
So start by understanding quantization.

91
00:08:16,120 --> 00:08:22,720
And the reason behind this is simply because we could make our models much smaller and faster without

92
00:08:22,720 --> 00:08:25,300
losing much model performance.

93
00:08:25,300 --> 00:08:27,040
So that's what we're going to do.

94
00:08:27,040 --> 00:08:33,460
We're going to see how to convert our model into an Onyx model that's from TensorFlow to Onyx, how

95
00:08:33,460 --> 00:08:39,940
to quantized the Onyx model, how to implement quantization, our training, how to also convert our

96
00:08:39,940 --> 00:08:41,950
model to TensorFlow Lite.

97
00:08:41,950 --> 00:08:44,680
And then from here, we're ready to deploy this model.

98
00:08:44,680 --> 00:08:48,070
But before deploying, we'll understand how APIs work.

99
00:08:48,070 --> 00:08:55,270
We're going to build our own API with fast API Python library, then we'll deploy our API to the cloud.

100
00:08:55,390 --> 00:09:00,730
And finally, we're going to carry out some load testing to see how many users could make use of our

101
00:09:00,730 --> 00:09:02,140
API at once.

102
00:09:02,770 --> 00:09:09,160
So now that we're able to deploy our model to the cloud, we are then going to move to another project

103
00:09:09,160 --> 00:09:10,750
that's object detection.

104
00:09:10,840 --> 00:09:17,530
And it should be noted that in this section we are going to build our own YOLO model from scratch.

105
00:09:17,530 --> 00:09:25,210
So we'll build this model, train it, and then we will test it for you to see how we could carry out

106
00:09:25,210 --> 00:09:28,000
object detection from scratch.

107
00:09:28,150 --> 00:09:34,570
The next project we shall work and will be that of image segmentation, and we shall build a simple

108
00:09:34,570 --> 00:09:37,570
unit model to solve this problem.

109
00:09:37,570 --> 00:09:43,180
And the section will implement techniques like data augmentation and class weighting to improve our

110
00:09:43,180 --> 00:09:44,770
model's performance.

111
00:09:44,770 --> 00:09:53,650
And then we'll move on to the people's counting problem where we shall make use of the CSR net to build

112
00:09:53,650 --> 00:09:59,050
a working system which gets inputs of people like this.

113
00:09:59,050 --> 00:09:59,470
And.

114
00:09:59,670 --> 00:10:03,840
Able to count the number of people we have in year from year.

115
00:10:03,870 --> 00:10:08,420
We will now finally have this project on Image Generation.

116
00:10:08,430 --> 00:10:10,200
It's actually a matter of two projects.

117
00:10:10,200 --> 00:10:17,460
The first one is Digits Generation with VA is that variational auto encoders and then for the next that

118
00:10:17,460 --> 00:10:24,180
is fifth generation with Gans would first of all start by understanding how the Gans work and then we'll

119
00:10:24,180 --> 00:10:30,960
make use of this Gans that specifically the this again to generate phase images.

120
00:10:31,380 --> 00:10:33,480
This course is constantly be updated.

121
00:10:33,480 --> 00:10:42,540
So once you're done with this course material, you could take out five different projects which you

122
00:10:42,540 --> 00:10:46,110
build end to end and then share that with the world.