1 00:00:00,240 --> 00:00:06,080 In this lesson we will convert our model our tensor flow model to tensor flow. 2 00:00:06,120 --> 00:00:07,180 Yes. 3 00:00:07,470 --> 00:00:13,170 And in the process we're going to learn how to set up a separate Python environment and also switch 4 00:00:13,170 --> 00:00:15,520 between development environments. 5 00:00:15,540 --> 00:00:20,460 Now we're gonna do that well regular tensor flow models can't run in the browser. 6 00:00:20,460 --> 00:00:26,520 For that our models need to be in the tensor flow J.S. format and this means that we have to convert 7 00:00:26,760 --> 00:00:33,770 our saved model and its PDB files and its variable files into a different format. 8 00:00:33,810 --> 00:00:35,440 The tensor flow J.S. format. 9 00:00:35,550 --> 00:00:37,290 So how we're gonna do that. 10 00:00:37,290 --> 00:00:42,930 Well as always our first stop is going to be the official documentation. 11 00:00:42,930 --> 00:00:49,590 So if I check out tensor flows website here under a model conversion I can see quite a lengthy and descriptive 12 00:00:49,620 --> 00:00:50,400 article. 13 00:00:50,400 --> 00:00:55,830 And as you can see here there are different instructions depending on your starting point. 14 00:00:55,890 --> 00:00:57,720 So there's a slightly different way to go about it. 15 00:00:57,720 --> 00:01:00,840 If you have a saved model versus if you have a Caris model. 16 00:01:00,890 --> 00:01:05,180 This is if you have a tensor flow hub model and because we're working with the saved model we're going 17 00:01:05,170 --> 00:01:12,240 to be making use of this tensor flow J.S. converter and this tensor flow JS converter lives on this 18 00:01:12,240 --> 00:01:18,270 getup repo right here and there's some instructions below but I'm think there is slightly complicated. 19 00:01:18,270 --> 00:01:23,340 So when I walk you through them the thing is and I'm going to say this upfront it can be a little bit 20 00:01:23,340 --> 00:01:25,150 tricky to get this to work. 21 00:01:25,200 --> 00:01:30,510 It certainly was for me and the reason is is that there are a lot of moving parts that all have to play 22 00:01:30,510 --> 00:01:37,680 nice with each other the kind of moving parts I'm talking about are tensor flow tensor flow yes Charisse 23 00:01:37,980 --> 00:01:39,180 and python. 24 00:01:39,180 --> 00:01:44,010 All of these have specific versions that the converter cares about and all of these different parts. 25 00:01:44,010 --> 00:01:47,120 Update periodically but not at the same time. 26 00:01:47,220 --> 00:01:52,770 So I actually had a little bit of trouble getting this to work at first because I ran into some incompatibilities 27 00:01:52,800 --> 00:01:57,660 between the different versions of python and the converter and the converter and versions of Karras 28 00:01:58,020 --> 00:01:59,500 and some other things. 29 00:01:59,550 --> 00:02:02,560 So this was actually a little bit difficult to troubleshoot. 30 00:02:02,760 --> 00:02:08,880 So as a little bit of a backup I'm going to add my output files under the lesson resources so if you 31 00:02:08,880 --> 00:02:14,490 run into any problems and you're having difficulty troubleshooting it you have those available and can 32 00:02:14,490 --> 00:02:19,800 continue on with the module so the question at this point is what's the solution. 33 00:02:19,840 --> 00:02:25,240 How do we get all these components and all the versioning to play nice with each other and to work. 34 00:02:25,240 --> 00:02:28,750 How do we lock down all the versions in a controlled way. 35 00:02:30,040 --> 00:02:33,800 And the answer to this problem is we need a separate environment. 36 00:02:33,970 --> 00:02:39,070 We need a special environment just for the tensor flow GSK converter. 37 00:02:39,070 --> 00:02:43,920 In other words the 10th floor JS converter will use its own version of Python. 38 00:02:44,020 --> 00:02:49,240 It's gonna use a very particular version of Caris and it's gonna use particular versions of all the 39 00:02:49,240 --> 00:02:56,020 other dependencies including its own version of tensor flow Jesus and all these versions may not be 40 00:02:56,110 --> 00:02:59,420 the latest versions but they're the ones that are gonna work. 41 00:02:59,440 --> 00:03:02,080 So how do we set up the separate environment. 42 00:03:02,170 --> 00:03:06,050 The good news is is that Anaconda makes this really really easy for us. 43 00:03:06,160 --> 00:03:13,030 It makes it really easy to set up separate Python environments in parallel and allowing you to switch 44 00:03:13,030 --> 00:03:14,290 in between them. 45 00:03:14,290 --> 00:03:19,750 If you look at the documentation right here we get a little bit of a description of how to go about 46 00:03:19,750 --> 00:03:20,340 it. 47 00:03:20,380 --> 00:03:23,330 Here we can see how we can go about creating an environment. 48 00:03:23,470 --> 00:03:29,830 And here we see how to activate and deactivate an environment all of these instructions pertain to the 49 00:03:29,830 --> 00:03:32,640 terminal or the windows command prompt. 50 00:03:32,660 --> 00:03:38,990 They're all we have to do is type in something like this kinda create which is the command. 51 00:03:39,220 --> 00:03:46,250 Then we provide the environment name that we want to create and then we provide a python version. 52 00:03:46,330 --> 00:03:54,220 So in our case we're gonna write kinda create hyphen and converter underscore environment and then Python 53 00:03:54,280 --> 00:03:56,140 equals three point six. 54 00:03:56,140 --> 00:04:03,400 Now you can actually do this in the anaconda Navigator as well if you prefer using agree as opposed 55 00:04:03,400 --> 00:04:10,510 to the terminal or the command prompt if you fire up the anaconda navigator and you click on environments 56 00:04:10,540 --> 00:04:11,750 right here. 57 00:04:11,770 --> 00:04:14,650 Then what you can see is you can simply click on this button here. 58 00:04:14,680 --> 00:04:23,070 What's this create then you can provide a name converter underscore environment and then you can select 59 00:04:23,080 --> 00:04:27,250 the python version right here and we're gonna be using three point six. 60 00:04:27,250 --> 00:04:32,440 Now at this point you can either click Create or if you want to follow along with me you can do it in 61 00:04:32,440 --> 00:04:34,210 the terminal on Windows. 62 00:04:34,330 --> 00:04:38,150 You'd be using the anaconda prompt which looks like this. 63 00:04:38,340 --> 00:04:46,410 Here you simply type kinda create hyphen n and then you provide the name of the environment you want 64 00:04:46,410 --> 00:04:57,100 to create converter underscore and space python is equal to three point six so I've got my line wrapping 65 00:04:57,100 --> 00:05:01,180 around him but it's all a single line that looks something like this. 66 00:05:01,180 --> 00:05:05,800 Make sure you don't have any typos and you're indeed using python version three point six. 67 00:05:05,800 --> 00:05:06,910 That's quite important. 68 00:05:06,910 --> 00:05:08,320 So hitting enter here. 69 00:05:08,380 --> 00:05:11,400 You can actually see that it's working in the background. 70 00:05:11,680 --> 00:05:15,910 And as it's doing this it's asking us whether or not we want to proceed. 71 00:05:15,910 --> 00:05:21,460 So type Y and hit enter and then you can keep an eye on the anaconda navigator even if you've got it 72 00:05:21,490 --> 00:05:26,810 open and you should actually see it working and creating that second environment. 73 00:05:28,640 --> 00:05:29,090 Brilliant. 74 00:05:29,540 --> 00:05:31,500 So now that you've set up this environment. 75 00:05:31,520 --> 00:05:32,630 How do you activate it. 76 00:05:32,630 --> 00:05:35,230 How do you get into it and how do you get out of it. 77 00:05:35,240 --> 00:05:38,540 The terminal actually shows us the instructions right here. 78 00:05:38,630 --> 00:05:44,960 So I can even copy this pasted in or just type it and kinda activate and then the name of the environment. 79 00:05:45,080 --> 00:05:51,710 And if you hit enter then you should see a change in your prompt instead of reading base in the beginning. 80 00:05:51,710 --> 00:05:56,510 You should see it read converter underscore environment at this point. 81 00:05:56,510 --> 00:05:59,190 Anything you type here will be a command. 82 00:05:59,280 --> 00:06:04,100 That's specific to this particular environment that you set up and took it out of it. 83 00:06:04,100 --> 00:06:10,500 You simply type Conda deactivate just like it says in the instruction above. 84 00:06:10,670 --> 00:06:18,840 And when you do that the command prompt changes back to base now that we've got our environment setup 85 00:06:19,230 --> 00:06:21,470 we're ready to take the next step. 86 00:06:21,480 --> 00:06:29,430 Now there's one little thing that all the windows users have to do first though and that thing is changing 87 00:06:29,580 --> 00:06:33,780 the maximum character limit for our path. 88 00:06:33,960 --> 00:06:41,100 And that's a special extra step for everybody using Windows because by default you've got a 260 character 89 00:06:41,100 --> 00:06:44,790 limit on the path to a particular file. 90 00:06:44,790 --> 00:06:51,870 So that's that piece of text at the very top in your folder that reads C colon slash users slash your 91 00:06:51,870 --> 00:06:54,960 username slash whatever sub folders you've got. 92 00:06:55,080 --> 00:06:57,240 And then the file name at the end. 93 00:06:57,240 --> 00:07:04,170 So by default Windows for historical reasons has had a limit on how many characters there can be in 94 00:07:04,170 --> 00:07:05,290 total here. 95 00:07:05,400 --> 00:07:11,730 And because this 260 character limit will give us a little bit of a problem when we're installing tensor 96 00:07:11,730 --> 00:07:12,090 flow. 97 00:07:12,090 --> 00:07:19,340 Yes we have to lift this limit beforehand and for that we have to go into the Windows Registry Editor. 98 00:07:19,680 --> 00:07:26,550 And the way you would do this on Windows is you simply search for the Registry Editor by typing Reg 99 00:07:26,640 --> 00:07:28,620 edit into the search ball. 100 00:07:28,830 --> 00:07:34,650 And when you click on it you give permission to make changes and then you should see your Windows Registry 101 00:07:34,650 --> 00:07:36,150 pop up at this point. 102 00:07:36,150 --> 00:07:39,960 You have to navigate to a particular place in the registry. 103 00:07:40,140 --> 00:07:48,960 So go to each key local machine then go to system then go to current controls set expand the folder 104 00:07:48,990 --> 00:07:54,460 that reads control and then scroll down a bit until you see filesystem 105 00:07:57,290 --> 00:08:03,600 and inside file system you should see an entry that reads long paths enabled. 106 00:08:03,710 --> 00:08:07,560 So at the very top of the Registry Editor you should see what I have right here. 107 00:08:08,060 --> 00:08:15,740 Each key local machine system current controls set control and then file system double click on long 108 00:08:15,740 --> 00:08:17,630 paths enabled. 109 00:08:17,780 --> 00:08:24,200 And if you've still got the limit in place on two hundred sixty characters then you should see a zero 110 00:08:24,200 --> 00:08:24,950 there. 111 00:08:25,130 --> 00:08:33,800 And what you need to do at this point is you need to change that zero to a one then click OK and then 112 00:08:33,950 --> 00:08:35,810 simply restart your computer. 113 00:08:38,320 --> 00:08:38,900 That way. 114 00:08:38,980 --> 00:08:42,170 The changes will take effect all right. 115 00:08:42,180 --> 00:08:48,750 So now that we've set up our converter environment and we've lifted the 260 character limit. 116 00:08:48,900 --> 00:08:52,910 If you're on Windows the next step is installing tensor flow. 117 00:08:52,980 --> 00:08:59,160 J.S. and the we we're gonna do this is first we're going to activate the converter environment that 118 00:08:59,160 --> 00:09:04,440 we've just set up and so we should see our prompt change and we should see the converter environment 119 00:09:04,500 --> 00:09:09,630 reflected in the beginning of the prompt and then we're going to install tensor flow Jesus and the way 120 00:09:09,630 --> 00:09:16,080 this is gonna work is we're going to issue a command called pip install followed by the package name 121 00:09:16,410 --> 00:09:20,860 and then the version of tensor flow Jesus that we're looking to install. 122 00:09:20,880 --> 00:09:24,270 So Pip stands for Pip install packages. 123 00:09:24,270 --> 00:09:27,460 And yes the abbreviation has the word Pip again. 124 00:09:27,780 --> 00:09:34,620 So there there's some more python in jokes for you but the command that we're going to type into the 125 00:09:34,620 --> 00:09:40,620 terminal or into the kind prompt if you're in windows is Pip install tensor flow G as W equals. 126 00:09:41,190 --> 00:09:47,300 And then we're gonna go for a very particular version which I know works and that's version 1.0. 127 00:09:47,310 --> 00:09:49,520 Two point three. 128 00:09:49,590 --> 00:09:59,580 So here's how it's going to look pip install tensor flow G.S. W equals one point two point three. 129 00:09:59,580 --> 00:10:04,310 So there's some line wrapping happening here but this is all one line one command. 130 00:10:05,070 --> 00:10:10,390 If it enter it's gonna take a little while for tens of low just to be installed. 131 00:10:10,800 --> 00:10:15,270 So there's gonna be some downloads happening and the way you know that it's finished is that you should 132 00:10:15,270 --> 00:10:16,770 see your prompt appear again. 133 00:10:16,770 --> 00:10:23,100 So if you see this blinking cursor in front of your name then you know it's done now. 134 00:10:23,110 --> 00:10:27,770 The next step is actually running the tensor flow J.S. converter because now you've got tense flow. 135 00:10:27,770 --> 00:10:28,860 Yes right. 136 00:10:28,870 --> 00:10:31,090 So how do we do that. 137 00:10:31,210 --> 00:10:37,630 The detailed instructions of course are on the Github page for the tensor flow gas converter and scrolling 138 00:10:37,630 --> 00:10:39,490 down we see that for a saved model. 139 00:10:39,490 --> 00:10:41,050 This is an example. 140 00:10:41,080 --> 00:10:43,340 Now how does this actually translate for us. 141 00:10:43,570 --> 00:10:48,430 Now even though these instructions are not bad it's even simpler for us actually. 142 00:10:48,460 --> 00:10:54,950 So the only thing that we have to type into the anaconda prompt or into the terminal is tense offload 143 00:10:55,030 --> 00:11:01,480 J.S. underscore converter which is the command and then we have to supply a couple of inputs. 144 00:11:01,480 --> 00:11:06,480 The first one is what the format is that we're going to be converting from. 145 00:11:06,910 --> 00:11:12,980 And that's a 10 serve flow saved model T F underscore saved on a screen model. 146 00:11:13,090 --> 00:11:17,110 The next thing that we have to provide is the output format that we're looking for. 147 00:11:17,110 --> 00:11:20,850 And this is a tensor flow J.S. graph model. 148 00:11:20,920 --> 00:11:26,620 Again the spelling has very important and you're going to have to spell everything exactly as you see 149 00:11:26,620 --> 00:11:29,650 here for this to work. 150 00:11:29,770 --> 00:11:32,490 Now you have to supply two more things. 151 00:11:32,530 --> 00:11:36,310 The first is where to find the saved model files. 152 00:11:36,310 --> 00:11:39,950 So these are the PDB files that we generated from Jupiter. 153 00:11:40,480 --> 00:11:44,730 And the next thing is where the output files should be stored. 154 00:11:44,740 --> 00:11:50,270 So this is the folder where the Converter will put the output files and that's it. 155 00:11:50,270 --> 00:11:56,420 Actually it's gonna be a little bit of typing but it's actually just these four things. 156 00:11:56,440 --> 00:11:56,860 All right. 157 00:11:56,880 --> 00:11:58,870 So how's this actually gonna work in practice. 158 00:11:59,200 --> 00:12:05,980 Well let me show you in the terminal it's gonna be tensor flow J.S. underscore converter. 159 00:12:05,980 --> 00:12:07,430 So that's the command. 160 00:12:07,780 --> 00:12:16,740 And then it's gonna be two hyphens with input underscore format equal to T F underscore saved underscore 161 00:12:17,320 --> 00:12:24,390 model then our second input that we're gonna provide is the output on a score format and that's gonna 162 00:12:24,400 --> 00:12:30,830 be equal to TFG s underscore graph on the score model. 163 00:12:30,850 --> 00:12:34,120 Now before you continue on just double check the spelling here. 164 00:12:34,330 --> 00:12:37,990 Make sure you've got no typos in this incantation. 165 00:12:37,990 --> 00:12:39,840 Otherwise we're going to have a problem. 166 00:12:39,880 --> 00:12:41,860 So everything's lowercase. 167 00:12:42,040 --> 00:12:44,520 You've got the underscores where they need to be. 168 00:12:44,560 --> 00:12:46,930 You've got double hyphens where they need to be. 169 00:12:47,440 --> 00:12:51,050 And if you've got some line wrapping Don't worry about it. 170 00:12:51,070 --> 00:12:54,420 So once you're happy with that you can provide the third input. 171 00:12:54,820 --> 00:12:57,890 And that's where those saved model files are saved. 172 00:12:57,890 --> 00:13:04,930 Now all you need to do is drag and drop the directory in to your command prompt and so you can just 173 00:13:05,290 --> 00:13:12,460 drag and drop it in and you'll get the full path of where to find that file and your variable files 174 00:13:12,580 --> 00:13:13,900 and so on. 175 00:13:13,900 --> 00:13:21,780 So that's the third input and the fourth input is where you want the converter to spit out its results. 176 00:13:21,850 --> 00:13:27,070 And so in this case I recommend creating a new folder in your projects folder and you can call this 177 00:13:27,070 --> 00:13:30,580 folder anything you like but I'm gonna go for TFG yes. 178 00:13:30,820 --> 00:13:37,090 And I want to take that folder and I'm going to drag and drop it into the command prompt again to get 179 00:13:37,090 --> 00:13:39,790 the full path to this folder. 180 00:13:39,910 --> 00:13:44,160 And once you're happy with that you can actually run the converter and let's see what happens. 181 00:13:44,350 --> 00:13:49,870 The converter actually doesn't take very much time to run and once it's done and you should see no error 182 00:13:49,870 --> 00:13:55,690 messages at this point you should see this TFG s folder that you've just created populated with two 183 00:13:55,690 --> 00:14:02,820 files the first file is gonna read something like group one short one of one. 184 00:14:02,890 --> 00:14:05,550 And the second file is gonna be your model. 185 00:14:05,560 --> 00:14:06,790 Don't Jason. 186 00:14:06,790 --> 00:14:09,950 Now again if you had any problems any errors. 187 00:14:10,090 --> 00:14:16,270 I'm actually gonna add these two files under the lesson resources so that you can download them in case 188 00:14:16,270 --> 00:14:22,180 you get stuck and you don't have to spend a just troubleshooting the tensor flow just convert them but 189 00:14:22,420 --> 00:14:24,000 I hope this worked for you. 190 00:14:24,010 --> 00:14:30,820 And if you're wondering what this is then you've got on one hand the actual graph and you can see this 191 00:14:30,820 --> 00:14:31,960 beautifully. 192 00:14:31,960 --> 00:14:38,170 If you open it in something like the Firefox browser which formats this Jason File very very nicely 193 00:14:38,770 --> 00:14:43,270 and then this other file here you've got the actual weights of your neural network. 194 00:14:43,360 --> 00:14:44,510 So here's your model. 195 00:14:44,710 --> 00:14:46,440 And these are gonna be your weights. 196 00:14:46,540 --> 00:14:51,640 Now these are gonna be much harder to inspect because this is a binary you're not actually gonna be 197 00:14:51,640 --> 00:14:54,070 able to open this in a program and make sense of it. 198 00:14:54,820 --> 00:14:55,990 So there you go. 199 00:14:55,990 --> 00:15:01,150 Congratulations on successfully converting a tense flow model to a tense flow J. 200 00:15:01,150 --> 00:15:03,790 S model in the next lesson. 201 00:15:03,880 --> 00:15:08,380 We're gonna be starting to look at the Web site that we're gonna build we're going to look at these. 202 00:15:08,380 --> 00:15:12,990 The project I'm going to start learning about h Tim L and CSX. 203 00:15:13,060 --> 00:15:13,770 I'll see you there.