WEBVTT 00:01.260 --> 00:03.180 -: Hey, I'm gonna show you how to use AI 00:03.180 --> 00:07.200 to tell what's performing inside of your creative images. 00:07.200 --> 00:10.845 So we're gonna use Google's vision API, 00:10.845 --> 00:15.060 and to do that we just need to run 00:15.060 --> 00:16.575 PIP store, Google Cloud Vision. 00:16.575 --> 00:19.710 It's gonna install the library that we need. 00:19.710 --> 00:23.605 And then we're gonna just import these libraries here. 00:23.605 --> 00:25.140 So when you're running these cells, 00:25.140 --> 00:27.840 it's literally just click into the cell and shift enter, 00:27.840 --> 00:29.700 or command enter if you're on a Mac 00:29.700 --> 00:31.473 or you can click play here. 00:32.400 --> 00:35.605 So what we're gonna do is we're gonna pull 00:35.605 --> 00:39.325 the campaign stats, the campaign results in first. 00:39.325 --> 00:42.600 So this is the performance data 00:42.600 --> 00:44.430 that you get from your campaign. 00:44.430 --> 00:48.215 So here we have like a specific ID for each creative 00:48.215 --> 00:50.525 that we ran, that each one of these is an image 00:50.525 --> 00:52.080 and we've got the reach impressions, 00:52.080 --> 00:55.065 clicks, spend, conversions, whatever else you need. 00:55.065 --> 00:57.960 So before we can do that, we need to upload it 00:57.960 --> 01:00.420 and you should be able to see that. 01:00.420 --> 01:03.240 We can see the campaign results, CSV here. 01:03.240 --> 01:04.650 So what we need to do, 01:04.650 --> 01:06.210 if you're in the notebook, 01:06.210 --> 01:08.667 it might just have something like this. 01:08.667 --> 01:12.660 You just click on the folder and then click upload, 01:12.660 --> 01:16.680 and then you can kind of choose the file to upload it there. 01:16.680 --> 01:19.050 You also want to do that for the images. 01:19.050 --> 01:24.050 So you're gonna get a link to all the images in this folder. 01:24.870 --> 01:27.330 I've got all the images here, 01:27.330 --> 01:29.280 look, they're on my computer. 01:29.280 --> 01:32.365 So you can see these are just the different ads that we ran, 01:32.365 --> 01:34.165 you know, for our campaign. 01:34.165 --> 01:35.970 And we wanna know, okay, 01:35.970 --> 01:39.240 is it pictures of women or men 01:39.240 --> 01:40.537 or is it kind of like... 01:40.537 --> 01:42.533 You know, what style of images 01:42.533 --> 01:44.220 worked really well for our campaigns? 01:44.220 --> 01:45.930 Or was it these kind of, you know, 01:45.930 --> 01:49.200 clay emoji style images that worked? 01:49.200 --> 01:54.200 And you know, obviously there's like 20 ads here. 01:54.240 --> 01:57.600 Yeah, you could go through and kinda manually tag them, 01:57.600 --> 01:59.575 but it might be interesting to see 01:59.575 --> 02:01.715 what the AI comes up with. 02:01.715 --> 02:04.800 So once you've uploaded, 02:04.800 --> 02:07.860 you can just run this and get the campaign results. 02:07.860 --> 02:09.305 And then what we're gonna do is we're gonna use Google 02:09.305 --> 02:11.790 Vision to tag each image 02:11.790 --> 02:14.820 and then join that to the campaign results. 02:14.820 --> 02:18.030 So what this is doing is just running through 02:18.030 --> 02:20.820 and once you've uploaded all the images, 02:20.820 --> 02:23.820 so you know, in this case, you'd click upload 02:23.820 --> 02:26.045 and then find all the images on your computer 02:26.045 --> 02:27.963 and then upload 'em that way. 02:28.855 --> 02:32.520 This is gonna make sure that you have all those images. 02:32.520 --> 02:35.100 So it's just gonna run through, 02:35.100 --> 02:38.880 look in this folder in the Google Colab notebook, 02:38.880 --> 02:41.790 and then it's just gonna kind of find everything 02:41.790 --> 02:44.245 that's a PNG in that folder. 02:44.245 --> 02:45.895 So don't worry too much about how this works, 02:45.895 --> 02:49.560 it's just gonna display them with the name here. 02:49.560 --> 02:51.753 So just gonna let that run. 02:52.710 --> 02:53.610 Here we go. 02:53.610 --> 02:54.810 Okay, so you can see what all the 02:54.810 --> 02:55.965 different images look like. 02:55.965 --> 02:57.690 This is useful because sometimes you 02:57.690 --> 02:58.830 need to remind yourself like 02:58.830 --> 03:02.475 What ad did we run for that campaign? 03:02.475 --> 03:03.308 Cool. 03:03.308 --> 03:05.550 The other thing you need to do is 03:05.550 --> 03:07.860 get a service account, JSON. 03:07.860 --> 03:11.310 So this is basically like a password, 03:11.310 --> 03:13.545 but it's a file that you can save. 03:13.545 --> 03:17.190 So if you go to this page here, 03:17.190 --> 03:18.023 there's a link, 03:18.865 --> 03:20.580 it tells you how to get started. 03:20.580 --> 03:22.500 So if you haven't set up a Google Cloud 03:22.500 --> 03:25.320 console project before, 03:25.320 --> 03:27.390 it's gonna walk you through how to do that. 03:27.390 --> 03:28.980 You need to enable billing. 03:28.980 --> 03:30.150 So it doesn't work 03:30.150 --> 03:32.160 unless you don't have a credit card behind it. 03:32.160 --> 03:34.080 It's not really gonna cost you very much 03:34.080 --> 03:36.690 unless you do something hilariously wrong. 03:36.690 --> 03:38.719 But you know, for me it's like, 03:38.719 --> 03:40.320 I think a couple cents. 03:40.320 --> 03:41.760 And you usually get some credit 03:41.760 --> 03:43.680 that you can use like a few hundred dollars 03:43.680 --> 03:46.660 when you set up the Google Cloud project. 03:46.660 --> 03:47.764 You can also... 03:47.764 --> 03:50.428 The next thing you need to do is enable the API. 03:50.428 --> 03:52.665 So literally, if you click on this, 03:52.665 --> 03:56.400 it will actually just kind of look at the project 03:56.400 --> 03:59.520 and then it's gonna confirm, you know, 03:59.520 --> 04:02.640 within your project if you are happy to enable it. 04:02.640 --> 04:05.105 So that's the next thing you do. 04:05.105 --> 04:06.840 And that's just kind of a fail safe 04:06.840 --> 04:08.705 to make sure you don't use the wrong APIs. 04:08.705 --> 04:13.530 And then the next thing you do is get a service account key. 04:13.530 --> 04:17.370 So if you click go to create service accounts, 04:17.370 --> 04:21.000 then that's gonna kind of give you the different projects 04:21.000 --> 04:22.290 and you're gonna choose that. 04:22.290 --> 04:23.940 And then you're gonna go through this flow 04:23.940 --> 04:26.850 and just create and continue. 04:26.850 --> 04:29.325 Once you've got a service account, 04:29.325 --> 04:33.250 then you basically just need to click into it 04:35.940 --> 04:39.660 and then you wanna go to Keys, 04:39.660 --> 04:43.380 and then you can download a key here. 04:43.380 --> 04:46.455 So you can go add a key, create a new key, 04:46.455 --> 04:49.695 and I'm gonna delete mine after this. 04:49.695 --> 04:52.383 So, you know, don't worry, 04:53.550 --> 04:55.865 you won't be able to access. 04:55.865 --> 04:57.630 Yeah, because I'll have it deleted 04:57.630 --> 04:59.580 by the time I put the video live 04:59.580 --> 05:02.520 and I recommend you don't share yours with anyone. 05:02.520 --> 05:04.860 But once you have the key... 05:04.860 --> 05:07.050 So like if I create a new key, 05:07.050 --> 05:08.970 it'll say like, create, 05:08.970 --> 05:12.630 and then it'll download some JSON to my folder. 05:12.630 --> 05:15.540 And then you basically just then upload it here. 05:15.540 --> 05:17.160 I've renamed it. 05:17.160 --> 05:19.255 So you can click into a file and go rename 05:19.255 --> 05:22.740 and I've renamed it "Service account JSON," 05:22.740 --> 05:23.855 just to make the code run. 05:23.855 --> 05:25.515 But you can just set yours 05:25.515 --> 05:29.310 to whatever your thing is named here. 05:29.310 --> 05:34.310 So if you run that fine and then you go in here. 05:34.710 --> 05:37.260 So this is the actual code that finds the labels. 05:37.260 --> 05:38.610 So this is just a test run. 05:38.610 --> 05:40.710 We'll try it with one of the images. 05:40.710 --> 05:42.720 In this case it's gonna be, 05:42.720 --> 05:44.043 let's say image 10. 05:44.970 --> 05:47.395 So or the, I guess that's the 11th image 05:47.395 --> 05:49.890 because it starts counting from zero. 05:49.890 --> 05:52.375 So it's gonna open the image. 05:52.375 --> 05:54.570 So it's gonna read the image, 05:54.570 --> 05:58.980 then it's gonna turn that into a Google vision image. 05:58.980 --> 06:00.660 And then when it's in that format, 06:00.660 --> 06:03.870 then you can get the label detection running. 06:03.870 --> 06:05.940 And there's actually a bunch of other things you can do 06:05.940 --> 06:07.350 with this, but label detection 06:07.350 --> 06:09.450 is the most immediately useful. 06:09.450 --> 06:11.700 And then once you get the response, 06:11.700 --> 06:14.940 so this calls the API for Google Vision 06:14.940 --> 06:16.470 and it takes a few seconds, 06:16.470 --> 06:18.810 you get the response and then you can get the labels 06:18.810 --> 06:20.250 as well as other things as well. 06:20.250 --> 06:22.770 But labels is all we need. 06:22.770 --> 06:24.570 And then what we're gonna do 06:24.570 --> 06:25.485 here is just to test, it's working. 06:25.485 --> 06:27.405 We're just gonna display the image 06:27.405 --> 06:28.920 with the labels underneath. 06:28.920 --> 06:31.203 So let's run that and see if it's working. 06:32.841 --> 06:33.674 Here we go. 06:33.674 --> 06:34.507 Yeah, it worked. 06:34.507 --> 06:35.640 So this was our image. 06:35.640 --> 06:37.920 And then we got the labels of hair, 06:37.920 --> 06:40.290 joint, head, shoulder, eye, 06:40.290 --> 06:41.310 flash photography, neck, 06:41.310 --> 06:42.336 sleeve, dress, thigh, 06:42.336 --> 06:44.970 and we can check other images as well. 06:44.970 --> 06:47.313 So lemme try another one. 06:48.990 --> 06:50.850 This is the first image. 06:50.850 --> 06:51.843 So zero. 06:52.830 --> 06:56.370 So this is a lemon, picked up the blue 06:56.370 --> 06:58.620 and the sphere and still life. 06:58.620 --> 07:01.260 And then you can just basically, you know, 07:01.260 --> 07:04.365 kind of try out all the different images if you like, 07:04.365 --> 07:06.510 kinda see what it comes up with. 07:06.510 --> 07:09.035 It's not always fully accurate. 07:09.035 --> 07:12.305 I think it's a good kind of directional understanding. 07:12.305 --> 07:14.640 Sometimes it doesn't get them right. 07:14.640 --> 07:16.505 These images themselves are actually 07:16.505 --> 07:18.300 generated by AI as well. 07:18.300 --> 07:22.470 So sometimes it does get confused on these 07:22.470 --> 07:27.470 more than it does with the real images that you know 07:28.380 --> 07:30.360 that you would use from your account. 07:30.360 --> 07:31.193 Cool. 07:31.193 --> 07:32.310 So now that we know it's working, 07:32.310 --> 07:34.650 we can just run this to process all the images. 07:34.650 --> 07:37.020 So it's gonna go through every single image 07:37.020 --> 07:38.940 and do the same thing again. 07:38.940 --> 07:41.280 Except instead of just printing the labels, 07:41.280 --> 07:42.810 although it will print them, 07:42.810 --> 07:45.360 it will also then add them to a list. 07:45.360 --> 07:47.115 And then we're gonna use that list afterwards 07:47.115 --> 07:49.535 to join the data together. 07:49.535 --> 07:51.300 So I'm not gonna run that again 07:51.300 --> 07:53.885 because it's already run here, 07:53.885 --> 07:55.380 takes a few minutes, 07:55.380 --> 07:57.870 but it depends on how many images you have. 07:57.870 --> 07:59.100 But you can see that, you know... 07:59.100 --> 08:00.960 Here it can tell that that's a peach, 08:00.960 --> 08:02.880 which is pretty cool. 08:02.880 --> 08:06.655 It thinks that this is a toy or maybe some office supplies. 08:06.655 --> 08:09.420 So you can see this isn't quite get it, 08:09.420 --> 08:12.000 but you know, these, I think these fashion photography ones 08:12.000 --> 08:15.120 seem to work a little bit better. 08:15.120 --> 08:18.003 So, you know, it's picking up quite a few things here. 08:18.870 --> 08:20.610 So it's quite interesting. 08:20.610 --> 08:22.775 So you can see all the logos as it comes through. 08:22.775 --> 08:26.670 And then if you scroll down to the end, 08:26.670 --> 08:27.840 now this is where we start 08:27.840 --> 08:29.340 to get the data into the right format. 08:29.340 --> 08:33.690 So what this does is it just goes through every label 08:33.690 --> 08:37.050 and then for every label it adds it with the creative. 08:37.050 --> 08:39.973 So you know, if C 39 had, you know, 08:39.973 --> 08:44.175 C 39 creative had the label of ball and solo photography 08:44.175 --> 08:47.100 and circle and sphere and electric blue. 08:47.100 --> 08:49.210 So this just showing the first five rows 08:50.070 --> 08:53.910 and then we're basically gonna then join that together 08:53.910 --> 08:58.020 with our other data, the metrics DF, 08:58.020 --> 08:59.850 which was the impression, 08:59.850 --> 09:01.220 click, spend, et cetera. 09:01.220 --> 09:03.900 So that just does merge. 09:03.900 --> 09:05.070 And there we go. 09:05.070 --> 09:09.415 Now we can see, 'cause C 39 had the label of ball, 09:09.415 --> 09:13.920 we've given that the metrics from C 39 09:13.920 --> 09:16.745 and it's given that for each one of the labels. 09:16.745 --> 09:19.290 And then what we're gonna do is kind of join it all together 09:19.290 --> 09:23.370 so that, you know, if multiple creators had, you know, 09:23.370 --> 09:25.230 the circle label, 09:25.230 --> 09:27.545 then they're gonna add up all of their metrics. 09:27.545 --> 09:29.583 So that's what this does here, 09:30.840 --> 09:32.280 it's gonna run this. 09:32.280 --> 09:33.780 And there we go. 09:33.780 --> 09:36.990 So this is kind of the beginning of the table 09:36.990 --> 09:39.840 and what this has done is it's kind of grouped 09:39.840 --> 09:44.280 by the creative and then just counted how many creatives 09:44.280 --> 09:47.190 it was in and kind of spat back 09:47.190 --> 09:48.900 which creatives it was in as well. 09:48.900 --> 09:53.430 So seeking the action figure was in C 44, 09:53.430 --> 09:55.020 an arm was tagged in C3, 09:55.020 --> 09:56.493 C 34, and C 25. 09:57.968 --> 10:01.530 So we're gonna use that in a second. 10:01.530 --> 10:04.745 The only thing left to do really is to 10:04.745 --> 10:06.665 attach all the metrics. 10:06.665 --> 10:10.590 So again, this is just going through each column 10:10.590 --> 10:14.010 and then getting just the metrics data, 10:14.010 --> 10:18.030 so from column to onwards and then grouping by 10:18.030 --> 10:21.510 and summing it and then kind of adding it into that column. 10:21.510 --> 10:24.420 So every creative, so you know, 10:24.420 --> 10:26.310 this one action figure, 10:26.310 --> 10:28.143 it was just in creative C 44. 10:29.400 --> 10:33.480 The reach impressions clicks will be just from C 44, 10:33.480 --> 10:37.020 but this one arm was in C 3, C 34, and C 25. 10:37.020 --> 10:38.970 So it has the combination, 10:38.970 --> 10:41.250 the sum of all those metrics. 10:41.250 --> 10:42.125 The other thing I did was just 10:42.125 --> 10:43.925 calculated the click-through rate. 10:43.925 --> 10:45.270 That's what I'm gonna look at for this. 10:45.270 --> 10:47.940 We don't have a lot of conversions, 10:47.940 --> 10:50.190 but here it's just the clicks divided 10:50.190 --> 10:52.050 by impressions times by a hundred. 10:52.050 --> 10:54.450 And that's what I'm kind of using to tell 10:54.450 --> 10:56.830 which creatives were good or not. 10:56.830 --> 10:57.663 Cool. 10:57.663 --> 10:59.130 So one thing you'll notice is that like, 10:59.130 --> 11:00.360 if it's only in one creative, 11:00.360 --> 11:01.590 it doesn't tell as much. 11:01.590 --> 11:03.570 So it's saying action figures, you know, 11:03.570 --> 11:05.490 ranked the same as art, 11:05.490 --> 11:07.560 1.18% click through rate. 11:07.560 --> 11:09.990 'Cause that was the click through rate for this one ad. 11:09.990 --> 11:12.750 So what's more interesting if it's in multiple creators, 11:12.750 --> 11:14.040 but it wouldn't be that interesting 11:14.040 --> 11:15.270 if it was in all the creatives 11:15.270 --> 11:17.070 'cause then that would just be the average, right? 11:17.070 --> 11:20.790 So what we want is any, you know, 11:20.790 --> 11:22.565 we don't want it in all creatives. 11:22.565 --> 11:27.095 So this is saying basically filter for any time 11:27.095 --> 11:31.110 where the creative count was equal 11:31.110 --> 11:33.675 or you know, it's basically it needs to be 11:33.675 --> 11:36.905 less than the length of creative names. 11:36.905 --> 11:38.090 And then this three 11:38.090 --> 11:41.430 or more filter basically, you know, 11:41.430 --> 11:46.430 needs to have at least three in the creative count call. 11:47.400 --> 11:49.037 Cool, so then the filter DF, 11:49.037 --> 11:51.450 I've also sorted it by values. 11:51.450 --> 11:53.700 So sorted by click through rate descending, 11:53.700 --> 11:55.620 and then that shows the head here. 11:55.620 --> 11:58.440 So you can run that and you can see, 11:58.440 --> 11:59.280 there we go. 11:59.280 --> 12:02.626 And now we can start to tell, you know, 12:02.626 --> 12:06.325 what types of tags are showing up as performant. 12:06.325 --> 12:09.420 And you could look at the head 12:09.420 --> 12:13.740 and see that this one has like, 12:13.740 --> 12:15.150 it's only in three creatives, 12:15.150 --> 12:17.730 but it's actually like quite a high click-through rate. 12:17.730 --> 12:19.200 When the circle thing, 12:19.200 --> 12:21.720 I think this is showing up in the labels, 12:21.720 --> 12:25.487 like in the specific, you know, 12:25.487 --> 12:29.700 the kind of the Claymation emoji thing. 12:29.700 --> 12:31.500 So that tends to show up there. 12:31.500 --> 12:33.505 So that's an interesting thing to look at. 12:33.505 --> 12:35.670 We also see neck, when we show the neck, 12:35.670 --> 12:36.930 the click rate is pretty high. 12:36.930 --> 12:38.640 Flash photography is good, 12:38.640 --> 12:40.080 sleeve, forehead. 12:40.080 --> 12:43.170 Okay, so that's kind of telling us roughly what's working. 12:43.170 --> 12:44.385 We could also look at what's not working. 12:44.385 --> 12:49.385 So is the pale and we can see the very end. 12:50.580 --> 12:52.410 So these are the things with the lowest click rate. 12:52.410 --> 12:56.160 So when it notices, there's a hairstyle that seems 12:56.160 --> 12:57.453 to perform worse. 12:58.440 --> 13:00.537 When there's a white color in there 13:00.537 --> 13:03.378 that seems to perform worse, 13:03.378 --> 13:05.385 then this electric blue color. 13:05.385 --> 13:08.520 So that's in some of the Claymation ones, 13:08.520 --> 13:09.600 it's performing worse. 13:09.600 --> 13:13.350 So you start to get a sense of what sorts of things 13:13.350 --> 13:14.580 are improving performance 13:14.580 --> 13:18.000 and what sorts of things are decreasing in performance. 13:18.000 --> 13:19.350 And these are just correlations. 13:19.350 --> 13:22.005 So you do need to test these hypotheses, 13:22.005 --> 13:24.185 but this is a great way to find new ideas 13:24.185 --> 13:25.833 of what's a test. 13:27.480 --> 13:29.330 All right, hopefully this was useful.