WEBVTT 00:00.690 --> 00:03.150 -: Okay, say you're designing a coffee shop, 00:03.150 --> 00:06.630 and you want to figure out what are the typical interiors 00:06.630 --> 00:07.800 for your competitors. 00:07.800 --> 00:09.120 So, you wanna understand what type 00:09.120 --> 00:12.030 of furniture they're using, whether they have plants, 00:12.030 --> 00:14.670 or whether there's a chalkboard on the wall 00:14.670 --> 00:16.590 for the menu, like this, 00:16.590 --> 00:18.270 and you want to do that analysis 00:18.270 --> 00:21.930 and see how often certain things come up versus not. 00:21.930 --> 00:24.930 So, one interesting way to do this 00:24.930 --> 00:28.350 is instead of manually looking through a bunch of images, 00:28.350 --> 00:31.260 you could tag them automatically with AI, 00:31.260 --> 00:35.070 because GPT Vision is actually pretty good at this now. 00:35.070 --> 00:38.880 What you need is you have to scrape Google Images, 00:38.880 --> 00:40.710 which is actually the hard part, (chuckles) 00:40.710 --> 00:43.370 but thankfully, there's a tool that does this for you, 00:43.370 --> 00:46.290 so if you just search for Google-Image-Scraper, 00:46.290 --> 00:49.590 yeah, literally just search for Google-Image-Scraper, 00:49.590 --> 00:51.780 and that was the first result, 00:51.780 --> 00:54.660 but however you get the data, it doesn't really matter, 00:54.660 --> 00:56.400 but I assure you it works. 00:56.400 --> 01:00.990 I just, this is in the main.py, 01:00.990 --> 01:04.380 so, when you download this to your computer from GitHub, 01:04.380 --> 01:06.780 then you have this code. 01:06.780 --> 01:10.260 You just put in here what are the different keywords 01:10.260 --> 01:11.130 that you wanna search for? 01:11.130 --> 01:13.500 I downloaded a bunch, and you can see a bunch 01:13.500 --> 01:18.500 of coffee shops in Beijing, coffee shops in Rome, 01:18.540 --> 01:20.700 and then also New York. 01:20.700 --> 01:22.650 We're gonna compare them, potentially. 01:22.650 --> 01:26.430 So, I'm gonna just run an abbreviated part of this analysis. 01:26.430 --> 01:29.550 I would say, by the way, it takes a while to download them, 01:29.550 --> 01:32.850 so, here, I downloaded a bunch, and it doesn't always work. 01:32.850 --> 01:37.260 So, you can see, you get 18 every time, 36 in Sydney, 01:37.260 --> 01:40.680 but once you literally just run Python main.py, 01:40.680 --> 01:43.680 you should see these images start appearing, 01:43.680 --> 01:45.210 but that's not the part I'm gonna teach you, 01:45.210 --> 01:47.850 because that is someone else's code, 01:47.850 --> 01:51.060 and I'll let them teach you that in the README, 01:51.060 --> 01:52.500 but what I'm gonna teach you is what to do 01:52.500 --> 01:54.030 when you have the data, right? 01:54.030 --> 01:58.380 This is a script for the GPT Vision API, 01:58.380 --> 02:00.990 and you just need to run these libraries here, 02:00.990 --> 02:02.940 but I have all the code written here 02:02.940 --> 02:06.417 to pull in what are the different coffee shops, right? 02:06.417 --> 02:08.850 You can see how many photos you've got, 02:08.850 --> 02:10.950 I just looked at the first five here. 02:10.950 --> 02:13.350 See, I've got 45 images from New York, 02:13.350 --> 02:15.420 got a bunch of Sydney, et cetera, 02:15.420 --> 02:18.840 so, what I'm gonna do is I'm just gonna 02:18.840 --> 02:21.360 just little bit just to, wait. 02:21.360 --> 02:24.750 Let's just do Sydney just for testing, 02:24.750 --> 02:27.540 because Vision API is expensive. 02:27.540 --> 02:29.430 If you're running it a bunch of times, 02:29.430 --> 02:30.877 it also takes a little bit of time, 02:30.877 --> 02:32.850 so I'm just gonna write it on these 33 images 02:32.850 --> 02:35.070 in Sydney, right? 02:35.070 --> 02:36.660 And there's two parts to this. 02:36.660 --> 02:38.100 One is inductive coding, 02:38.100 --> 02:40.260 and then the second is deductive coding. 02:40.260 --> 02:42.510 Inductive coding is when you're looking at these, 02:42.510 --> 02:43.343 and you look for patterns, 02:43.343 --> 02:46.860 and you compare them, you go, "Okay, that one has a coffee." 02:46.860 --> 02:50.010 You know, that one has a chalkboard for the menu, 02:50.010 --> 02:52.200 and then there's other ones that have a chalkboard 02:52.200 --> 02:53.490 for menus, as well, 02:53.490 --> 02:55.260 so, maybe that's something that... 02:55.260 --> 02:56.850 Oh, here we go, there's another chalkboard, 02:56.850 --> 02:58.380 so, I've noticed that pattern, 02:58.380 --> 03:01.140 and now that I know there's a pattern, 03:01.140 --> 03:02.430 I wanna find that everywhere. 03:02.430 --> 03:04.980 So, there's another chalkboard, 03:04.980 --> 03:07.440 there's another chalkboard, so, that's deductive coding. 03:07.440 --> 03:10.740 Inductive coding is coming up with the patterns, 03:10.740 --> 03:13.200 comparing and contrasting to find what's different 03:13.200 --> 03:14.580 or what's similar, 03:14.580 --> 03:16.770 and then deductive coding is then going 03:16.770 --> 03:19.830 and finding those patterns everywhere else. 03:19.830 --> 03:24.630 So, we're gonna do those two things here with GPT Vision. 03:24.630 --> 03:25.620 All right, so, the first thing we're doing 03:25.620 --> 03:26.730 is inductive coding, 03:26.730 --> 03:29.970 and we need to be able to send an image to OpenAI, 03:29.970 --> 03:32.700 so, the way that you send an image 03:32.700 --> 03:35.100 is with the Base64 encoding. 03:35.100 --> 03:37.110 This literally just takes an image, 03:37.110 --> 03:39.090 and then turns it into some text 03:39.090 --> 03:41.493 that you can send along with your API call. 03:42.360 --> 03:44.910 And all this code does is it just looks in the folder, 03:44.910 --> 03:47.460 and then it just chooses two images at random 03:47.460 --> 03:49.620 from the different folders, 03:49.620 --> 03:51.750 so, we just have the one folder right now, 03:51.750 --> 03:53.370 but if we have multiple folders, 03:53.370 --> 03:55.290 it would look at all of them. 03:55.290 --> 03:56.970 Oh, I did mention, I forget to mention one thing. 03:56.970 --> 03:59.160 When you're using a Google-Image-Scraper, 03:59.160 --> 04:00.870 you'll probably need to download, 04:00.870 --> 04:04.980 I had to download this mac-x64 Chrome for testing, 04:04.980 --> 04:09.060 but when it's creeping these images, you need that in there. 04:09.060 --> 04:11.220 Here, I'll just put us in a folder called webdriver, 04:11.220 --> 04:13.519 and then it opens that. 04:13.519 --> 04:17.010 I'm just mentioning it because I ran into it. 04:17.010 --> 04:19.350 All right, so, back to the inductive coding. 04:19.350 --> 04:20.550 So, it's choosing an image pair, 04:20.550 --> 04:22.620 it's choosing two images at random, 04:22.620 --> 04:25.050 and then it's comparing them. 04:25.050 --> 04:27.000 So, it's getting the Base64 string, 04:27.000 --> 04:29.257 and then it's using this prompt here. 04:29.257 --> 04:32.190 "Compare these two coffee shops using inductive coding. 04:32.190 --> 04:34.410 What features, attributes, or elements of design 04:34.410 --> 04:35.550 are similar or different?" 04:35.550 --> 04:38.250 And then I'm asking for a structured response here. 04:38.250 --> 04:41.010 That's just how you pass it into GPT Vision, by the way. 04:41.010 --> 04:44.010 You just client.chat.completions, 04:44.010 --> 04:45.780 and then you pass the prompt, 04:45.780 --> 04:48.570 but then you also pass in the two images, as well, 04:48.570 --> 04:50.850 in the same content string. 04:50.850 --> 04:52.590 I'm just using the user prompt here, 04:52.590 --> 04:55.770 but you could also have a system prompt at the top. 04:55.770 --> 04:58.290 So, just run this, see what we've got. 04:58.290 --> 05:00.810 So, it's selected two of these images at random, 05:00.810 --> 05:03.720 number 25 and number 31, 05:03.720 --> 05:06.030 and you can see it takes a little bit of time 05:06.030 --> 05:07.830 for the image response, 05:07.830 --> 05:10.410 as this is OpenAI's biggest model. 05:10.410 --> 05:12.960 Here we go, so now, we've got some labels 05:12.960 --> 05:14.820 that there's some things that they share. 05:14.820 --> 05:18.960 So, coffee shop A and B both have modern designs, 05:18.960 --> 05:20.970 they both have wooden elements. 05:20.970 --> 05:23.790 Only Coffee Shop B has natural light, 05:23.790 --> 05:26.730 and only coffee shop A has spacious interior, right? 05:26.730 --> 05:28.290 So, that's really interesting. 05:28.290 --> 05:30.897 Now, we have some labels, we have a dictionary of labels, 05:30.897 --> 05:33.510 and we can run that and then save it. 05:33.510 --> 05:38.510 So, if we look at, we should have here some cleaned labels. 05:39.180 --> 05:42.510 So, you can see, yeah, there's a bunch of labels here 05:42.510 --> 05:44.370 for these different things, and it's rewritten them. 05:44.370 --> 05:47.490 We've got the actual kind of folder that they're in, 05:47.490 --> 05:49.593 the label, and then the font. 05:50.430 --> 05:51.870 Cool, so now, you have your label dictionary, 05:51.870 --> 05:53.820 and that's useful in itself, 05:53.820 --> 05:55.890 because you could run that multiple times 05:55.890 --> 05:57.630 across multiple images, 05:57.630 --> 05:59.220 so, we just compared two of those images. 05:59.220 --> 06:01.620 You can compare every single image against each other. 06:01.620 --> 06:02.790 Just take a little bit of time. 06:02.790 --> 06:04.110 That, you can run again, as well, 06:04.110 --> 06:06.870 and get a bigger dictionary of data. 06:06.870 --> 06:10.410 Now, it comes to the deductive coding, 06:10.410 --> 06:13.080 and this is the one that's more time consuming, 06:13.080 --> 06:14.160 more expensive. 06:14.160 --> 06:17.774 The reason why I've set this up as asynchronous 06:17.774 --> 06:20.040 is 'cause otherwise it takes a very long time to run, 06:20.040 --> 06:22.140 'cause you're going through every single image, and saying, 06:22.140 --> 06:23.797 and every single label, and going, 06:23.797 --> 06:25.620 "Is this label in this image? 06:25.620 --> 06:27.660 Does this coffee shop have wooden elements? 06:27.660 --> 06:29.820 Does this coffee shop have wooden elements? 06:29.820 --> 06:31.800 Does this coffee shop have wooden elements," right? 06:31.800 --> 06:33.750 And then you're going through all the other labels as well. 06:33.750 --> 06:36.210 Does this coffee shop have suspended lighting in it? 06:36.210 --> 06:38.490 Does this coffee shop have seating areas? 06:38.490 --> 06:41.310 There are different tricks for doing this faster or slower, 06:41.310 --> 06:44.280 but for me, I found if I do asynchronous, 06:44.280 --> 06:48.870 means I can get all the labels at once for a single image, 06:48.870 --> 06:51.450 and they all, you know, run quite quickly, 06:51.450 --> 06:53.310 and so, it runs a little bit faster. 06:53.310 --> 06:56.070 So, again, here, we're just passing one image in, 06:56.070 --> 06:57.600 and we're just passing the user prompt, 06:57.600 --> 06:58.740 which is much simpler. 06:58.740 --> 07:01.230 If the label applies to this image, return one, 07:01.230 --> 07:04.320 otherwise return zero, only output one or zero. 07:04.320 --> 07:05.670 All right, if we get a one, 07:05.670 --> 07:08.400 then we are gonna return the label, 07:08.400 --> 07:11.760 and if we don't, then we're not, we're gonna return none. 07:11.760 --> 07:15.210 So, this what this does, this is the asynchronous code. 07:15.210 --> 07:16.800 We just set up a bunch of tasks, 07:16.800 --> 07:18.480 so, we run it for every image, 07:18.480 --> 07:20.430 and then we wait until all the tasks 07:20.430 --> 07:21.450 are done for that image, 07:21.450 --> 07:24.570 it's processing all the labels at the same time, 07:24.570 --> 07:28.080 and OpenAI can handle multiple API calls at the same time. 07:28.080 --> 07:30.210 What will happen is instead of waiting 07:30.210 --> 07:31.590 for each one to come back, 07:31.590 --> 07:33.510 it can process all of them at the same time. 07:33.510 --> 07:36.840 We're gonna get through, say, if we're doing 10 labels, 07:36.840 --> 07:38.790 we'll get through all 10 at the same time 07:38.790 --> 07:40.260 rather than waiting. 07:40.260 --> 07:42.990 If one takes a minute, we're not waiting 10 minutes, 07:42.990 --> 07:44.700 we're getting all of them in one minute, 07:44.700 --> 07:47.220 so, that's the benefit of asynchronous. 07:47.220 --> 07:49.170 Cool, and then we're just gonna return the labels 07:49.170 --> 07:52.320 if the, you know, if the label is in there, right? 07:52.320 --> 07:54.360 And then actually running deductive code, 07:54.360 --> 07:57.167 we just pull open the cleaned_labels.json 07:57.167 --> 07:58.110 that we had before, 07:58.110 --> 07:59.940 and then we have the cleaned labels 07:59.940 --> 08:03.300 which loads the JSON from that dictionary we had, 08:03.300 --> 08:05.040 and then it finds the values. 08:05.040 --> 08:07.950 I've also limited it to just five labels per coffee shop. 08:07.950 --> 08:10.560 Otherwise, it'd take a bit of time. 08:10.560 --> 08:13.110 Cool, all right, so, let me just run this, 08:13.110 --> 08:14.970 and we're processing coffeeshopsydney. 08:14.970 --> 08:16.230 You can see that all three of those 08:16.230 --> 08:18.150 came back at the same time, 08:18.150 --> 08:20.640 but some of them are taking a little bit longer, 08:20.640 --> 08:22.770 but we're not waiting for one to finish 08:22.770 --> 08:24.300 before we get the other. 08:24.300 --> 08:25.133 Cool, so, here we go. 08:25.133 --> 08:27.150 So now, we're processing coffeeshopsydney, 08:29.280 --> 08:32.010 and we'll see how quickly they come back. 08:32.010 --> 08:33.520 Yeah, this is Sydney30. 08:33.520 --> 08:36.330 Okay, now, we're processing Sydney10, 08:36.330 --> 08:37.950 we're getting hanging light fixtures. 08:37.950 --> 08:40.470 So, this is only coming back as positive 08:40.470 --> 08:43.473 if it does have that label, if that makes sense. 08:45.660 --> 08:49.950 Cool, that's where we'll move into analysis but with Vision. 08:49.950 --> 08:52.260 It's really powerful once you figure 08:52.260 --> 08:54.240 how you can apply this to pretty much anything, 08:54.240 --> 08:56.190 like, anytime you have a lot of images, 08:56.190 --> 08:59.520 like maybe in your ads, or on your website, 08:59.520 --> 09:02.640 you can do some analysis to figure out what labels, 09:02.640 --> 09:04.560 and you can define the labels, as well. 09:04.560 --> 09:07.680 You could tell it what sorts of things you're looking for, 09:07.680 --> 09:09.390 give it a list of things. 09:09.390 --> 09:11.100 However you want to do it, that's fine, too. 09:11.100 --> 09:12.420 You could just add new stuff 09:12.420 --> 09:14.220 to this cleaned labels thing, right? 09:14.220 --> 09:17.790 You could just add in some new ideas, right? 09:17.790 --> 09:19.860 Like, I wanna find... 09:19.860 --> 09:21.480 So, this is customers_queuing, right? 09:21.480 --> 09:23.820 I want to find out if there's a barista 09:23.820 --> 09:25.050 that has a tattoo, right? 09:25.050 --> 09:27.960 Like, you could just add that as another option. 09:27.960 --> 09:32.063 Copy, this one has tattooed barista, right? 09:35.160 --> 09:37.350 And it is really funny if the barista's present in this, 09:37.350 --> 09:39.150 but then we're gonna see if they have tattoos, 09:39.150 --> 09:40.260 visible tattoos. 09:40.260 --> 09:41.460 That's how this works, 09:41.460 --> 09:43.110 and then what you're gonna get back 09:43.110 --> 09:45.510 is you're gonna get a file name like this. 09:45.510 --> 09:49.770 This is just a big list of all the different labels, right? 09:49.770 --> 09:53.100 So, it says, coffeeshopnewyork42 09:53.100 --> 09:55.080 has the label, Street View. 09:55.080 --> 09:58.410 This one has the label, Patrons, present, 09:58.410 --> 10:01.110 and then you can do some analysis after this. 10:01.110 --> 10:03.690 Basically, figure out, okay, how many times, 10:03.690 --> 10:06.840 yeah, how many labels does this apply to? 10:06.840 --> 10:09.600 So, how many of the coffee shops has this label? 10:09.600 --> 10:10.860 I'll give you an example. 10:10.860 --> 10:12.240 I'll show you an example here, 10:12.240 --> 10:14.160 because that's still processing, 10:14.160 --> 10:16.080 because it's going through 33 images. 10:16.080 --> 10:19.193 Here's an example from New York that I ran earlier. 10:19.193 --> 10:23.340 So, you can get a list of all the labels into a data frame, 10:23.340 --> 10:25.380 this is like an Excel spreadsheet, 10:25.380 --> 10:27.810 and then you can print out what percentage 10:27.810 --> 10:30.510 of the total images have these different labels. 10:30.510 --> 10:35.130 So here, we're finding that 41% of the patrons, 10:35.130 --> 10:39.330 40% of the coffee shop images show patrons present. 10:39.330 --> 10:41.760 So, to me, you wanna really pay attention 10:41.760 --> 10:44.100 to the ones that are on almost everything, 10:44.100 --> 10:46.200 'cause that's almost like a rule of the game. 10:46.200 --> 10:48.900 If you don't have patrons present in your coffee shop image, 10:48.900 --> 10:50.160 that's probably really bad, 10:50.160 --> 10:53.070 and you definitely want to at least have a stance on that. 10:53.070 --> 10:54.300 Like, maybe have a good reason 10:54.300 --> 10:56.400 for not having patrons present in the image 10:56.400 --> 10:59.160 of your coffee shop when you're uploading that image 10:59.160 --> 11:01.620 to, you know, Google my business, 11:01.620 --> 11:03.390 but you should at least address that. 11:03.390 --> 11:05.880 Have a strategic choice here. 11:05.880 --> 11:09.150 Similarly, for things that are not very popular, 11:09.150 --> 11:10.680 you also might want to think about these 11:10.680 --> 11:11.760 as a differentiator. 11:11.760 --> 11:15.120 So, exposed brick interior, you could argue 11:15.120 --> 11:17.190 that that's not very fashionable anymore, 11:17.190 --> 11:19.530 and that's why only 10% of people have had it, 11:19.530 --> 11:22.590 or you could say, "This is gonna be fashionable. 11:22.590 --> 11:24.390 I can see that this is a trend, 11:24.390 --> 11:27.090 so, I'm gonna start using this," right? 11:27.090 --> 11:30.000 That's how you do this type of analysis. 11:30.000 --> 11:31.590 Obviously, what it doesn't tell you is, like, 11:31.590 --> 11:34.950 which choices to make and what designs to have, 11:34.950 --> 11:37.587 but if you're, you know, not a designer like me, 11:37.587 --> 11:40.830 and you want to build a bit more data to make your decision, 11:40.830 --> 11:42.270 this can help inform your decision, 11:42.270 --> 11:45.540 give you a map of what are the different creative decisions 11:45.540 --> 11:48.900 you need to make when you're doing something designed 11:48.900 --> 11:50.490 or visual, right? 11:50.490 --> 11:51.750 Hopefully that's useful. 11:51.750 --> 11:54.330 This is my little hack for being a data guy, 11:54.330 --> 11:56.850 but also playing in the world of branding. 11:56.850 --> 11:58.550 Hopefully, you find it useful too.