I just decided to do Kaggle competitions more seriously and I started with this competition (it's not my first deep learning project in all btw. I've worked on finished comps to learn and build new models but this one was going to be the first serious one for me). I'm fairly familiar with computer vision in general but it was my first time working on an Image Captioning task (just read a tutorial months ago).
\nIt was really disappointing for me that it took me about a whole day (say 14 or more hours of coding in total) to build a working pipeline. I was mostly using others' kernels and codes to just build a pipeline that works (and also learning what is happening in there!) and categorize the code into different modules to start from there and improve upon it, but this task that I thought would be easy, took me this lot of time! I want to ask you guys who are more experienced than me that is this okay?!
\nI'm actually a medical student and last year I started learning python programming and deep learning and I don't have a coding background beside this. So, I was wondering if it is normal that such a task takes this lot of time OR the problem is my coding and I need to do it more efficiently.
\nI'll be really glad if you let me know your experiences and comments!
I just decided to do Kaggle competitions more seriously and I started with this competition (it's not my first deep learning project in all btw. I've worked on finished comps to learn and build new models but this one was going to be the first serious one for me). I'm fairly familiar with computer vision in general but it was my first time working on an Image Captioning task (just read a tutorial months ago).
\nIt was really disappointing for me that it took me about a whole day (say 14 or more hours of coding in total) to build a working pipeline. I was mostly using others' kernels and codes to just build a pipeline that works (and also learning what is happening in there!) and categorize the code into different modules to start from there and improve upon it, but this task that I thought would be easy, took me this lot of time! I want to ask you guys who are more experienced than me that is this okay?!
\nI'm actually a medical student and last year I started learning python programming and deep learning and I don't have a coding background beside this. So, I was wondering if it is normal that such a task takes this lot of time OR the problem is my coding and I need to do it more efficiently.
\nI'll be really glad if you let me know your experiences and comments!
Yes.
\nIts normal…. Some time for me it takes even longer if I am not familiar with the method.. Just make sure to document and organize your code. In future if you encounter similar project you can reuse a lot of your old code =)
Such competitions usually go on for several months. And on some of them I didn't have time to create a model that would suit me and didn't send anything. So I'm really jealous that you created a working model in just one day.
", "rawMarkdown": "Such competitions usually go on for several months. And on some of them I didn't have time to create a model that would suit me and didn't send anything. So I'm really jealous that you created a working model in just one day.", "votes": null }, { "id": "1248413", "postDate": "03/22/2021 15:29:37", "content": "Thanks for reassuring and sharing your experience. Yeah I found out that I do really need to document literally \"everything\"! Because there are many nuances and details in this task compared to say vanilla image classification!
", "rawMarkdown": "Thanks for reassuring and sharing your experience. Yeah I found out that I do really need to document literally \"everything\"! Because there are many nuances and details in this task compared to say vanilla image classification!", "votes": null }, { "id": "1248417", "postDate": "03/22/2021 15:32:18", "content": "Yeah it seems to be a really long and complicated competition in a sense that there are a lot of things to try out.
\nActually more than 80 percent of the pipeline I prepared was others' effort translated into my coding habits :) so I think I did not make a break through to do it in one day and I'm sure you can do what I did in the same or even less time :)
It's important to have a toolkit and standard working pipeline that you can adapt on different competitions.
\nYou will make it flexible and robust through competitions and you'll gain lot of time.
\nDon't worry if it takes long time to build.
", "rawMarkdown": "It's important to have a toolkit and standard working pipeline that you can adapt on different competitions. \n\nYou will make it flexible and robust through competitions and you'll gain lot of time. \n\nDon't worry if it takes long time to build.", "votes": null }, { "id": "1248443", "postDate": "03/22/2021 15:52:09", "content": "Thanks for pointing it out. So I'll look at it as an investment for similar competitions or personal projects! I think this mindset will help me out.
", "rawMarkdown": "Thanks for pointing it out. So I'll look at it as an investment for similar competitions or personal projects! I think this mindset will help me out.", "votes": null }, { "id": "1248455", "postDate": "03/22/2021 15:59:44", "content": "You may find inspiration in this great repo by @analokamus
\nhttps://github.com/analokmaus/kuma_utils
", "rawMarkdown": "You may find inspiration in this great repo by @analokamus \n\n[https://github.com/analokmaus/kuma_utils](https://github.com/analokmaus/kuma_utils)", "votes": null }, { "id": "1248465", "postDate": "03/22/2021 16:04:30", "content": "Thanks! I was looking for such a thing.
", "rawMarkdown": "Thanks! I was looking for such a thing.", "votes": null }, { "id": "1248661", "postDate": "03/22/2021 18:29:24", "content": "I usually do a minimal viable product sort of thing where I am at least able to train something, not an actual pipeline. This can take anywhere from 30m -2h, but once I move on to a pipeline the mvp makes it really easy to set up. The rest is just dependent on how big of a dataset I have to manage. Building the pipeline can take anywhere from 30m - 4h. If I reuse old code it can be considerably faster depending on the overlap.
", "rawMarkdown": "I usually do a minimal viable product sort of thing where I am at least able to train something, not an actual pipeline. This can take anywhere from 30m -2h, but once I move on to a pipeline the mvp makes it really easy to set up. The rest is just dependent on how big of a dataset I have to manage. Building the pipeline can take anywhere from 30m - 4h. If I reuse old code it can be considerably faster depending on the overlap.", "votes": null }, { "id": "1248726", "postDate": "03/22/2021 19:36:35", "content": "Thanks for sharing your way of doing it. Yeah, I need to divide what I do into smaller steps. In this way I can build a much more re-usable code base.
", "rawMarkdown": "Thanks for sharing your way of doing it. Yeah, I need to divide what I do into smaller steps. In this way I can build a much more re-usable code base.", "votes": null }, { "id": "1249790", "postDate": "03/23/2021 14:16:10", "content": "It takes hundreds of hours to do well in a competition.
\nOne article on the Zillow prize winners
\n\n\n", "rawMarkdown": "It takes **hundreds of hours** to do well in a competition. \n\nOne article on the Zillow prize winners\n\nhttps://www.geekwire.com/2019/meet-zillow-prize-winners-get-1m-bragging-rights-zestimate-beating-algorithm/\n\n\n> “I worked about five hours a day, seven days a week for two months,” Shabazi told GeekWire via email. “Our team had a bit of a late start but once we got going, we were obsessed. I got excited every time we made a small improvement and even thought about the problem in the shower.” Meyer said he spent well over 400 hours of free time on the competition. “My poor computer was running 24/7 during both rounds,” he said.", "votes": null }, { "id": "1249921", "postDate": "03/23/2021 16:16:33", "content": "“I worked about five hours a day, seven days a week for two months,” Shabazi told GeekWire via email. “Our team had a bit of a late start but once we got going, we were obsessed. I got excited every time we made a small improvement and even thought about the problem in the shower.” Meyer said he spent well over 400 hours of free time on the competition. “My poor computer was running 24/7 during both rounds,” he said.
\n
Wow! That's really amazing! Thanks for sharing this. I literally had no idea that how much the winners put on competitions.
", "rawMarkdown": "Wow! That's really amazing! Thanks for sharing this. I literally had no idea that how much the winners put on competitions.", "votes": null }, { "id": "1255889", "postDate": "03/29/2021 10:14:43", "content": "if you want to survive in kaggle, you need to build a \"pipeline for minimum viable model\" in less than 4 hrs.
\nin the worst case, in one day.
note: this is for the \"minimum viable model\". you can improve on it later.
\nas a programmer, you should reuse your code and keep a library of it.
\nyou should develop some methodology or operations so that you can reuse your old work, etc. at first it may take some time to start your first Kaggle project, but over time you should find yourself using less time to start a Kaggle project.
\nalso, you pipeline should be bug-free (with self check) with proper recording/logging, etc.
\nas an example, this is what a kaggle plan and timetable would look like. (i came across this while goggling around)
\nhttps://github.com/rxng8/Chest-Xray-Abnormalities-Detection
", "rawMarkdown": "if you want to survive in kaggle, you need to build a \"pipeline for minimum viable model\" in less than 4 hrs.\nin the worst case, in one day.\n\nnote: this is for the \"minimum viable model\". you can improve on it later.\n\nas a programmer, you should reuse your code and keep a library of it.\n\nyou should develop some methodology or operations so that you can reuse your old work, etc. at first it may take some time to start your first Kaggle project, but over time you should find yourself using less time to start a Kaggle project.\n\nalso, you pipeline should be bug-free (with self check) with proper recording/logging, etc.\n\n---\nas an example, this is what a kaggle plan and timetable would look like. (i came across this while goggling around)\n\nhttps://github.com/rxng8/Chest-Xray-Abnormalities-Detection", "votes": null }, { "id": "1255953", "postDate": "03/29/2021 11:54:07", "content": "Thanks for your insightful comment. I do really felt the need to write re-usable code actually and during the previous week I tried to reach a consensus on my general template for the future competitions.
\nSadly I felt this need a little late but I'm happy that at least now I know I need such a thing!
\nI'm still struggling with logging and documenting by the way; I think I'm going to use Weights&Biases for this task but I always prefer everything local. So, I was thinking about tensorboard as well.
\nAlso thank you for the link you provided; I'll check it out.
Yes.
\nIts normal…. Some time for me it takes even longer if I am not familiar with the method.. Just make sure to document and organize your code. In future if you encounter similar project you can reuse a lot of your old code =)
Thanks for reassuring and sharing your experience. Yeah I found out that I do really need to document literally \"everything\"! Because there are many nuances and details in this task compared to say vanilla image classification!
", "votes": null, "replies": [] } ] }, { "id": 1248379, "author_name": "sapr3s", "author_url": "", "post_date": "03/22/2021 14:58:21", "content": "Such competitions usually go on for several months. And on some of them I didn't have time to create a model that would suit me and didn't send anything. So I'm really jealous that you created a working model in just one day.
", "votes": null, "replies": [ { "id": 1248417, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/22/2021 15:32:18", "content": "Yeah it seems to be a really long and complicated competition in a sense that there are a lot of things to try out.
\nActually more than 80 percent of the pipeline I prepared was others' effort translated into my coding habits :) so I think I did not make a break through to do it in one day and I'm sure you can do what I did in the same or even less time :)
It's important to have a toolkit and standard working pipeline that you can adapt on different competitions.
\nYou will make it flexible and robust through competitions and you'll gain lot of time.
\nDon't worry if it takes long time to build.
", "votes": null, "replies": [ { "id": 1248443, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/22/2021 15:52:09", "content": "Thanks for pointing it out. So I'll look at it as an investment for similar competitions or personal projects! I think this mindset will help me out.
", "votes": null, "replies": [] }, { "id": 1248455, "author_name": "serigne", "author_url": "", "post_date": "03/22/2021 15:59:44", "content": "You may find inspiration in this great repo by @analokamus
\nhttps://github.com/analokmaus/kuma_utils
", "votes": null, "replies": [] }, { "id": 1248465, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/22/2021 16:04:30", "content": "Thanks! I was looking for such a thing.
", "votes": null, "replies": [] } ] }, { "id": 1248661, "author_name": "eladwar", "author_url": "", "post_date": "03/22/2021 18:29:24", "content": "I usually do a minimal viable product sort of thing where I am at least able to train something, not an actual pipeline. This can take anywhere from 30m -2h, but once I move on to a pipeline the mvp makes it really easy to set up. The rest is just dependent on how big of a dataset I have to manage. Building the pipeline can take anywhere from 30m - 4h. If I reuse old code it can be considerably faster depending on the overlap.
", "votes": null, "replies": [ { "id": 1248726, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/22/2021 19:36:35", "content": "Thanks for sharing your way of doing it. Yeah, I need to divide what I do into smaller steps. In this way I can build a much more re-usable code base.
", "votes": null, "replies": [] } ] }, { "id": 1249790, "author_name": "marketneutral", "author_url": "", "post_date": "03/23/2021 14:16:10", "content": "It takes hundreds of hours to do well in a competition.
\nOne article on the Zillow prize winners
\n\n\n", "votes": null, "replies": [ { "id": 1249921, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/23/2021 16:16:33", "content": "“I worked about five hours a day, seven days a week for two months,” Shabazi told GeekWire via email. “Our team had a bit of a late start but once we got going, we were obsessed. I got excited every time we made a small improvement and even thought about the problem in the shower.” Meyer said he spent well over 400 hours of free time on the competition. “My poor computer was running 24/7 during both rounds,” he said.
\n
Wow! That's really amazing! Thanks for sharing this. I literally had no idea that how much the winners put on competitions.
", "votes": null, "replies": [] } ] }, { "id": 1255889, "author_name": "hengck23", "author_url": "", "post_date": "03/29/2021 10:14:43", "content": "if you want to survive in kaggle, you need to build a \"pipeline for minimum viable model\" in less than 4 hrs.
\nin the worst case, in one day.
note: this is for the \"minimum viable model\". you can improve on it later.
\nas a programmer, you should reuse your code and keep a library of it.
\nyou should develop some methodology or operations so that you can reuse your old work, etc. at first it may take some time to start your first Kaggle project, but over time you should find yourself using less time to start a Kaggle project.
\nalso, you pipeline should be bug-free (with self check) with proper recording/logging, etc.
\nas an example, this is what a kaggle plan and timetable would look like. (i came across this while goggling around)
\nhttps://github.com/rxng8/Chest-Xray-Abnormalities-Detection
", "votes": null, "replies": [ { "id": 1255953, "author_name": "moeinshariatnia", "author_url": "", "post_date": "03/29/2021 11:54:07", "content": "Thanks for your insightful comment. I do really felt the need to write re-usable code actually and during the previous week I tried to reach a consensus on my general template for the future competitions.
\nSadly I felt this need a little late but I'm happy that at least now I know I need such a thing!
\nI'm still struggling with logging and documenting by the way; I think I'm going to use Weights&Biases for this task but I always prefer everything local. So, I was thinking about tensorboard as well.
\nAlso thank you for the link you provided; I'll check it out.