Hi all.I have knowledge in Data Science,Machine learning.Need some mentor guidance wrt beginning journey in Computer vision.Any help is appreciated on how to start solving the MNIST dataset as a beginner
", "messages": [ { "id": 1139275, "postDate": "2021-01-05T09:34:41.783Z", "content": "Here are some useful resources for this competition:
\nHope it helps!
", "rawMarkdown": "Here are some useful resources for this competition:\n\n- [https://www.kaggle.com/c/digit-recognizer/discussion/61480](https://www.kaggle.com/c/digit-recognizer/discussion/61480) : How to score 97%, 98%, 99%, and 100%\n- [https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6](https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6) : Introduction to CNN Keras - 0.997 (top 6%)\n- [https://www.kaggle.com/arthurtok/interactive-intro-to-dimensionality-reduction](https://www.kaggle.com/arthurtok/interactive-intro-to-dimensionality-reduction) : Interactive Intro to Dimensionality Reduction\n\nHope it helps!", "votes": 3, "replies": [ { "id": 1140212, "postDate": "2021-01-05T21:19:49.790Z", "content": "@fedesoriano those are definitely good resources! Thanks for posting them.
", "rawMarkdown": "@fedesoriano those are definitely good resources! Thanks for posting them." }, { "id": 1140219, "postDate": "2021-01-05T21:31:52.687Z", "content": "You're welcome!! ;)
", "rawMarkdown": "You're welcome!! ;)" } ] }, { "id": 1138412, "postDate": "2021-01-04T17:11:37.380Z", "content": "Hi all.I have knowledge in Data Science,Machine learning.Need some mentor guidance wrt beginning journey in Computer vision.Any help is appreciated on how to start solving the MNIST dataset as a beginner
", "rawMarkdown": "Hi all.I have knowledge in Data Science,Machine learning.Need some mentor guidance wrt beginning journey in Computer vision.Any help is appreciated on how to start solving the MNIST dataset as a beginner\n", "votes": 2 }, { "id": 1138655, "postDate": "2021-01-04T20:59:36.883Z", "content": "A great way of approaching a solution is to take a look at the Notebooks section of the competition. Change the ordering of the notebooks from Hotness to Best Score and start examining some of the things that other people have done to solve the problem (my advice is to ignore any notebooks that have a performance of 1.0 as they are likely to have just found an open version of the dataset and used that to submit a perfect score). When you find a notebook that you like, you can use the Copy and Edit button to copy it to your workspace to make edits on it.
\nFor more specific advice:
\nlabel column that tells you what the digit is actually supposed to be. The remaining columns comprise the actual pixel data for the image (flattened from a 28 x 28 array into a single 784 element array). The values of the pixels range from 0 (black) to 255 (white). train_test_split function from SciKit Learn (or similar), you can generate a validation set to see how well your approach works. From there, you should look at what your classifier got wrong and why it got it wrong. That will help you decide what to do next (for example, maybe try augmenting the data, transform it, or clean it further if need be).Note that these steps aren't necessarily sequential. You'll often find yourself moving back and forth between various steps as you learn more about the dataset. For this particular challenge, I've created two notebooks that may help you get started:
\n\nThey go over loading the data, building a classifier, testing the results, and steps to try next. As I said above, looking at other notebooks is also very helpful. Good luck and have fun learning!
", "rawMarkdown": "A great way of approaching a solution is to take a look at the [Notebooks](https://www.kaggle.com/c/digit-recognizer/notebooks) section of the competition. Change the ordering of the notebooks from _Hotness_ to _Best Score_ and start examining some of the things that other people have done to solve the problem (my advice is to ignore any notebooks that have a performance of 1.0 as they are likely to have just found an open version of the dataset and used that to submit a perfect score). When you find a notebook that you like, you can use the _Copy and Edit_ button to copy it to your workspace to make edits on it.\n\nFor more specific advice:\n\n1. Solving the MNIST problem involves first understanding how the data is organized in the datasets provided. In the training dataset, there is a `label` column that tells you what the digit is actually supposed to be. The remaining columns comprise the actual pixel data for the image (flattened from a 28 x 28 array into a single 784 element array). The values of the pixels range from 0 (black) to 255 (white). \n2. At this point, you will probably want to dig a little deeper into the data we've been provided for training. Things you may want to look at include how many examples of each digit we have to look at and what those digits actually look like. For example, are the digits all oriented the same way? Are they located in the same spot in each image? Do they have the same thickness?\n3. You will then need to choose a classifier or approach to take for the problem. There are many different solutions that you can try that involve simple image classifiers such as K nearest neighbor and naive bayes, to more complicated ones such as convoluted neural networks (CNNs). I recommend starting simple to establish a baseline of performance that you can compare to.\n4. You should then test out your classifier approach (as you probably already know given that you said you are already knowledgeable in Machine Learning and Data Science). If you take your training data and split it up using a `train_test_split` function from SciKit Learn (or similar), you can generate a validation set to see how well your approach works. From there, you should look at what your classifier got wrong and why it got it wrong. That will help you decide what to do next (for example, maybe try augmenting the data, transform it, or clean it further if need be).\n5. Once you have a good classifier strategy working, you can try and fine tune it to get better performance.\n\nNote that these steps aren't necessarily sequential. You'll often find yourself moving back and forth between various steps as you learn more about the dataset. For this particular challenge, I've created two notebooks that may help you get started:\n\n* [Beginner Guide to Digit Recognition](https://www.kaggle.com/craigmthomas/beginner-guide-to-digit-recognition)\n* [Beginner Guide to Digit Recognition (Part 2)](https://www.kaggle.com/craigmthomas/beginner-guide-to-digit-recognition-part-2)\n\nThey go over loading the data, building a classifier, testing the results, and steps to try next. As I said above, looking at other notebooks is also very helpful. Good luck and have fun learning!" }, { "id": 1138946, "postDate": "2021-01-05T04:34:11.413Z", "content": "Hi Craig thanks for comment .It was really helpful
", "rawMarkdown": "Hi Craig thanks for comment .It was really helpful" } ], "comments": [ { "id": 1139275, "author_name": "fedesoriano", "author_url": "", "post_date": "2021-01-05T09:34:41.783000", "content": "Here are some useful resources for this competition:
\nHope it helps!
", "votes": 3, "replies": [ { "id": 1140212, "author_name": "Craig Thomas", "author_url": "", "post_date": "2021-01-05T21:19:49.790000", "content": "@fedesoriano those are definitely good resources! Thanks for posting them.
", "votes": 0, "replies": [] }, { "id": 1140219, "author_name": "fedesoriano", "author_url": "", "post_date": "2021-01-05T21:31:52.687000", "content": "You're welcome!! ;)
", "votes": 0, "replies": [] } ] }, { "id": 1138655, "author_name": "Craig Thomas", "author_url": "", "post_date": "2021-01-04T20:59:36.883000", "content": "A great way of approaching a solution is to take a look at the Notebooks section of the competition. Change the ordering of the notebooks from Hotness to Best Score and start examining some of the things that other people have done to solve the problem (my advice is to ignore any notebooks that have a performance of 1.0 as they are likely to have just found an open version of the dataset and used that to submit a perfect score). When you find a notebook that you like, you can use the Copy and Edit button to copy it to your workspace to make edits on it.
\nFor more specific advice:
\nlabel column that tells you what the digit is actually supposed to be. The remaining columns comprise the actual pixel data for the image (flattened from a 28 x 28 array into a single 784 element array). The values of the pixels range from 0 (black) to 255 (white). train_test_split function from SciKit Learn (or similar), you can generate a validation set to see how well your approach works. From there, you should look at what your classifier got wrong and why it got it wrong. That will help you decide what to do next (for example, maybe try augmenting the data, transform it, or clean it further if need be).Note that these steps aren't necessarily sequential. You'll often find yourself moving back and forth between various steps as you learn more about the dataset. For this particular challenge, I've created two notebooks that may help you get started:
\n\nThey go over loading the data, building a classifier, testing the results, and steps to try next. As I said above, looking at other notebooks is also very helpful. Good luck and have fun learning!
", "votes": 0, "replies": [] }, { "id": 1138946, "author_name": "Aviraj Roy", "author_url": "", "post_date": "2021-01-05T04:34:11.413000", "content": "Hi Craig thanks for comment .It was really helpful
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "1139275": "Here are some useful resources for this competition:\n\n- [https://www.kaggle.com/c/digit-recognizer/discussion/61480](https://www.kaggle.com/c/digit-recognizer/discussion/61480) : How to score 97%, 98%, 99%, and 100%\n- [https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6](https://www.kaggle.com/yassineghouzam/introduction-to-cnn-keras-0-997-top-6) : Introduction to CNN Keras - 0.997 (top 6%)\n- [https://www.kaggle.com/arthurtok/interactive-intro-to-dimensionality-reduction](https://www.kaggle.com/arthurtok/interactive-intro-to-dimensionality-reduction) : Interactive Intro to Dimensionality Reduction\n\nHope it helps!", "1138412": "Hi all.I have knowledge in Data Science,Machine learning.Need some mentor guidance wrt beginning journey in Computer vision.Any help is appreciated on how to start solving the MNIST dataset as a beginner\n", "1138655": "A great way of approaching a solution is to take a look at the [Notebooks](https://www.kaggle.com/c/digit-recognizer/notebooks) section of the competition. Change the ordering of the notebooks from _Hotness_ to _Best Score_ and start examining some of the things that other people have done to solve the problem (my advice is to ignore any notebooks that have a performance of 1.0 as they are likely to have just found an open version of the dataset and used that to submit a perfect score). When you find a notebook that you like, you can use the _Copy and Edit_ button to copy it to your workspace to make edits on it.\n\nFor more specific advice:\n\n1. Solving the MNIST problem involves first understanding how the data is organized in the datasets provided. In the training dataset, there is a `label` column that tells you what the digit is actually supposed to be. The remaining columns comprise the actual pixel data for the image (flattened from a 28 x 28 array into a single 784 element array). The values of the pixels range from 0 (black) to 255 (white). \n2. At this point, you will probably want to dig a little deeper into the data we've been provided for training. Things you may want to look at include how many examples of each digit we have to look at and what those digits actually look like. For example, are the digits all oriented the same way? Are they located in the same spot in each image? Do they have the same thickness?\n3. You will then need to choose a classifier or approach to take for the problem. There are many different solutions that you can try that involve simple image classifiers such as K nearest neighbor and naive bayes, to more complicated ones such as convoluted neural networks (CNNs). I recommend starting simple to establish a baseline of performance that you can compare to.\n4. You should then test out your classifier approach (as you probably already know given that you said you are already knowledgeable in Machine Learning and Data Science). If you take your training data and split it up using a `train_test_split` function from SciKit Learn (or similar), you can generate a validation set to see how well your approach works. From there, you should look at what your classifier got wrong and why it got it wrong. That will help you decide what to do next (for example, maybe try augmenting the data, transform it, or clean it further if need be).\n5. Once you have a good classifier strategy working, you can try and fine tune it to get better performance.\n\nNote that these steps aren't necessarily sequential. You'll often find yourself moving back and forth between various steps as you learn more about the dataset. For this particular challenge, I've created two notebooks that may help you get started:\n\n* [Beginner Guide to Digit Recognition](https://www.kaggle.com/craigmthomas/beginner-guide-to-digit-recognition)\n* [Beginner Guide to Digit Recognition (Part 2)](https://www.kaggle.com/craigmthomas/beginner-guide-to-digit-recognition-part-2)\n\nThey go over loading the data, building a classifier, testing the results, and steps to try next. As I said above, looking at other notebooks is also very helpful. Good luck and have fun learning!", "1138946": "Hi Craig thanks for comment .It was really helpful" } }