I don't know if it's feasable but you can use a YOLO model pretrained on Street View House Numbers Dataset. So the model can identify the labels and you can sum them up. I know that handwritten digits are a bit weirder than house numbers but maybe it's worth a try. :)
", "messages": [ { "id": "1717841", "postDate": "03/10/2022 09:11:00", "content": "I don't know if it's feasable but you can use a YOLO model pretrained on Street View House Numbers Dataset. So the model can identify the labels and you can sum them up. I know that handwritten digits are a bit weirder than house numbers but maybe it's worth a try. :)
", "rawMarkdown": "I don't know if it's feasable but you can use a YOLO model pretrained on [Street View House Numbers Dataset](http://ufldl.stanford.edu/housenumbers/). So the model can identify the labels and you can sum them up. I know that handwritten digits are a bit weirder than house numbers but maybe it's worth a try. :)", "votes": null }, { "id": "1717924", "postDate": "03/10/2022 10:49:15", "content": "Dear @bnsbb1999
\nAnother simple idea could be to slightly modify the workflow outlined by @remekkinas in the notetbook \"STEP 2 - Find Numbers - NO MODEL REQUIRED\" to output the digits resized to 28, 28, then use a classifier that has been trained on the classic MNIST dataset.
One could even try using a quick and easy gradient-boosing classifier (for example see \"MNIST with no neural network\", amongst many others) to test things out before moving on to much more heavyweight models such as YOLO.
\nOnce you have a classifier that has been trained on the classic MNIST dataset it can be \"containerized\" as a kaggle public(*) dataset, perhaps using pickle. If you are not familiar with pickle an example of how to to that see the notebook \"All in a pickle: Saving the Titanic\", or you can save a keras model in H5 format.
\nAll the best,
\ncarl
(*) The competition rules state \"You can… …use pre-trained models which are publicly available to everyone.\"
", "rawMarkdown": "Dear @bnsbb1999 \n\nAnother simple idea could be to slightly modify the workflow outlined by @remekkinas in the notetbook [\"STEP 2 - Find Numbers - NO MODEL REQUIRED\"](https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required) to output the digits resized to `28, 28`, then use a classifier that has been trained on the classic MNIST dataset. \n\nOne could even try using a quick and easy gradient-boosing classifier (for example see [\"MNIST with no neural network\"](https://www.kaggle.com/carlmcbrideellis/mnist-with-no-neural-network), amongst many others) to test things out before moving on to much more heavyweight models such as YOLO.\n\nOnce you have a classifier that has been trained on the classic MNIST dataset it can be \"containerized\" as a kaggle public(*) dataset, perhaps using pickle. If you are not familiar with pickle an example of how to to that see the notebook [\"All in a pickle: Saving the Titanic\"](https://www.kaggle.com/carlmcbrideellis/all-in-a-pickle-saving-the-titanic), or you can [save a keras model in H5 format](https://www.tensorflow.org/guide/keras/save_and_serialize).\n\nAll the best,\ncarl\n\n(\\*) The competition rules state \"*You can... ...use pre-trained models which are publicly available to everyone.*\"", "votes": null } ], "comments": [ { "id": 1717924, "author_name": "carlmcbrideellis", "author_url": "", "post_date": "03/10/2022 10:49:15", "content": "Dear @bnsbb1999
\nAnother simple idea could be to slightly modify the workflow outlined by @remekkinas in the notetbook \"STEP 2 - Find Numbers - NO MODEL REQUIRED\" to output the digits resized to 28, 28, then use a classifier that has been trained on the classic MNIST dataset.
One could even try using a quick and easy gradient-boosing classifier (for example see \"MNIST with no neural network\", amongst many others) to test things out before moving on to much more heavyweight models such as YOLO.
\nOnce you have a classifier that has been trained on the classic MNIST dataset it can be \"containerized\" as a kaggle public(*) dataset, perhaps using pickle. If you are not familiar with pickle an example of how to to that see the notebook \"All in a pickle: Saving the Titanic\", or you can save a keras model in H5 format.
\nAll the best,
\ncarl
(*) The competition rules state \"You can… …use pre-trained models which are publicly available to everyone.\"
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1717841": "I don't know if it's feasable but you can use a YOLO model pretrained on [Street View House Numbers Dataset](http://ufldl.stanford.edu/housenumbers/). So the model can identify the labels and you can sum them up. I know that handwritten digits are a bit weirder than house numbers but maybe it's worth a try. :)", "1717924": "Dear @bnsbb1999 \n\nAnother simple idea could be to slightly modify the workflow outlined by @remekkinas in the notetbook [\"STEP 2 - Find Numbers - NO MODEL REQUIRED\"](https://www.kaggle.com/remekkinas/step-2-find-numbers-no-model-required) to output the digits resized to `28, 28`, then use a classifier that has been trained on the classic MNIST dataset. \n\nOne could even try using a quick and easy gradient-boosing classifier (for example see [\"MNIST with no neural network\"](https://www.kaggle.com/carlmcbrideellis/mnist-with-no-neural-network), amongst many others) to test things out before moving on to much more heavyweight models such as YOLO.\n\nOnce you have a classifier that has been trained on the classic MNIST dataset it can be \"containerized\" as a kaggle public(*) dataset, perhaps using pickle. If you are not familiar with pickle an example of how to to that see the notebook [\"All in a pickle: Saving the Titanic\"](https://www.kaggle.com/carlmcbrideellis/all-in-a-pickle-saving-the-titanic), or you can [save a keras model in H5 format](https://www.tensorflow.org/guide/keras/save_and_serialize).\n\nAll the best,\ncarl\n\n(\\*) The competition rules state \"*You can... ...use pre-trained models which are publicly available to everyone.*\"" }, "source": "meta" }