If you embed the 'PCA/classifier' combo, i can understand how you transfer the trained data.
\n\nBut suppose you have that pytorch trained kernel... \n1° How do you transfer that kernel to a tablet app... fe\n2° Will it be fast enough, since we are training here 'hours', is the trained kernel something that will go fast enough on a tablet ? Actually as fast as you are writing ?\n3° How will it additionally/incrementally learn from the user ? And will that be fast enough ?
", "messages": [ { "id": "677678", "postDate": "11/20/2019 13:33:26", "content": "If you embed the 'PCA/classifier' combo, i can understand how you transfer the trained data.
\n\nBut suppose you have that pytorch trained kernel... \n1° How do you transfer that kernel to a tablet app... fe\n2° Will it be fast enough, since we are training here 'hours', is the trained kernel something that will go fast enough on a tablet ? Actually as fast as you are writing ?\n3° How will it additionally/incrementally learn from the user ? And will that be fast enough ?
", "rawMarkdown": "If you embed the 'PCA/classifier' combo, i can understand how you transfer the trained data.\n\nBut suppose you have that pytorch trained kernel... \n1° How do you transfer that kernel to a tablet app... fe\n2° Will it be fast enough, since we are training here 'hours', is the trained kernel something that will go fast enough on a tablet ? Actually as fast as you are writing ?\n3° How will it additionally/incrementally learn from the user ? And will that be fast enough ?", "votes": null }, { "id": "680055", "postDate": "11/23/2019 23:30:01", "content": "Trained model is just a bunch of weights/parameters. It's trained so it can give accurate predictions to new data (e.g. provided by an user in applications). So basically you don't need to train anything when you embed it to a real life application. You only need to be able to produce new predictions (which is obviously way faster than training).
\n\nThat model will not learn anything else from the user. It can only use its knowledge (from training) to process incoming data. Moreover, it will only perform well on data that is similar to your training set i.e. if you train a face recognition model with a set of constrained indoor face images, it may not peform well with \"in the wild\" images and poor phone cameras.
", "rawMarkdown": "Trained model is just a bunch of weights/parameters. It's trained so it can give accurate predictions to new data (e.g. provided by an user in applications). So basically you don't need to train anything when you embed it to a real life application. You only need to be able to produce new predictions (which is obviously way faster than training).\n\nThat model will not learn anything else from the user. It can only use its knowledge (from training) to process incoming data. Moreover, it will only perform well on data that is similar to your training set i.e. if you train a face recognition model with a set of constrained indoor face images, it may not peform well with \"in the wild\" images and poor phone cameras.", "votes": null } ], "comments": [ { "id": 680055, "author_name": "rincewind007", "author_url": "", "post_date": "11/23/2019 23:30:01", "content": "Trained model is just a bunch of weights/parameters. It's trained so it can give accurate predictions to new data (e.g. provided by an user in applications). So basically you don't need to train anything when you embed it to a real life application. You only need to be able to produce new predictions (which is obviously way faster than training).
\n\nThat model will not learn anything else from the user. It can only use its knowledge (from training) to process incoming data. Moreover, it will only perform well on data that is similar to your training set i.e. if you train a face recognition model with a set of constrained indoor face images, it may not peform well with \"in the wild\" images and poor phone cameras.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "677678": "If you embed the 'PCA/classifier' combo, i can understand how you transfer the trained data.\n\nBut suppose you have that pytorch trained kernel... \n1° How do you transfer that kernel to a tablet app... fe\n2° Will it be fast enough, since we are training here 'hours', is the trained kernel something that will go fast enough on a tablet ? Actually as fast as you are writing ?\n3° How will it additionally/incrementally learn from the user ? And will that be fast enough ?", "680055": "1. Load the parameters into a model.\n2. You don't trained the model om a tablet you only do interference.\n3. No, New data doesn't affect the modell.", "684062": "Trained model is just a bunch of weights/parameters. It's trained so it can give accurate predictions to new data (e.g. provided by an user in applications). So basically you don't need to train anything when you embed it to a real life application. You only need to be able to produce new predictions (which is obviously way faster than training).\n\nThat model will not learn anything else from the user. It can only use its knowledge (from training) to process incoming data. Moreover, it will only perform well on data that is similar to your training set i.e. if you train a face recognition model with a set of constrained indoor face images, it may not peform well with \"in the wild\" images and poor phone cameras." }, "source": "meta" }