1). When we talk about EfficientNet , which you take into account to choose the model, for example I chose the EfficientNetB7 because I understand that it has higher resolution and more parameters.
\n2). If I have more parameters, does my model improve or is it indifferent?
\n3). How do you deal with the problem of image balancing, according to what I have read in several notebooks there is a lot of noise in the images.
\nThanks ..
", "messages": [ { "id": 1106306, "postDate": "2020-12-08T17:39:24.307Z", "content": "1). When we talk about EfficientNet , which you take into account to choose the model, for example I chose the EfficientNetB7 because I understand that it has higher resolution and more parameters.
\n2). If I have more parameters, does my model improve or is it indifferent?
\n3). How do you deal with the problem of image balancing, according to what I have read in several notebooks there is a lot of noise in the images.
\nThanks ..
", "rawMarkdown": "1). When we talk about EfficientNet , which you take into account to choose the model, for example I chose the EfficientNetB7 because I understand that it has higher resolution and more parameters.\n\n2). If I have more parameters, does my model improve or is it indifferent?\n\n3). How do you deal with the problem of image balancing, according to what I have read in several notebooks there is a lot of noise in the images. \n\nThanks ..", "votes": 3 }, { "id": 1106445, "postDate": "2020-12-08T21:00:44.463Z", "content": "As you said, you choose your EfficientNet based on input image resolution and complexity of the network (you can understand more about this network here: https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/). But it's important to remind that image resolution and the number of parameters doesn't mean better results. When you increase the resolution you add more spatial features from the image into the model, turning your feature space more dense. This can be important to add separability to the decision boundaries, but can also turn this boundary more complex to be fitted, this may vary a lot based on the problem. The rule of thumb is that a complex model tends to generate better results, but also is more complicated to be trained and needs more computational power. I always prefer to have a simpler model that achieves good results.
\nAbout imbalance, I didn't work on that yet, but you can check several notebooks where people address this kind of problem, you can for example subsample your data or create an ensemble classifier.
", "rawMarkdown": "As you said, you choose your EfficientNet based on input image resolution and complexity of the network (you can understand more about this network here: https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/). But it's important to remind that image resolution and the number of parameters doesn't mean better results. When you increase the resolution you add more spatial features from the image into the model, turning your feature space more dense. This can be important to add separability to the decision boundaries, but can also turn this boundary more complex to be fitted, this may vary a lot based on the problem. The rule of thumb is that a complex model tends to generate better results, but also is more complicated to be trained and needs more computational power. I always prefer to have a simpler model that achieves good results. \n\nAbout imbalance, I didn't work on that yet, but you can check several notebooks where people address this kind of problem, you can for example subsample your data or create an ensemble classifier.", "votes": 2, "replies": [ { "id": 1106469, "postDate": "2020-12-08T21:30:40.890Z", "content": "thank you for providing information. Greetings
", "rawMarkdown": "thank you for providing information. Greetings" } ] }, { "id": 1106493, "postDate": "2020-12-08T21:54:06.187Z", "rawMarkdown": "", "votes": 2, "isDeleted": true, "replies": [ { "id": 1106537, "postDate": "2020-12-08T22:55:11.373Z", "content": "Thank you for your contribution, in question 3 I refer to the balance of the data, what kind of approach to improve accuracy, since one of the classes is majority, for example modify the original dataset by removing the majority class or insert images from the previous dataset.
", "rawMarkdown": "Thank you for your contribution, in question 3 I refer to the balance of the data, what kind of approach to improve accuracy, since one of the classes is majority, for example modify the original dataset by removing the majority class or insert images from the previous dataset. " } ] } ], "comments": [ { "id": 1106445, "author_name": "Alvaro Leandro Cavalcante Carneiro", "author_url": "", "post_date": "2020-12-08T21:00:44.463000", "content": "As you said, you choose your EfficientNet based on input image resolution and complexity of the network (you can understand more about this network here: https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/). But it's important to remind that image resolution and the number of parameters doesn't mean better results. When you increase the resolution you add more spatial features from the image into the model, turning your feature space more dense. This can be important to add separability to the decision boundaries, but can also turn this boundary more complex to be fitted, this may vary a lot based on the problem. The rule of thumb is that a complex model tends to generate better results, but also is more complicated to be trained and needs more computational power. I always prefer to have a simpler model that achieves good results.
\nAbout imbalance, I didn't work on that yet, but you can check several notebooks where people address this kind of problem, you can for example subsample your data or create an ensemble classifier.
", "votes": 2, "replies": [ { "id": 1106469, "author_name": "Carlos Alberto Gómez Prado", "author_url": "", "post_date": "2020-12-08T21:30:40.890000", "content": "thank you for providing information. Greetings
", "votes": 0, "replies": [] } ] }, { "id": 1106493, "author_name": "", "author_url": "", "post_date": "2020-12-08T21:54:06.187000", "content": "", "votes": 2, "replies": [ { "id": 1106537, "author_name": "Carlos Alberto Gómez Prado", "author_url": "", "post_date": "2020-12-08T22:55:11.373000", "content": "Thank you for your contribution, in question 3 I refer to the balance of the data, what kind of approach to improve accuracy, since one of the classes is majority, for example modify the original dataset by removing the majority class or insert images from the previous dataset.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "1106306": "1). When we talk about EfficientNet , which you take into account to choose the model, for example I chose the EfficientNetB7 because I understand that it has higher resolution and more parameters.\n\n2). If I have more parameters, does my model improve or is it indifferent?\n\n3). How do you deal with the problem of image balancing, according to what I have read in several notebooks there is a lot of noise in the images. \n\nThanks ..", "1106445": "As you said, you choose your EfficientNet based on input image resolution and complexity of the network (you can understand more about this network here: https://keras.io/examples/vision/image_classification_efficientnet_fine_tuning/). But it's important to remind that image resolution and the number of parameters doesn't mean better results. When you increase the resolution you add more spatial features from the image into the model, turning your feature space more dense. This can be important to add separability to the decision boundaries, but can also turn this boundary more complex to be fitted, this may vary a lot based on the problem. The rule of thumb is that a complex model tends to generate better results, but also is more complicated to be trained and needs more computational power. I always prefer to have a simpler model that achieves good results. \n\nAbout imbalance, I didn't work on that yet, but you can check several notebooks where people address this kind of problem, you can for example subsample your data or create an ensemble classifier.", "1106493": "" } }