Before starting to experiments with new architectures and training from scratch, I'd like to get a decent score using pre-trained models (e.g., ResNet34 and DenseNet). Unfortunately, my attempts so far have all failed.
\n\nTwo things I'm not sure I'm doing right, are normalization and the way I pass the images to the network. Regarding the latter, I simply define a new first convolutional layer that takes the 6 channels as input. I'm training it from scratch.\nRegarding the normalization, I simply use the statistics from RxRx. The problem is that models like ResNet were trained with RGB images, so although it's technically possible to do this, I do not know whether it's correct.
\n\nI tried to train ResNet34 both from scratch and using a pre-trained model. Sadly, the performances of both are comparable and I obtain very low accuracies. Something I did not expect, also considering that others are using the same model and are obtaining much higher accuracies!
\n\nOne more thing I'm worried about, is batch size. My machine (GTX 1060), with ResNet34, 6 channels and 128x128 images cannot use a batch size larger than 16. Is it too small?
\n\nDoes anybody have any suggestion regarding both input normalization and how to handle 6 channels rather than RGB images?
", "messages": [ { "id": "580765", "postDate": "07/20/2019 17:53:52", "content": "Before starting to experiments with new architectures and training from scratch, I'd like to get a decent score using pre-trained models (e.g., ResNet34 and DenseNet). Unfortunately, my attempts so far have all failed.
\n\nTwo things I'm not sure I'm doing right, are normalization and the way I pass the images to the network. Regarding the latter, I simply define a new first convolutional layer that takes the 6 channels as input. I'm training it from scratch.\nRegarding the normalization, I simply use the statistics from RxRx. The problem is that models like ResNet were trained with RGB images, so although it's technically possible to do this, I do not know whether it's correct.
\n\nI tried to train ResNet34 both from scratch and using a pre-trained model. Sadly, the performances of both are comparable and I obtain very low accuracies. Something I did not expect, also considering that others are using the same model and are obtaining much higher accuracies!
\n\nOne more thing I'm worried about, is batch size. My machine (GTX 1060), with ResNet34, 6 channels and 128x128 images cannot use a batch size larger than 16. Is it too small?
\n\nDoes anybody have any suggestion regarding both input normalization and how to handle 6 channels rather than RGB images?
", "rawMarkdown": "Before starting to experiments with new architectures and training from scratch, I'd like to get a decent score using pre-trained models (e.g., ResNet34 and DenseNet). Unfortunately, my attempts so far have all failed.\n\nTwo things I'm not sure I'm doing right, are normalization and the way I pass the images to the network. Regarding the latter, I simply define a new first convolutional layer that takes the 6 channels as input. I'm training it from scratch.\nRegarding the normalization, I simply use the statistics from RxRx. The problem is that models like ResNet were trained with RGB images, so although it's technically possible to do this, I do not know whether it's correct.\n\nI tried to train ResNet34 both from scratch and using a pre-trained model. Sadly, the performances of both are comparable and I obtain very low accuracies. Something I did not expect, also considering that others are using the same model and are obtaining much higher accuracies!\n\nOne more thing I'm worried about, is batch size. My machine (GTX 1060), with ResNet34, 6 channels and 128x128 images cannot use a batch size larger than 16. Is it too small?\n\nDoes anybody have any suggestion regarding both input normalization and how to handle 6 channels rather than RGB images?", "votes": null }, { "id": "590833", "postDate": "08/02/2019 17:10:48", "content": "Sorry, how have you normalized your images? Can you explaine please? How should we use pixel_stats.csv? Thanks.
", "rawMarkdown": "Sorry, how have you normalized your images? Can you explaine please? How should we use pixel_stats.csv? Thanks.", "votes": null }, { "id": "590852", "postDate": "08/02/2019 17:54:16", "content": "@joven1997 I simply took the means and standard deviations from here and then I used Normalize from PyTorch. I did not use pixel_stats.csv.
Thanks!
", "rawMarkdown": "Thanks!", "votes": null }, { "id": "593448", "postDate": "08/06/2019 16:08:04", "content": "@lorenzofabbri92 did you normalize before applying ToTensor or after it?
", "rawMarkdown": "lorenzofabbri92 did you normalize before applying ToTensor or after it?", "votes": null }, { "id": "593452", "postDate": "08/06/2019 16:10:13", "content": "I'm not using ToTensor since the output of my Dataset is a tensor already.
And in what range your pixel values are on the output of your Dataset ? [0,255] or [0,1] ?
@rafailfridman Well, in neither of those... I guess it depends on the input to ToTensor.
Sorry, how have you normalized your images? Can you explaine please? How should we use pixel_stats.csv? Thanks.
", "votes": null, "replies": [ { "id": 590852, "author_name": "lorenzofabbri92", "author_url": "", "post_date": "08/02/2019 17:54:16", "content": "@joven1997 I simply took the means and standard deviations from here and then I used Normalize from PyTorch. I did not use pixel_stats.csv.
Thanks!
", "votes": null, "replies": [] } ] }, { "id": 593448, "author_name": "rafailfridman", "author_url": "", "post_date": "08/06/2019 16:08:04", "content": "@lorenzofabbri92 did you normalize before applying ToTensor or after it?
", "votes": null, "replies": [ { "id": 593452, "author_name": "lorenzofabbri92", "author_url": "", "post_date": "08/06/2019 16:10:13", "content": "I'm not using ToTensor since the output of my Dataset is a tensor already.
And in what range your pixel values are on the output of your Dataset ? [0,255] or [0,1] ?
@rafailfridman Well, in neither of those... I guess it depends on the input to ToTensor.