![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3924325%2F9b620b9faafb03201f3b8bae8a624dc2%2FSkrmklipp.jpg?generation=1599118018387236&alt=media\")
At the risk of being a stupid question:
\nDue to the fact that a spectrogram is not in color/rgb, shouldn't one also change the number of channels to 1, from default 3, before training a model for better performance?
\nOr can we use the 3 channels, e.g. add 2 more variants of the spectrogram to the input.
\nMono_to_color code seen in public notebooks, seems to be one way to solve it, the best way?
At the risk of being a stupid question:
\nDue to the fact that a spectrogram is not in color/rgb, shouldn't one also change the number of channels to 1, from default 3, before training a model for better performance?
\nOr can we use the 3 channels, e.g. add 2 more variants of the spectrogram to the input.
\nMono_to_color code seen in public notebooks, seems to be one way to solve it, the best way?
I think the real reason why people use 3 channels instead of 1 is the \"Image-net\" or \"noisy-student\" pretrained weights are trained on RGB, and it's common belief that using pretrained weights is better than training from scratch.
", "rawMarkdown": "I think the real reason why people use 3 channels instead of 1 is the \"Image-net\" or \"noisy-student\" pretrained weights are trained on RGB, and it's common belief that using pretrained weights is better than training from scratch.", "votes": null }, { "id": "996229", "postDate": "09/03/2020 06:49:53", "content": "The other good way to make use of 3 channels is add delta and delta delta. I believe it's very common way in audio classification task.
", "rawMarkdown": "The other good way to make use of 3 channels is add delta and delta delta. I believe it's very common way in audio classification task.", "votes": null }, { "id": "996273", "postDate": "09/03/2020 07:29:56", "content": "OK thanks, will look into that!
\nAlso found an example in a research paper from a university that changed the channel from 3 to 1 to different models.
\n\"the number of channels in ResNet has been reduced to 1 (because spectrograms do not have RGB colors); as the number of input features decreased, the model complexity can also be lowered\"
\nInception-v3: \"The same weight initialization that
\nwas applied to ResNet has also been applied here. The number of channels was
\ndecreased before the data flows into Inception module 5c.\"
Inception-v3 outperformed the others, due to the number of parameters.
\nBut little confused if they only changed the resnet's parameters/weight, Inception parm size look the same but not for resnet.
\n\"In our experiment, the parameter size of ResNet-18 is
\n9.12 MB, the parameter size of ResNet-34 is 10.41 MB, and the parameter size
\nof Inception-v3 is 92.3 MB.\"
In all cases, still interesting to read some testing on the subject.
\nOther interesting reading, they also use these Spec. param.
\n
\"Bird Sound Classification using Convolutional
\nNeural Networks\"
\nhttp://ceur-ws.org/Vol-2380/paper_68.pdf
What I do in Pytorch is to input a BxHxW (B: batch size, H:Height, W: width) tensor to my model, and I create the 3 channels in the forward method without duplicating data:
\n def forward(self, x):\n x = x.unsqueeze(1)\n x = x.expand(-1, 3, -1, -1)\n ...\nI think the real reason why people use 3 channels instead of 1 is the \"Image-net\" or \"noisy-student\" pretrained weights are trained on RGB, and it's common belief that using pretrained weights is better than training from scratch.
", "votes": null, "replies": [] }, { "id": 996229, "author_name": "bamps53", "author_url": "", "post_date": "09/03/2020 06:49:53", "content": "The other good way to make use of 3 channels is add delta and delta delta. I believe it's very common way in audio classification task.
", "votes": null, "replies": [ { "id": 996273, "author_name": "kirderf", "author_url": "", "post_date": "09/03/2020 07:29:56", "content": "OK thanks, will look into that!
\nAlso found an example in a research paper from a university that changed the channel from 3 to 1 to different models.
\n\"the number of channels in ResNet has been reduced to 1 (because spectrograms do not have RGB colors); as the number of input features decreased, the model complexity can also be lowered\"
\nInception-v3: \"The same weight initialization that
\nwas applied to ResNet has also been applied here. The number of channels was
\ndecreased before the data flows into Inception module 5c.\"
Inception-v3 outperformed the others, due to the number of parameters.
\nBut little confused if they only changed the resnet's parameters/weight, Inception parm size look the same but not for resnet.
\n\"In our experiment, the parameter size of ResNet-18 is
\n9.12 MB, the parameter size of ResNet-34 is 10.41 MB, and the parameter size
\nof Inception-v3 is 92.3 MB.\"
In all cases, still interesting to read some testing on the subject.
\nOther interesting reading, they also use these Spec. param.
\n
\"Bird Sound Classification using Convolutional
\nNeural Networks\"
\nhttp://ceur-ws.org/Vol-2380/paper_68.pdf
What I do in Pytorch is to input a BxHxW (B: batch size, H:Height, W: width) tensor to my model, and I create the 3 channels in the forward method without duplicating data:
\n def forward(self, x):\n x = x.unsqueeze(1)\n x = x.expand(-1, 3, -1, -1)\n ...\n