Good Morning,
I would need to recreate the model showed below.
.
In the paper [here] (https://arxiv.org/pdf/1412.3128.pdf ) it is written that "the convolutional layers are interspersed with normalization and maxpooling layers at various stages and it pretrained using ImageNet.
I would like to implement this model or alternatively I would like to know how to adapt the pretrained version of Alexnet (since the model is derived from Alexnet). The model is used to perform regression on bounding boxes. Do you have suggestions?
To create a custom model, you could have a look at the CIFAR10 tutorial, which defines a custom model architecture.
On the other hand, you could also use the pretrained torchvision.models.alexnet and change the last output layer for your regression task. Do replace the last linear layer, you could use this code snippet:
model = models.alexnet()
model.classifier[6] = nn.Linear(4096, nb_output)
where nb_output defines the number of regression outputs you are dealing with.
The Finetuning tutorial gives you more information about how to further fine tune the model.
Note that the tutorial is dealing with a classification task, so you would have to at least change the criterion to e.g. nn.MSELoss.
