what must I change in the Alexnet model for working at images with dimensions of 28x28? the kernel size?
Hi @jsouza, LeNet (http://d2l.ai/chapter_convolutional-neural-networks/lenet.html) is designed for 28x28 size image.
Thank you so much for your feedback @goldpiggy. I will verify the LeNet model.
My question is based on the experiment (private) that applied the Alexnet model without modifications at a 28x28 size image. In this experiment, the model has been closing at 10% accuracy. I have solved it by resizing the image for 224X24 size. I also have solved it by initializing the weight and bias with normal distribution, where mean = 0.0 and standard deviation equal to 0.01. What do other modifications can be done?
I’m sorry my English, I am learning.
Hi @jsouza, great questions. There are tons of tricks/models that we talked about through the whole books. No hurries, you will learn them gradually by using more advanced architectures, more advanced optimization strategy etc.
Hello @goldpiggy, thank you again. Also, the site’s information, could you indicate me to papers that talk about the subject?
Here are all the reference papers we used through the book. https://d2l.ai/chapter_references/zreferences.html
Hello Ma’am, firstly, thank you so much to you and your team for such a great and insightful content on CNN. I am new to the subject and was stuck on the 2nd question given in the exercises for 7.1. That is, to design an AlexNet model that could work directly on 28*28 images. I would be really helped if you could provide any sort of help on this question, a hint maybe.
For exercise 4 of this section, are these standard questions in the field, or are you guys just asking general questions that could force the reader to delve deep into really picking apart AlexNet and wondering why it has so much going on in the Fully Connected layers rather than the Convolutional layer?
Hi @smizerex. If we look into the details of AlexNet, it maybe surprising to find out the 95% of its parameters are coming from the last 3 FCN. That’s why the researchers invent novel architectures such ResNet (which has much less parameters but better performance).
Thats what I was getting at, but wasnt going to outright say it, diminishing the exercise smashing! thank you @goldpiggy