Basically it is divided into two parts: In the first part I created a class to define the model graph of AlexNet together with a function to load the pretrained weights and in the second part how to actually use this class to finetune AlexNet on a new dataset. Well and I think the main reason for this article is that working on a project like this, helps me to better understand TensorFlow in general.Īfter finishing to write this article I ended up having written another very long post.After some time with Keras, I recently switched to pure TensorFlow and now I want to be able to finetune the same network as previously, but using just TensorFlow. prototxt, the model structure with blobs…) to work with Caffe. But to be honest, I found it quite cumbersome (e.g. In the past, I used mainly Caffe to finetune convolutional networks.As the models get deeper they naturally need more computational time, which in some projects I can’t afford. Another reason is that for a lot of my personal projects AlexNet works quite well and there is no reason to switch to any of the more heavy-weight models to gain probably another.For the AlexNet model, we have to do a bit more on our own. Although the idea behind finetuning is the same, the major difference is, that Tensorflow (as well as Keras) already ship with VGG or Inception classes and include the weights (pretrained on ImageNet).
Albeit there exist many How-To’s, most of the newer once are covering finetuning VGG or Inception Models and not AlexNet.Well here are some reasons why I thought it’s worth doing anyway:
“AlexNet?” you might say, “So 2012’ish!” you might say. This time about finetuning AlexNet in pure TensorFlow 1.0.
#Finetune bandcamp code#
The links below in this article are still pointing to the code explained here in this article.Īfter over one year I finally found time and leisure to write my next article. Read my other blogpost for an explanation of this new feature coming with TensorFlows version >= 1.12rc0. I updated the code of the repository to work with TensorFlows new input pipeline.
0 kommentar(er)
