[Week5 — Eat & Count]
In our last blog, we talked about how to predict a given food photo’s kind and what was our convolutional neural network architecture. We implemented an AlexNet architecture and tried to train it from the beginning and we get some poor results.
This week first, we tried to implement AlexNet with a library which is TFLearn.[1] TFLearn is a deep learning library built on top of Tensorflow. This time we added the AlexNet’s layers from TFLearn’s own functions. Then we started to train our network.
We tried this network on 10 different label classes and each class has 750 train data, 250 validation data. After about 15 hours, this model gives us about %50 accuracy on this classes.If we did not stop it, the accuracy would have been increased but it takes lots of time. Then we increased the batch size and we can train more images at the same time. It decreases the time but still it spends lots of time. After a research, we found pre-trained models and if we use them in our model, It can train the model faster.
[2] Pre-trained models use the ready to use weights and by this way, we do not need to train our weights from the beginning. It brings the weights of the convolutional layers. We need to train the last convolution layers and and fully-connected layers, this allows us to train our model much faster.
On our next blogs, we will be talking about how we implement those pre-trained models and the results of our experiments and how we extract our second data set which is about the foods calorie values.
References
