[Week3 — Eat&Count]

In this week, we searched some deep learning frameworks for our project. We needed a framework that satisfies our requirements such as easy to find sources about the framework, flexible, good at performance, suitable for work at Python etc. We found some frameworks, Caffe, Tensorflow, Theano and Torch [1] . Each framework has advantages and disadvantages.

First, we considered to not use Torch Framework because it is not suitable for Python. When we consider performance last three frameworks all uses the gpu for evaluation, then all of them were good. In Caffe Framework, language modelling were low level, but on the other hand, in Tensorflow and Theano Frameworks, we can abstract our architecture with using symbolic graphs. So that they are easier to use for us compare to Caffe. At last, when we compare these two frameworks, Theano is hard to navigate, debug and refactor because of whole code packaged as python string. But in Tensorflow, the code converted into C++ code and it is more cleaner modular architecture. In the direction of these results, we decided to use Tensorflow Framework and we started to our experiments with this framework.

In Tensorflow architecture, it represents computations as graphs. Every layers are made up of computations. At the beginning, we create a computation graph and run it using sessions.

As a beginning, we followed its own tutorials [2] and we used MNIST dataset to classify handwritten digits. We tried a network with 2 CONV layers and got about %99.02 accuracy with it. Tensorflow allows us to use high-leveled built in function and they made it easy us to handle back propagation and forward pass.

On our next blog, we will get some experiments on our dataset which is Food-101 and we will be talking about the results of these experiments.

References

[1] Evaluation of Deep Learning Toolkits- https://github.com/zer0n/deepframeworks/blob/master/README.md

[2] Tensorflow official website - https://www.tensorflow.org