Computer Vision News - February 2018

Inception-v3 is a variant of Inception-v2 with the auxiliary classifier units normalized per batch. A few other minor changes include using the RMSprop optimizer and label smoothing. Inception-v4 is a variant of Inception-v3. Inspired by the ResNet architecture, residual connection was added to the inception units. As you can see below, a residual connection was added alongside the existing Inception-v3 units, passing the input as-is to the unit after it. TF-Slim: TensorFlow (see here and here ) is an open-source software library for dataflow programming across a range of tasks. It is mainly used for machine learning with an emphasis on deep learning research and production. It provides good control of your network structure and functioning, with access to the inner workings of your deep neural network, such as the ability to directly update weights and gradients. Among the more advanced tools included in TensorFlow are highly flexible higher-level constructs, including Estimator, Experiment, and Dataset , which help you set-up your learning process in an easier way. Additional tools include: queues for computing tensors asynchronously in a graph, parallel- thread computation to speed up training. TensorFlow ships with powerful debugging tools, providing insight into internal structure and states. TF-Slim is a super-lightweight library written using a Functional Programming style , that can be used right alongside any of TensorFlow's native operations. TF- Slim is a library that makes building, training and evaluating neural networks simple: the use of argument scoping and high level layers and regularizers allows the user to define models much more compactly. These tools increase both readability and maintainability; they also reduce the likelihood of error. Slim makes it easy to extend widely used computer vision models (e.g. VGG, AlexNet), 14 Computer Vision News Tool We Tried for You: TensorFlow-Slim