December 2016 Summaries

Rescale's platform offers a comprehensive file management solution for cloud storage, utilizing major public cloud vendors with features like optimized transfer tools and encryption. It allows users to manage decrypted file content exclusively via Rescale's tools, but also supports integration with existing data pipelines and customer-controlled storage locations for simulations. The platform enables efficient compute resource allocation close to storage, supporting multiple cloud providers. A detailed example using Amazon S3 illustrates how users can set up a Design of Experiments (DOE) job, uploading input and output files to specified S3 directories, and configuring access through IAM policies. This setup involves creating local configuration files and scripts, which are encrypted and stored in Rescale-managed storage, while the core input and output files remain in user-specified buckets. Future developments aim to enhance integration with customer storage accounts.

The text explores the integration of the Hylang programming language, a Lisp-like dialect for the Python interpreter, with Emacs, focusing on the efficiency of REPL-centric interaction in programming. It highlights the advantages of using Emacs for Lisp languages, where code evaluation can be done directly at the cursor, fostering a seamless feedback loop ideal for prototyping and experimentation. The author demonstrates updating a Keras training job tutorial using Hylang instead of raw Python, providing code snippets and discussing the benefits of concise syntax and threading macros in Hylang. The text also addresses challenges such as debugging and the quirks of REPL integration but emphasizes the power and enjoyment of using Hylang within the Python ecosystem, particularly for those already familiar with Lisp languages. Ultimately, it illustrates how Hylang can be effectively used for various applications, including web development and machine learning, while offering a practical example and video demonstration of Emacs interaction.

Style transfer neural networks offer a way to apply artistic styles to images by leveraging existing image classification networks as loss functions, which helps train a new network to blend the semantic features of a target image with the textures of a style image. The process involves two steps: training the style transfer network and then applying it to new images. Using JC Johnson’s fast-neural-style implementation and pre-trained VGG16 network, the training is efficiently conducted on Rescale's cloud platform with K80 GPUs. The tutorial provided details on executing a job on Rescale, which involves uploading necessary files such as the fast-neural-style software, a dataset, the target image, and the style image, and then running specific scripts to build and apply the style transfer model. Once trained, the model can be reused to style additional images without re-running the training process, and the results are accessible through Rescale's platform.

Deep Learning is rapidly advancing fields like autonomous driving, robotics, machine translation, and facial recognition due to its ability to process large data volumes and recognize patterns. Choosing the right deep learning approach requires considering factors such as data volume, labeling, and the need for model transparency. Deep learning often necessitates specialized GPU hardware, posing challenges for organizations lacking such infrastructure. Rescale, in partnership with IBM, offers a Deep Learning Cloud that provides an integrated platform for designing and training deep neural networks, leveraging IBM Cloud’s NVIDIA servers. This platform streamlines data management, GPU access, and software updates, facilitating efficient model training and deployment. Users can upload data, configure training clusters, monitor progress, and review results through Rescale’s tools, enhancing the scalability and accessibility of deep learning applications.