Deep Learning with TensorFlow on Qubole

Deep learning works on large volumes of unstructured data such as human speech, text, and images to enable powerful use cases such as speech-to-text transcription, voice identification, image classification, facial or object recognition, analysis of sentiment or intent from text, and many more. In the last few years, TensorFlow has become a very popular deep learning framework for image recognition and speech detection use cases. All deep learning methods, including TensorFlow, require large volumes of data to train the model. Today, the most significant challenge in deep learning is the ever-increasing training time — as models get more complicated, the size of training data continues to increase. In order to address this challenge, cloud providers have launched instance types with many graphics processing units (GPUs) in a single node. However, using all of the GPUs in a single training job is not trivial. Qubole’s TensorFlow engine has been built to run on distributed Graphics Processing Units (GPUs) on Amazon Web Services. In this webinar we will: - Discuss how Qubole has achieved single-node, multi-GPU parallelization using native Tensorflow and Keras with Tensorflow as a backend. - Present results from our studies that show how training time varies with the number of GPUs in the cluster. - Run through a demo of a TensorFlow use case on Qubole.