...
Full Bio
Follow on
How Android users will be easily Fascinated by Artificial Intelligence?
yesterday
Clear your fundamentals with some Artificial Intelligence Courses
2 days ago
See How Hadoop easily distorted Big Data
2 days ago
Chatbots have taken over humans. But How?
2 days ago
Keep in Mind some points before Learning Machine Learning
3 days ago
Every Programmer should strive for reading these 5 books
522078 views
Highest Paid Programming Languages With Highest Market Demand
121716 views
Python Programming Language can easily be acquired easily. How?
103446 views
2019: New programming languages which might enhance your skills in future
97740 views
A clear picture of C, C++, and Java through famous Websites
97311 views
Demonstration of Deep Learning models Apache, Sparx, and TensorFlow
Neural networks have seen spectacular progress during the last few years and they are now the state of the art in image recognition and automated translation. TensorFlow is a new framework released by Google for numerical computations and neural networks. In this blog post, we are going to demonstrate how to use TensorFlow and Spark together to train and apply deep learning models.
You might be wondering: What's Apache Spark's use here when most high-performance deep learning implementations are single-node only? To answer this question, we walk through two use cases and explain how you can use Spark and a cluster of machines to improve deep learning pipelines with TensorFlow:
Hyperparameter Tuning: use Spark to find the best set of hyperparameters for neural network training, leading to 10X reduction in training time and 34% lower error rate.
Deploying models at scale: use Spark to apply a trained neural network model on a large amount of data.
Hyperparameter Tuning
An example of a deep learning machine learning (ML) technique is artificial neural networks. They take a complex input, such as an image or an audio recording, and then apply complex mathematical transforms on these signals. The output of this transform is a vector of numbers that is easier to manipulate by other ML algorithms. Artificial neural networks perform this transformation by mimicking the neurons in the visual cortex of the human brain (in a much-simplified form).
Just as humans learn to interpret what they see, artificial neural networks need to be trained to recognize specific patterns that are â??interesting'. For example, these can be simple patterns such as edges, circles, but they can be much more complicated. Here, we are going to use a classical dataset put together by NIST and train a neural network to recognize these digits:
The TensorFlow library automates the creation of training algorithms for neural networks of various shapes and sizes. The actual process of building a neural network, however, is more complicated than just running some function on a dataset. There are typically a number of very important hyperparameters (configuration parameters in layman's terms) to set, which affects how the model is trained. Picking the right parameters leads to high performance, while bad parameters can lead to prolonged training and bad performance. In practice, machine learning practitioners rerun the same model multiple times with different hyperparameters in order to find the best set. This is a classical technique called hyperparameter tuning.
When building a neural network, there are many important hyperparameters to choose carefully. For example:
- A number of neurons in each layer: Too few neurons will reduce the expression power of the network, but too many will substantially increase the running time and return noisy estimates.
- Learning rate: If it is too high, the neural network will only focus on the last few samples seen and disregard all the experience accumulated before. If it is too low, it will take too long to reach a good state.
The interesting thing here is that even though TensorFlow itself is not distributed, the hyperparameter tuning process is "embarrassingly parallel" and can be distributed using Spark. In this case, we can use Spark to broadcast the common elements such as data and model description and then schedule the individual repetitive computations across a cluster of machines in a fault-tolerant manner.
How does using Spark improve accuracy? The accuracy with the default set of hyperparameters is 99.2%. Our best result with hyperparameter tuning has a 99.47% accuracy on the test set, which is a 34% reduction of the test error. Distributing the computations scaled linearly with the number of nodes added to the cluster: using a 13-node cluster, we were able to train 13 models in parallel, which translates into a 7x speedup compared to training the models one at a time on one machine. Here is a graph of the computation times (in seconds) with respect to the number of machines on the cluster:
This shows a typical tradeoff curve for neural networks:
- The learning rate is critical: if it is too low, the neural network does not learn anything (high test error). If it is too high, the training process may oscillate randomly and even diverge in some configurations.
- The number of neurons is not as important for getting a good performance, and networks with many neurons are much more sensitive to the learning rate. This is Occam's Razor principle: simpler model tend to be "good enough" for most purposes. If you have the time and resource to go after the missing 1% test error, you must be willing to invest a lot of resources in training and to find the proper hyperparameters that will make the difference.
By using a sparse sample of parameters, we can zero in on the most promising sets of parameters.
How do we use it?
Since TensorFlow can use all the cores on each worker, we only run one task at one time on each worker and we batch them together to limit contention. The TensorFlow library can be installed on Spark clusters as a regular Python library, following the instructions on the TensorFlow website. The following notebooks below show how to install TensorFlow and let users rerun the experiments of this blog post:
- Distributed processing of images using TensorFlow
- Testing the distribution processing of images using TensorFlow
Source: HOB