Curious to see all the traction this has received. I wanted to empthasise again that the intention here was not a speed-comparison but a Rosetta Stone of deep-learning frameworks. This example repo looks at small-dataset in RAM (generator function that yields mini-batches of numpy arrays, with no pre-processing or data manipulation) and applies a standard CNN to it. Results may drastically vary once we have more advanced pipelines.
The notebooks are not specifically written for speed, instead they aim to create an easy comparison between the frameworks. However, any suggestions on improving the training-time are welcome
Notebooks are run on Nvidia K80 GPU, on Microsoft Azure Data Science Virtual Machine for Linux (Ubuntu), where frameworks have been updated to the latest version
Create a Rosetta Stone of deep-learning frameworks to allow data-scientists to easily leverage their expertise from one framework to another (by translating, rather than learning from scratch). Also, to make the models more transparent to comparisons in terms of training-time and default-options.
A lot of online tutorials use very-low level APIs, which are very verbose, and don't make much sense (given higher-level helpers being available) for most use-cases unless one plans to create new layers. Here we try to apply the highest-level API possible, conditional on being to override conflicting defaults, to allow an easier comparison between frameworks. It will demonstrated that the code structure becomes very similar once higher-level APIs are used and can be roughly represented as:
- Load data; x_train, x_test, y_train, y_test = cifar_for_library(channel_first=?, one_hot=?)
- Generate CNN symbol (usually no activation on final dense-layer)
- Specify loss (cross-entropy comes bundles with softmax), optimiser and initialise weights + sessions
- Train on mini-batches from train-set using custom iterator (common data-source for all frameworks)
- Predict on fresh mini-batches from test-set
- Evaluate accuracy
Since we are essentially comparing a series of deterministic mathematical operations (albeit with a random initialization), it does not make sense to me to compare the accuracy across frameworks and instead they are reported as checks we want to match, to make sure we are comparing the same model architecture.
| DL Library | Test Accuracy (%) | Training Time (s) |
|---|---|---|
| Tensorflow (1.2.1) | 72 | 300 |
| CNTK (2.1) | 77 | 168 |
| MXNet (0.11.0) | 75 | 153 |
| PyTorch (0.2.0_1) | 73 | 165 |
| Chainer (2.0.2) | 78 | 256 |
| Keras (2.0.6) (TF) | 77 | 408 |
| Keras (2.0.6) (CNTK) | 76 | 588 |
| Caffe2 | 75 | 312 |
| Lasagne (0.2.dev1) (Theano 0.10.0beta1) | 73 | 416 |
EDIT 2: Enabling cudnn's auto-tuner for PyTorch as suggested by sousmith almost halved the time from 351 to 165s. Thanks to botev for adding the theano + lasagne notebook
EDIT 1: I received some comments that perhaps that reason MXNet is so much faster than the others is because I use its own custom generator. Below is a notebook where I use the same generator as with other frameworks and the result does not change
| DL Library | Test Accuracy (%) | Training Time (s) |
|---|---|---|
| MXNet Custom Generator | 77 | 159 |