This is the companion code for DL benchmarking study reported in the paper Comparative Study of Deep Learning Software Frameworks by Soheil Bahrampour, Naveen Ramakrishnan, Lukas Schott, and Mohak Shah. The paper can be found here http://arxiv.org/abs/1511.06435. The code allows the users to reproduce and extend the results reported in the study. The code provides timings of forward run and forward+backward (gradient computation) run of several deep learning architecture using Caffe, Neon, TensorFlow, Theano, and Torch. The deep learning architectures used includes LeNet, AlexNet, LSTM, and a stacked AutoEncoder. Please cite the above paper when reporting, reproducing or extending the results.

Here you can find a set of new timings obtained using cuDNNv4 on a single M40 GPU on the same experiments performed in the paper. The result are reported using Caffe-Nvidia 0.14.5, Neon 1.5.4, Tensoflow 0.9.0rc0, Theano 0.8.2, and Torch7. Note that Neon does not use cuDNN.

1. LeNet using batch size of 64 (Extension of Table 3 in the paper)

2. Alexnet using batch size of 256 (Extension of Table 4 in the paper)

3. LSTM using batch size of 16 (Extension of Table 6 in the paper)

4. Stacked auto-encoder with encoder dimensions of 400, 200, 100 using batch size of 64 (Extension of Table 5 in the paper)

5. Stacked auto-encoder with encoder dimensions of 800, 1000, 2000 using batch size of 64 (Extension of Table 7 in the paper)

See the readme file within each folder to run the experiments.

DL-benchmarks is open-sourced under the MIT license. See the LICENSE file for details.

For a list of other open source components included in DL-benchmarks, see the file 3rd-party-licenses.txt.