It took a long time to start running on a GPU¶

Try to disable opencv to use a GPU: build opencv from source with GPU module disabled.

It’s slow on a single GPU¶

Check the following:

1. Ensure that your CUDA/driver version is not too old.
2. Build with USE_CUDNN=1. This often increases speed 50+%. Try to use the newest version.
3. Set export MXNET_CUDNN_AUTOTUNE_DEFAULT=1 before running. This often increases speed 10%-15%.
4. If you are using Tesla GPUs by nvidia-smi -e 0, disable ECC. You might need root permission and have to reboot.
5. For Tesla cards by nvidia-smi -ac ??, set to the maximal clock. For details, see this blog.
6. No throttle reason nvidia-smi -q -d PERFORMANCE is often caused by temperature.

No increase in speed when using more than one GPU or computer¶

Check the following:

1. Does your neural network already run fast, such as >1000 example/sec or >10 batches/sec? If yes, it’s unlikely to speed up any further by adding more resources because of the communication overhead.
2. Are you using a small batch size? Try to increase it.
3. Are you using more than 4 GPUs? Try using --kv-store=device.

Abnormal CPU memory usage¶

This might be due to the data pre-fetch. See issue 2111.

How to Copy Part of Parameters to Another Model¶

Most MXNet model consists two parts, the argument arrays and symbol. You can simply copy the argument array to the argument array of another model. For example, in the Python model API, you can do this:

copied_model =  mx.model.FeedForward(ctx=mx.gpu(), symbol=new_symbol,
                                     arg_params=old_arg_params, aux_params=old_aux_params,
                                     allow_extra_params=True);

For information about copying model parameters from an existing old_arg_params, see this notebook. More notebooks please refer to dmlc/mxnet-notebooks.

How to Extract the Feature Map of a Certain Layer¶

See this notebook. More notebooks please refer to dmlc/mxnet-notebooks.

What Is the Relationship Between MXNet and CXXNet, Minerva, and Purine2?¶

MXNet is created in collaboration by authors from the three projects. MXNet reflects what we have learned from these projects. It combines the important aspects of the existing projects: general efficiency, flexibility, and memory efficiency.

MXNet also contains new approaches that allow you to combine different ways of programming and write CPU/GPU applications that are more memory efficient than CXXNet and Purine, and more flexible than Minerva.

What Is the Relationship Between MXNet and TensorFlow?¶

Both MXNet and TensorFlow use computation graph abstraction, which was initially used by Theano, then adopted by other packages, such as CGT, Caffe2, and Purine. Currently, TensorFlow adopts an optimized symbolic API. MXNet supports a mixed approach, with a dynamic dependency scheduler to combine symbolic and imperative programming. In short, MXNet is lightweight and mixed with flexibility from imperative programming, while using a computation graph to make it very fast and memory efficient.