Ishitori commented on a change in pull request #11274: [MXNET-547] Tutorial explaining how to use the profiler URL: https://github.com/apache/incubator-mxnet/pull/11274#discussion_r195571787
########## File path: docs/tutorials/python/profiler.md ########## @@ -0,0 +1,198 @@ +# Profiling MXNet Models + +It is often helpful to understand what operations take how much time while running a model. This helps optimize the model to run faster. In this tutorial, we will learn how to profile MXNet models to measure their running time and memory consumption using the MXNet profiler. + +## The incorrect way to profile + +If you have just begun using MXNet, you might be tempted to measure the execution time of your model using Python's `time` module like shown below: + +```python +from time import time +from mxnet import autograd, nd +import mxnet as mx + +start = time() +x = nd.random_uniform(shape=(2000,2000)) +y = nd.dot(x, x) +print('Time for matrix multiplication: %f sec\n' % (time() - start)) + +start = time() +print(y.asnumpy()) +print('Time for printing the output: %f sec' % (time() - start)) +``` + + +**Time for matrix multiplication: 0.005051 sec**<!--notebook-skip-line--> + +[[501.1584 508.29724 495.65237 ... 492.84705 492.69092 490.0481 ]<!--notebook-skip-line--> + + [508.81058 507.1822 495.1743 ... 503.10526 497.29315 493.67917]<!--notebook-skip-line--> + + [489.56598 499.47015 490.17722 ... 490.99945 488.05008 483.28836]<!--notebook-skip-line--> + + ...<!--notebook-skip-line--> + + [484.0019 495.7179 479.92142 ... 493.69952 478.89194 487.2074 ]<!--notebook-skip-line--> + + [499.64932 507.65094 497.5938 ... 493.0474 500.74512 495.82712]<!--notebook-skip-line--> + + [516.0143 519.1715 506.354 ... 510.08878 496.35608 495.42523]]<!--notebook-skip-line--> + +**Time for printing the output: 0.167693 sec**<!--notebook-skip-line--> + + +From the output above, it seems as if printing the output takes lot more time that multiplying two large matrices. That doesn't feel right. + +This is because, in MXNet, all operations are executed asynchronously. So, when `nd.dot(x, x)` returns, the matrix multiplication is not complete, it has only been queued for execution. `asnumpy` in `print(y.asnumpy())` however, waits for the result to be computed and hence takes longer time. + +While it is possible to use `NDArray.waitall()` before and after operations to get running time of operations, it is not a scalable method to measure running time of multiple sets of operations, especially in a Sequential or Hybrid network. + +## The correct way to profile + +The correct way to measure running time of MXNet models is to use MXNet profiler. In the rest of this tutorial, we will learn how to use the MXNet profiler to measure the running time and memory consumption of MXNet models. + +To use the profiler, you need to build MXNet with `USE_PROFILER` enabled. For example, this command will build the CPU version of MXNet on Linux, + +``` +make -j $(nproc) USE_OPENCV=1 USE_BLAS=openblas USE_PROFILER=1 +``` + +Check [this](http://mxnet.incubator.apache.org/install/index.html?device=Linux&language=Python&processor=CPU) page for more information on building from source for various environments. + +After building with `USE_PROFILER=True` and installing, you can import the profiler and configure it from Python code. + +```python +from mxnet import profiler +profiler.set_config(profile_all=True, aggregate_stats=True, filename='profile_output.json') +``` + +`profile_all` enables all types of profiling. You can also individually enable the following types of profiling: + +- `profile_symbolic` (boolean): whether to profile symbolic operators +- `profile_imperative` (boolean): whether to profile imperative operators +- `profile_memory` (boolean): whether to profile memory usage +- `profile_api` (boolean): whether to profile the C API + +`aggregate_stats` aggregates statistics in memory which can then be printed to console by calling `profiler.dumps()`. + +### Setup: Build a model + +Let's build a small convolutional neural network that we can use for profiling. + +```python +from mxnet import gluon +net = gluon.nn.HybridSequential() +with net.name_scope(): + net.add(gluon.nn.Conv2D(channels=20, kernel_size=5, activation='relu')) + net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2)) + net.add(gluon.nn.Conv2D(channels=50, kernel_size=5, activation='relu')) + net.add(gluon.nn.MaxPool2D(pool_size=2, strides=2)) + net.add(gluon.nn.Flatten()) + net.add(gluon.nn.Dense(512, activation="relu")) + net.add(gluon.nn.Dense(10)) +``` + +We need data that we can run through the network for profiling. We'll use the MNIST dataset. + +```python +from mxnet.gluon.data.vision import transforms +train_data = gluon.data.DataLoader(gluon.data.vision.MNIST(train=True).transform_first(transforms.ToTensor()), + batch_size=64, shuffle=True) +``` + +Let's define a method that will run one training iteration given data and label. + +```python +# Use GPU is available Review comment: Typo? "Use GPU **if** available" ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services