You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@mxnet.apache.org by jx...@apache.org on 2017/12/14 19:27:09 UTC
[incubator-mxnet] branch master updated: Fix examples of profiler
and cpp-package (#9065)
This is an automated email from the ASF dual-hosted git repository.
jxie pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/incubator-mxnet.git
The following commit(s) were added to refs/heads/master by this push:
new bfa0989 Fix examples of profiler and cpp-package (#9065)
bfa0989 is described below
commit bfa0989e5aabe199e3dc661e552a62f67dd4d185
Author: reminisce <wu...@gmail.com>
AuthorDate: Thu Dec 14 11:27:06 2017 -0800
Fix examples of profiler and cpp-package (#9065)
* Fix profiler examples
* Fix cpp-package example
* More readme
---
docs/tutorials/c++/basics.md | 17 ++++++++++++-----
example/profiler/README.md | 23 +++++++++++++++++++++++
example/profiler/profiler_executor.py | 2 +-
example/profiler/profiler_imageiter.py | 6 +++---
example/profiler/profiler_matmul.py | 7 +++----
example/profiler/profiler_ndarray.py | 12 ++++++++++--
6 files changed, 52 insertions(+), 15 deletions(-)
diff --git a/docs/tutorials/c++/basics.md b/docs/tutorials/c++/basics.md
index cdf1a28..d3231e7 100644
--- a/docs/tutorials/c++/basics.md
+++ b/docs/tutorials/c++/basics.md
@@ -16,8 +16,8 @@ Except linking the MXNet shared library, the C++ package itself is a header-only
which means all you need to do is to include the header files. Among the header files,
`op.h` is special since it is generated dynamically. The generation should be done when
[building the C++ package](http://mxnet.io/get_started/build_from_source.html#build-the-c++-package).
-After that, you also need to copy the shared library (`libmxnet.so` in linux,
-`libmxnet.dll` in windows) from `/path/to/mxnet/lib` to the working directory.
+It is important to note that you need to **copy the shared library** (`libmxnet.so` in Linux and MacOS,
+`libmxnet.dll` in Windows) from `/path/to/mxnet/lib` to the working directory.
We do not recommend you to use pre-built binaries because MXNet is under heavy development,
the operator definitions in `op.h` may be incompatible with the pre-built version.
@@ -49,7 +49,7 @@ auto val_iter = MXDataIter("MNISTIter")
.CreateDataIter();
```
-The data have been successfully loaded, we can now easily construct various models to identify
+The data have been successfully loaded. We can now easily construct various models to identify
the digits with the help of C++ package.
@@ -159,7 +159,12 @@ while (val_iter.Next()) {
```
You can find the complete code in `mlp_cpu.cpp`. Use `make mlp_cpu` to compile it,
- and `./mlp_cpu` to run it.
+ and `./mlp_cpu` to run it. If it complains that the shared library `libmxnet.so` is not found
+ after typing `./mlp_cpu`, you will need to specify the path to the shared library in
+ the environment variable `LD_LIBRARY_PATH` in Linux and `DYLD_LIBRARY_PATH`
+ in MacOS. For example, if you are using MacOS, typing
+ `DYLD_LIBRARY_PATH+=. ./mlp_cpu` would solve the problem. It basically tells the system
+ to find the shared library under the current directory since we have just copied it here.
GPU Support
-----------
@@ -186,4 +191,6 @@ data_batch.label.CopyTo(&args["label"]);
NDArray::WaitAll();
```
-By replacing the former code to the latter one, we successfully port the code to GPU. You can find the complete code in `mlp_gpu.cpp`. Compilation is similar to the cpu version. (Note: The shared library should be built with GPU support on)
+By replacing the former code to the latter one, we successfully port the code to GPU.
+You can find the complete code in `mlp_gpu.cpp`. Compilation is similar to the cpu version.
+Note that the shared library must be built with GPU support enabled.
diff --git a/example/profiler/README.md b/example/profiler/README.md
new file mode 100644
index 0000000..7d3c42b
--- /dev/null
+++ b/example/profiler/README.md
@@ -0,0 +1,23 @@
+# MXNet Profiler Examples
+
+This folder contains examples of using MXNet profiler to generate profiling results in json files.
+Please refer to [this link](http://mxnet.incubator.apache.org/faq/perf.html?highlight=profiler#profiler)
+for visualizing profiling results and make sure that you have installed a version of MXNet compiled
+with `USE_PROFILER=1`.
+
+- profiler_executor.py. To run this example, simply type `python profiler_executor.py` in terminal.
+It will generate a json file named `profile_executor_5iter.json`.
+
+- profiler_imageiter.py. You first need to create a file named `test.rec`,
+which is an image dataset file before running this example.
+Please follow
+[this tutorial](https://mxnet.incubator.apache.org/faq/recordio.html?highlight=rec%20file#create-a-dataset-using-recordio)
+on how to create `.rec` files using an existing tool in MXNet. After you created 'test.rec',
+type `python profiler_imageiter.py` in terminal. It will generate `profile_imageiter.json`.
+
+- profiler_matmul.py. This example profiles matrix multiplications on GPU. Please make sure
+that you have installed a GPU enabled version of MXNet before running this example. Type
+`python profiler_matmul.py` and it will generate `profile_matmul_20iter.json`.
+
+- profiler_ndarray.py. This examples profiles a series of `NDArray` operations. Simply type
+`python profiler_ndarray.py` in terminal and it will generate `profile_ndarray.json`.
\ No newline at end of file
diff --git a/example/profiler/profiler_executor.py b/example/profiler/profiler_executor.py
index 26e3e1b..117a8df 100644
--- a/example/profiler/profiler_executor.py
+++ b/example/profiler/profiler_executor.py
@@ -17,7 +17,7 @@
import mxnet as mx
import argparse
-import os, sys
+import os
import time
import numpy as np
from mxnet import profiler
diff --git a/example/profiler/profiler_imageiter.py b/example/profiler/profiler_imageiter.py
index e16b9b7..77ca412 100644
--- a/example/profiler/profiler_imageiter.py
+++ b/example/profiler/profiler_imageiter.py
@@ -15,16 +15,15 @@
# specific language governing permissions and limitations
# under the License.
+from __future__ import print_function
import os
# uncomment to set the number of worker threads.
# os.environ["MXNET_CPU_WORKER_NTHREADS"] = "4"
-from __future__ import print_function
import time
import mxnet as mx
-import numpy as np
-def run_imageiter(path_rec, n, batch_size = 32):
+def run_imageiter(path_rec, n, batch_size=32):
data = mx.img.ImageIter(batch_size=batch_size,
data_shape=(3, 224, 224),
@@ -39,6 +38,7 @@ def run_imageiter(path_rec, n, batch_size = 32):
mx.nd.waitall()
print(batch_size*n/(time.time() - tic))
+
if __name__ == '__main__':
mx.profiler.profiler_set_config(mode='all', filename='profile_imageiter.json')
mx.profiler.profiler_set_state('run')
diff --git a/example/profiler/profiler_matmul.py b/example/profiler/profiler_matmul.py
index 1b1cf74..a23545c 100644
--- a/example/profiler/profiler_matmul.py
+++ b/example/profiler/profiler_matmul.py
@@ -18,9 +18,8 @@
from __future__ import print_function
import mxnet as mx
import argparse
-import os, sys
import time
-import numpy as np
+
def parse_args():
parser = argparse.ArgumentParser(description='Set network parameters for benchmark test.')
@@ -30,18 +29,18 @@ def parse_args():
parser.add_argument('--end_profiling_iter', type=int, default=70)
return parser.parse_args()
+
args = parse_args()
if __name__ == '__main__':
mx.profiler.profiler_set_config(mode='symbolic', filename=args.profile_filename)
print('profile file save to {0}'.format(args.profile_filename))
-
A = mx.sym.Variable('A')
B = mx.sym.Variable('B')
C = mx.symbol.dot(A, B)
- executor = C.simple_bind(mx.gpu(1), 'write', A=(4096, 4096), B=(4096, 4096))
+ executor = C.simple_bind(mx.gpu(0), 'write', A=(4096, 4096), B=(4096, 4096))
a = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
b = mx.random.uniform(-1.0, 1.0, shape=(4096, 4096))
diff --git a/example/profiler/profiler_ndarray.py b/example/profiler/profiler_ndarray.py
index 67ea87b..5c233c6 100644
--- a/example/profiler/profiler_ndarray.py
+++ b/example/profiler/profiler_ndarray.py
@@ -82,6 +82,7 @@ def random_ndarray(dim):
data = mx.nd.array(np.random.uniform(-10, 10, shape))
return data
+
def test_ndarray_elementwise():
np.random.seed(0)
nrepeat = 10
@@ -99,6 +100,7 @@ def test_ndarray_elementwise():
check_with_uniform(mx.nd.square, 1, dim, np.square, rmin=0)
check_with_uniform(lambda x: mx.nd.norm(x).asscalar(), 1, dim, np.linalg.norm)
+
def test_ndarray_negate():
npy = np.random.uniform(-10, 10, (2,3,4))
arr = mx.nd.array(npy)
@@ -170,6 +172,7 @@ def test_ndarray_scalar():
d = -c + 2
assert(np.sum(d.asnumpy()) < 1e-5)
+
def test_ndarray_pickle():
np.random.seed(0)
maxdim = 5
@@ -222,8 +225,7 @@ def test_ndarray_slice():
def test_ndarray_slice_along_axis():
arr = mx.nd.array(np.random.uniform(-10, 10, (3, 4, 2, 3)))
- sub_arr = mx.nd.zeros((3, 2, 2, 3))
- arr._copy_slice_to(1, 1, 3, sub_arr)
+ sub_arr = arr.slice(begin=(None, 1), end=(None, 3))
# test we sliced correctly
assert same(arr.asnumpy()[:, 1:3, :, :], sub_arr.asnumpy())
@@ -242,6 +244,7 @@ def test_clip():
assert B1[i] >= -2
assert B1[i] <= 2
+
def test_dot():
a = np.random.uniform(-3, 3, (3, 4))
b = np.random.uniform(-3, 3, (4, 5))
@@ -251,8 +254,10 @@ def test_dot():
C = mx.nd.dot(A, B)
assert reldiff(c, C.asnumpy()) < 1e-5
+
def test_reduce():
sample_num = 200
+
def test_reduce_inner(numpy_reduce_func, nd_reduce_func):
for i in range(sample_num):
ndim = np.random.randint(1, 6)
@@ -285,8 +290,10 @@ def test_reduce():
test_reduce_inner(lambda data, axis, keepdims:_np_reduce(data, axis, keepdims, np.min),
mx.nd.min)
+
def test_broadcast():
sample_num = 1000
+
def test_broadcast_to():
for i in range(sample_num):
ndim = np.random.randint(1, 6)
@@ -307,6 +314,7 @@ def test_broadcast():
assert err < 1E-8
test_broadcast_to()
+
if __name__ == '__main__':
mx.profiler.profiler_set_config(mode='all', filename='profile_ndarray.json')
mx.profiler.profiler_set_state('run')
--
To stop receiving notification emails like this one, please contact
['"commits@mxnet.apache.org" <co...@mxnet.apache.org>'].