- Python unit testing framework aka PyUnit
- Numpy Testing assertions
- Skip examples and jump straight to the Instructions (Go through the examples at least once)
DNNC framework is easy to get started since most of APIs closely resemble numpy. For the sake of simplicity, team has chosen to go with Row Major form of tensors for the beta releases. DNNC supports upto 4D tensors in the beta release
>>> import dnnc as dc
>>> a=dc.arange(5) ; # create a vector of 5 elements
>>> print(a)
[0.000000 1.000000 2.000000 3.000000 4.000000]
>>> a[1] # print second element of the array.
1.0
>>> a[1] = 100.1 # assign second element a new value 100.1
>>> print(a) # print and check new value.
[0.000000 100.099998 2.000000 3.000000 4.000000]
>>> a=dc.array([[10,11,12],[20,21,22]]) # create 2D arrary from python list
>>> a.shape() # check shape.
(2, 3)
>>> print(a) # print 2D arrary.
[[10.000000 11.000000 12.000000]
[12.000000 20.000000 21.000000]]
>>> a.reshape(dc.ivec([3,2])) # reshape 2x3 matrix to 3x2 matrix
>>> a.shape()
(3, 2)
>>> print(a)
[[10.000000 11.000000]
[20.000000 21.000000]
[0.000000 0.000000]]Other tensor functions
>>> a.<tab><tab>
a.broadcast( a.empty( a.name( a.this
a.data( a.flatten( a.rank( a.to_proto(
a.dtype( a.length( a.reshape( a.to_string(
a.eigen_to_numpy( a.load( a.shape( a.transpose( Matrix Multiplication Example
>>> a=dc.array([[10,11,12],[20,21,22]]) # create 'a' 2x3 matrix
>>> b=dc.array([[10,11,12],[20,21,22]]) # create 'b' 2x3 matrix
>>> dc.reshape(b,(3,2)) # reshape matrix 'b' to 3x2
>>> b.shape()
(3, 2)
>>> y=dc.matmul(a,b) # multiply 'a' and 'b'.
>>> print(y)
[[484.000000 594.000000]
[914.000000 1124.000000]]Other DNNC APIs
>>> dc.<tab><tab>
dc.add( dc.matmul(
dc.arange( dc.ones(
dc.array( dc.random(
dc.dTensor( dc.reshape(
dc.empty( dc.zeros(
dc.ivec dc.fvec
dc.iTensor( dc.fTensor(
dc.thresholded_relu(import dnnc as dc
import numpy as np
import unittest
class MatMulTest(unittest.TestCase):
def setUp(self):
self.len = 12
self.np_a = np.random.randn(self.len).astype(np.float32)
self.np_b = np.random.randn(self.len).astype(np.float32)
self.dc_a = dc.array(list(self.np_a));
self.dc_b = dc.array(list(self.np_b));
def test_MatMul_1D (self):
npr = np.matmul(self.np_a, self.np_b)
dcr = dc.matmul(self.dc_a, self.dc_b)
np.testing.assert_allclose(npr, np.array(dcr.data()[0]).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_MatMul_2D (self):
np_a = np.reshape(self.np_a, (3,4))
np_b = np.reshape(self.np_b, (4,3))
dc_a = dc.reshape(self.dc_a, (3,4));
dc_b = dc.reshape(self.dc_b, (4,3));
npr = np.matmul(np_a, np_b);
dcr = dc.matmul(dc_a, dc_b);
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_MatMul_3D (self):
np_a = np.reshape(self.np_a, (2,2,3))
np_b = np.reshape(self.np_b, (2,3,2))
dc_a = dc.reshape(self.dc_a, (2,2,3));
dc_b = dc.reshape(self.dc_b, (2,3,2));
npr = np.matmul(np_a, np_b);
dcr = dc.matmul(dc_a, dc_b);
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
if __name__ == '__main__':
unittest.main()-
I have moved your work on the python interface i.e. dnnc.i and dnnc_api.cpp to a temporary folder
-
Because not all the operators work with the interface, so be sure to only push the repository when your
makecommand doesn't fail. This is very important, as this could take your fellow developers' valuable time trying to fix the bug, rather than implementing test cases. -
So only push when you add the header files in the interface, and that works with python interface.
-
So below is the cycle you have to go through, in the next days.
-
Add one of your operator in the dnnc.i and dnnc_api.cpp.
-
Go to the dnnc-operator/swig folder, and run
makecommand to compile the newly added operators in the interface. -
If everything works fine, then you successfully added your operator in the dnnc python interface. And you can check it by importing dnnc in python bash.
-
Now go to the dnnc-operator/test/swig folder, here you will see the MatMul.py and Add.py examples. Add your operators like the above examples.
-
After adding your code, run the code from the same directory like this.
python MatMul.pyor you can also use
python -m unittest MatMul.MatMulTest -
If you want to see Verbose output of your execution, add
-vat the end of your command. Likepython MatMul.py -v -
It will show which test cases work, and which doesn't and shows the reason too.
-
To run all of the unittests at once, use this.
python -m unittest *py -v -
Then update the README to show which test case is not working, or if all are working.
-
Now look at the code, here are some things, if you need to know about the lines of the implementation.
-
import os,sys
# adding "dnnc-operators/swig" folder to the import directory, we will import "dnnc.py" from here
sys.path.append(os.path.abspath('..'+os.path.sep+'..'+os.path.sep+'swig'));
import dnnc as dc # import dnnc
import numpy as np # import numpy to cross check against dnnc
import unittest # to test our testcases
class AddTest(unittest.TestCase):
# This is like __init__ method of class, but for unittest, it just
# declares the variables everytime before testcases are tested.
def setUp(self):
# declare the total size of the tensor
self.len = 24
# use numpy random to generate random numbers
self.np_a = np.random.randn(self.len).astype(np.float32)
self.np_b = np.random.randn(self.len).astype(np.float32)
#self.np_a = np.arange(self.len).astype(np.float32)
#self.np_b = np.arange(self.len).astype(np.float32)
# set the same numbers as input to dnnc, so cross checking happens on same input
self.dc_a = dc.array(list(self.np_a));
self.dc_b = dc.array(list(self.np_b));
# First test case for !D
# remember, for unittest module to know which method should be tested, you have to give
# test_as prefix to the method
def test_Add1D (self):
# npr is the tensor which stores the result of numpy add of those two input tensors
npr = np.add(self.np_a, self.np_b)
# dcr is the tensor which stores the result of dnnc add of those same input tensors
dcr = dc.add(self.dc_a, self.dc_b)
# assert_allclose checks the two tensor values, with tolerance of
# atol (Absolute tolerance), rtol (Relative tolerance)
np.testing.assert_allclose(npr, np.array(dcr.data()[0]).astype(np.float32),
rtol=1e-3, atol=1e-3)
# Second test case for 2D
def test_Add2D (self):
#for 2d we have to reshape each of np_a, np_b, dc_a, dc_b
np_a = np.reshape(self.np_a, (6,4))
np_b = np.reshape(self.np_b, (6,4))
dc_a = dc.reshape(self.dc_a, (6,4));
dc_b = dc.reshape(self.dc_b, (6,4));
npr = np.add(np_a, np_b);
dcr = dc.add(dc_a, dc_b);
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_Add3D (self):
np_a = np.reshape(self.np_a, (2,4,3))
np_b = np.reshape(self.np_b, (2,4,3))
dc_a = dc.reshape(self.dc_a, (2,4,3));
dc_b = dc.reshape(self.dc_b, (2,4,3));
npr = np.add(np_a, np_b);
dcr = dc.add(dc_a, dc_b);
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_Add4D (self):
np_a = np.reshape(self.np_a, (2,2,2,3))
np_b = np.reshape(self.np_b, (2,2,2,3))
dc_a = dc.reshape(self.dc_a, (2,2,2,3));
dc_b = dc.reshape(self.dc_b, (2,2,2,3));
npr = np.add(np_a, np_b);
dcr = dc.add(dc_a, dc_b);
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
# when the program gets called
if __name__ == '__main__':
unittest.main()