Skip to content

unittests: move test_minnn.py to using unittests instead of manually written tests #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 5 commits into
base: main
Choose a base branch
from
Open

unittests: move test_minnn.py to using unittests instead of manually written tests #1

Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
50ba3df
first pass at making unittests, not working with args yet
gsireesh Sep 20, 2021
0630260
changed asserts and moved to use of np testing
gsireesh Oct 3, 2021
5f3f758
add ifmain for running from console
gsireesh Oct 3, 2021
85d4ebd
add missing newline
gsireesh Oct 3, 2021
c6d2b85
remove old tests
gsireesh Oct 3, 2021
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
228 changes: 101 additions & 127 deletions test_minnn.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,12 +5,15 @@
# -- usage is simply: python test_minnn.py <your-minnn.py>

import os
import unittest
from unittest import TestCase

import numpy as np

import minnn as test_mn

np.random.seed(123)

# --
# helpers

def search_file(dir: str, filename: str):
rets = []
Expand All @@ -20,132 +23,105 @@ def search_file(dir: str, filename: str):
assert len(rets) == 1
return rets[0]

def load_module(f: str, m: str):
import importlib.util
spec = importlib.util.spec_from_file_location(m, f)
ret = importlib.util.module_from_spec(spec)
spec.loader.exec_module(ret)
return ret

def is_allclose(x, y):
return np.allclose(x, y, rtol=1.e-3, atol=1.e-5)
# return np.abs(x-y).sum().item() <= 1e-5
return np.testing.assert_allclose(x, y, rtol=1.e-3, atol=1.e-5)

# --
# testings!
def test_accumulate_grad(test_mn):
g0 = np.asarray([1., 2.])
g1 = np.asarray([3., 4.])
t = test_mn.astensor([0., 0.])
t.accumulate_grad(g0)
t.accumulate_grad(g1)
assert is_allclose(t.get_dense_grad(), np.asarray([4., 6.]))
# --

def test_accumulate_grad_sparse(test_mn):
g0 = np.asarray([1., 2.])
g1 = np.asarray([3., 4.])
t = test_mn.astensor([[0., 0.], [0., 0.], [0., 0.]])
t.accumulate_grad_sparse([(0, g0), (2, g1), (2, g0)])
assert is_allclose(t.get_dense_grad(), np.asarray([g0, [0.,0.], g0+g1]))
# --

def test_xavier_uniform(test_mn):
shape = [100, 500]
y = test_mn.Initializer.xavier_uniform(shape)
assert list(y.shape) == shape
assert abs(y.mean()) <= 1e-2
assert abs(y.max()-0.1) <= 1e-2
assert abs(y.min()+0.1) <= 1e-2
# --

def test_momentum_update(test_mn):
m1 = test_mn.Model()
p1 = m1.add_parameters([2], 'constant', val=0.)
p2 = m1.add_parameters([2,2], 'constant', val=0.)
# --
t1 = test_mn.MomentumTrainer(m1, lrate=0.1, mrate=0.9)
for _ in range(5):
g = np.array([1.,1.])
p1.accumulate_grad(g)
t1.update()
assert is_allclose(p1.data, np.array([-0.131441, -0.131441])), \
"This is the values using our implementation, there can be other versions for momentum_update, you can choose to use others!"
for i in range(6):
g = np.array([1.,1.])
p2.accumulate_grad_sparse([((i%2), g.copy())])
t1.update()
assert is_allclose(p2.data, np.array([[-0.1002459, -0.1002459], [-0.078051, -0.078051]])), \
"This is the values using our implementation, there can be other versions for momentum_update, you can choose to use others!"
# --

def test_lookup(test_mn):
t = test_mn.astensor([[1., 2.], [3., 4.], [5., 6.]])
v = test_mn.lookup(t, [1,0,1])
assert is_allclose(v.data, np.asarray([[3., 4.], [1., 2.], [3., 4.]]))
v.accumulate_grad(np.asarray([[1., 1.], [1., 1.], [2., 3.]]))
v.op.backward()
assert is_allclose(t.get_dense_grad(), np.asarray([[1., 1.], [3., 4.], [0., 0.]]))

def test_dot(test_mn):
w, h = test_mn.astensor([[0., 1.], [2., 3.]]), test_mn.astensor([1., 2.])
v = test_mn.dot(w, h)
assert is_allclose(v.data, np.asarray([2., 8.]))
v.accumulate_grad(np.asarray([1., 3.]))
v.op.backward()
assert is_allclose(w.get_dense_grad(), np.asarray([[1.,2.],[3.,6.]]))
assert is_allclose(h.get_dense_grad(), np.asarray([6.,10.]))

def test_tanh(test_mn):
x = test_mn.astensor([0., 1., 2., 3.])
v = test_mn.tanh(x)
assert is_allclose(v.data, np.asarray([0., 0.76159416, 0.96402758, 0.99505475]))
v.accumulate_grad(np.asarray([1.,2.,3.,4.]))
v.op.backward()
assert is_allclose(x.get_dense_grad(), np.asarray([1., 0.83994868, 0.21195247, 0.03946415]))
# --

def test_avg(test_mn):
x = test_mn.astensor([[0., 1., 2.], [3., 5., 11.]])
v = test_mn.avg(x, 0)
assert is_allclose(v.data, np.asarray([1.5, 3., 6.5]))
v.accumulate_grad(np.asarray([1.,2.,3.]))
v.op.backward()
assert is_allclose(x.get_dense_grad(), np.asarray([[0.5,1.0,1.5], [0.5,1.0,1.5]]))
# --

def test_max(test_mn):
x = test_mn.astensor([[0., 5., 2.], [-2., 6., 1.]])
v = test_mn.max(x, 0)
assert is_allclose(v.data, np.asarray([0., 6., 2.]))
v.accumulate_grad(np.asarray([1.,2.,3.]))
v.op.backward()
assert is_allclose(x.get_dense_grad(), np.asarray([[1., 0., 3.], [0., 2., 0.]]))
# --
# helpers

# --
def main(minnn: str):
test_mn = load_module(minnn, 'mn')
# --
# note: to change weights in this table
test_table = [
("accumulate_grad", 1.,),
("accumulate_grad_sparse", 1.,),
("xavier_uniform", 1.,),
("momentum_update", 1.),
("lookup", 1.),
("dot", 1.),
("tanh", 1.),
("avg", 1.),
("max", 1.),
]
for name, weight in test_table:
test_f = globals()[f"test_{name}"]
test_f(test_mn)
print(f"Pass {name}!")
return
class TestMinnn(TestCase):


def test_accumulate_grad(self):
g0 = np.asarray([1., 2.])
g1 = np.asarray([3., 4.])
t = test_mn.astensor([0., 0.])
t.accumulate_grad(g0)
t.accumulate_grad(g1)
is_allclose(t.get_dense_grad(), np.asarray([4., 6.]))
# --

def test_accumulate_grad_sparse(self):
g0 = np.asarray([1., 2.])
g1 = np.asarray([3., 4.])
t = test_mn.astensor([[0., 0.], [0., 0.], [0., 0.]])
t.accumulate_grad_sparse([(0, g0), (2, g1), (2, g0)])
is_allclose(t.get_dense_grad(), np.asarray([g0, [0., 0.], g0 + g1]))
# --

def test_xavier_uniform(self):
shape = [100, 500]
y = test_mn.Initializer.xavier_uniform(shape)
self.assertEqual(list(y.shape), shape)
self.assertLessEqual(abs(y.mean()), 1e-2)
self.assertLessEqual(abs(y.max() - 0.1), 1e-2)
self.assertLessEqual(abs(y.min() + 0.1), 1e-2)
# --

def test_momentum_update(self):
m1 = test_mn.Model()
p1 = m1.add_parameters([2], 'constant', val=0.)
p2 = m1.add_parameters([2, 2], 'constant', val=0.)
# --
t1 = test_mn.MomentumTrainer(m1, lrate=0.1, mrate=0.9)
for _ in range(5):
g = np.array([1., 1.])
p1.accumulate_grad(g)
t1.update()
is_allclose(p1.data, np.array([-0.131441, -0.131441])), \
"This is the values using our implementation, there can be other versions for momentum_update, you can choose to use others!"
for i in range(6):
g = np.array([1., 1.])
p2.accumulate_grad_sparse([((i % 2), g.copy())])
t1.update()
is_allclose(p2.data, np.array([[-0.1002459, -0.1002459], [-0.078051, -0.078051]])), \
"This is the values using our implementation, there can be other versions for momentum_update, you can choose to use others!"
# --

def test_lookup(self):
t = test_mn.astensor([[1., 2.], [3., 4.], [5., 6.]])
v = test_mn.lookup(t, [1, 0, 1])
is_allclose(v.data, np.asarray([[3., 4.], [1., 2.], [3., 4.]]))
v.accumulate_grad(np.asarray([[1., 1.], [1., 1.], [2., 3.]]))
v.op.backward()
is_allclose(t.get_dense_grad(), np.asarray([[1., 1.], [3., 4.], [0., 0.]]))

def test_dot(self):
w, h = test_mn.astensor([[0., 1.], [2., 3.]]), test_mn.astensor([1., 2.])
v = test_mn.dot(w, h)
is_allclose(v.data, np.asarray([2., 8.]))
v.accumulate_grad(np.asarray([1., 3.]))
v.op.backward()
is_allclose(w.get_dense_grad(), np.asarray([[1., 2.], [3., 6.]]))
is_allclose(h.get_dense_grad(), np.asarray([6., 10.]))

def test_tanh(self):
x = test_mn.astensor([0., 1., 2., 3.])
v = test_mn.tanh(x)
is_allclose(v.data, np.asarray([0., 0.76159416, 0.96402758, 0.99505475]))
v.accumulate_grad(np.asarray([1., 2., 3., 4.]))
v.op.backward()
is_allclose(x.get_dense_grad(), np.asarray([1., 0.83994868, 0.21195247, 0.03946415]))
# --

def test_avg(self):
x = test_mn.astensor([[0., 1., 2.], [3., 5., 11.]])
v = test_mn.avg(x, 0)
is_allclose(v.data, np.asarray([1.5, 3., 6.5]))
v.accumulate_grad(np.asarray([1., 2., 3.]))
v.op.backward()
is_allclose(x.get_dense_grad(), np.asarray([[0.5, 1.0, 1.5], [0.5, 1.0, 1.5]]))
# --

def test_max(self):
x = test_mn.astensor([[0., 5., 2.], [-2., 6., 1.]])
v = test_mn.max(x, 0)
is_allclose(v.data, np.asarray([0., 6., 2.]))
v.accumulate_grad(np.asarray([1., 2., 3.]))
v.op.backward()
is_allclose(x.get_dense_grad(), np.asarray([[1., 0., 3.], [0., 2., 0.]]))
# --

# --
# extra one to test acc
def eval_acc(f1: str, f2: str):
with open(f1) as fd:
Expand All @@ -157,7 +133,5 @@ def eval_acc(f1: str, f2: str):
print(f"Acc = {good}/{len(preds1)}={good/len(preds1)}")
# --

if __name__ == '__main__':
# main("minnn.py")
import sys
main(*sys.argv[1:])
if __name__ == "__main__":
unittest.main()