stu1130 commented on a change in pull request #12285: [WIP] Change the way
NDArrayIter handle the last batch
URL: https://github.com/apache/incubator-mxnet/pull/12285#discussion_r212102352
##########
File path: tests/python/unittest/test_io.py
##########
@@ -87,82 +87,68 @@ def test_Cifar10Rec():
for i in range(10):
assert(labelcount[i] == 5000)
-
-def test_NDArrayIter():
+def _init_NDArrayIter_data():
data = np.ones([1000, 2, 2])
- label = np.ones([1000, 1])
+ labels = np.ones([1000, 1])
for i in range(1000):
data[i] = i / 100
- label[i] = i / 100
- dataiter = mx.io.NDArrayIter(
- data, label, 128, True, last_batch_handle='pad')
- batchidx = 0
- for batch in dataiter:
- batchidx += 1
- assert(batchidx == 8)
- dataiter = mx.io.NDArrayIter(
- data, label, 128, False, last_batch_handle='pad')
- batchidx = 0
- labelcount = [0 for i in range(10)]
- for batch in dataiter:
- label = batch.label[0].asnumpy().flatten()
- assert((batch.data[0].asnumpy()[:, 0, 0] == label).all())
- for i in range(label.shape[0]):
- labelcount[int(label[i])] += 1
+ labels[i] = i / 100
+ return data, labels
- for i in range(10):
- if i == 0:
- assert(labelcount[i] == 124)
- else:
- assert(labelcount[i] == 100)
+def _test_last_batch_handle(data, labels):
+ idx = 0
+ last_batch_handle_list = ['pad', 'discard' , 'roll_over']
+ labelcount_list = [(124, 100), (100, 96), (100, 96)]
+ batch_count_list = [8, 7, 7]
+
+ for idx in range(len(last_batch_handle_list)):
+ dataiter = mx.io.NDArrayIter(
+ data, labels, 128, False,
last_batch_handle=last_batch_handle_list[idx])
+ batch_count = 0
+ labelcount = [0 for i in range(10)]
+ for batch in dataiter:
+ label = batch.label[0].asnumpy().flatten()
+ assert((batch.data[0].asnumpy()[:, 0, 0] == label).all()),
last_batch_handle_list[idx]
+ for i in range(label.shape[0]):
+ labelcount[int(label[i])] += 1
+ batch_count += 1
+ # assert result
+ assert(labelcount[0] == labelcount_list[idx][0]),
last_batch_handle_list[idx]
+ assert(labelcount[8] == labelcount_list[idx][1]),
last_batch_handle_list[idx]
+
+ assert batch_count == batch_count_list[idx]
+ # shuffle equals True for sanity test
+ dataiter = mx.io.NDArrayIter(
+ data, labels, 128, True,
last_batch_handle=last_batch_handle_list[idx])
+ batch_count = 0
+ for _ in dataiter:
+ batch_count += 1
+ assert batch_count == batch_count_list[idx]
+def test_NDArrayIter():
+ data, labels = _init_NDArrayIter_data()
+ _test_last_batch_handle(data, labels)
def test_NDArrayIter_h5py():
if not h5py:
return
- data = np.ones([1000, 2, 2])
- label = np.ones([1000, 1])
- for i in range(1000):
- data[i] = i / 100
- label[i] = i / 100
+ data, labels = _init_NDArrayIter_data()
try:
- os.remove("ndarraytest.h5")
+ os.remove('ndarraytest.h5')
except OSError:
pass
- with h5py.File("ndarraytest.h5") as f:
- f.create_dataset("data", data=data)
- f.create_dataset("label", data=label)
-
- dataiter = mx.io.NDArrayIter(
- f["data"], f["label"], 128, True, last_batch_handle='pad')
- batchidx = 0
- for batch in dataiter:
- batchidx += 1
- assert(batchidx == 8)
-
- dataiter = mx.io.NDArrayIter(
- f["data"], f["label"], 128, False, last_batch_handle='pad')
- labelcount = [0 for i in range(10)]
- for batch in dataiter:
- label = batch.label[0].asnumpy().flatten()
- assert((batch.data[0].asnumpy()[:, 0, 0] == label).all())
- for i in range(label.shape[0]):
- labelcount[int(label[i])] += 1
+ with h5py.File('ndarraytest.h5') as f:
Review comment:
Good Catch. will implement that
----------------------------------------------------------------
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:
[email protected]
With regards,
Apache Git Services
