create div2k dataset class #547
Conversation
|
|
||
| class Div2k(Base): | ||
| classes = [] | ||
| num_classes = 0 |
There was a problem hiding this comment.
I guess setting num_classes = 0 is for base class restriction.
It's not clear. Please add comment.
| target_file = self.files[i] | ||
| image = load_image(target_file) | ||
|
|
||
| return image, None |
There was a problem hiding this comment.
Same as above. None label looks tricky.
|
@tsawada We need readability reviewer. |
blueoil/generate_lmnet_config.py
Outdated
| @@ -127,7 +131,6 @@ def _blueoil_to_lmnet(blueoil_config): | |||
| } | |||
| dataset = {} | |||
|
|
|||
|
|
|||
There was a problem hiding this comment.
Why do we remove a line here?
There was a problem hiding this comment.
My autopep8 removes unnecessary line break such as double empty line in functions.
Is this line necessary?
There was a problem hiding this comment.
this line does not seem necessary but this diff is not related to the PR and thus should go into another PR.
| files = sorted(dataset.files) | ||
|
|
||
| dataset_files = os.path.join(environment.DATA_DIR, "DIV2K/{}/*.png".format(type["dir"])) | ||
| expected = sorted(glob(dataset_files)) |
There was a problem hiding this comment.
expected should be hard coded
| def test_length(set_test_environment, subset): | ||
| dataset = Div2k(subset) | ||
|
|
||
| assert len(dataset.files) == len(dataset) |
| def test_num_per_epoch(set_test_environment, subset): | ||
| dataset = Div2k(subset) | ||
|
|
||
| assert len(dataset.files) == dataset.num_per_epoch |
There was a problem hiding this comment.
should be hardcoded. Probably we should not use parametrize.
| def test_get_item(set_test_environment, subset): | ||
| dataset = Div2k(subset) | ||
|
|
||
| for i in range(len(dataset)): |
There was a problem hiding this comment.
either
for image, _ in dataset:
assert ...
or
assert all(isinstance(image, np.ndarray) for image, _ in dataset)
There was a problem hiding this comment.
@tsawada oops, thank you!!!!
blueoil/generate_lmnet_config.py
Outdated
| @@ -127,7 +131,6 @@ def _blueoil_to_lmnet(blueoil_config): | |||
| } | |||
| dataset = {} | |||
|
|
|||
|
|
|||
There was a problem hiding this comment.
this line does not seem necessary but this diff is not related to the PR and thus should go into another PR.
| def test_get_item(set_test_environment): | ||
| dataset = Div2k("train") | ||
|
|
||
| assert all([isinstance(image, np.ndarray) for image, _ in dataset]) |
There was a problem hiding this comment.
no need to use list ([]) here. Generator comprehension is better as it'll fail early.
There was a problem hiding this comment.
I did not know that!
thank you!!
|
|
||
|
|
||
| def test_train_files(set_test_environment): | ||
| dataset = Div2k("train") |
There was a problem hiding this comment.
I think we don't need to bind everything into local variables.
assert sorted(Div2k("train")) == expected is short and easy to understand.
Same for other tests.
|
ARIGATO GOZAMASU! |
Motivation and Context
I want to implement a super-resolution network named DCSCN.
DNSCN uses the DIV2K dataset for train.
Description
How has this been tested?
Screenshots (if appropriate):
Types of changes
Checklist: