Add option to specify resampling method in load_img #7975
Conversation
keras/preprocessing/image.py
Outdated
| of `PIL.Image.NEAREST`, `PIL.Image.BOX`, `PIL.Image.BILINEAR`, | ||
| `PIL.Image.HAMMING`, `PIL.Image.BICUBIC` or `PIL.Image.LANCZOS`. If | ||
| omitted, or if the image has mode "1" or "P", | ||
| it is set `PIL.Image.NEAREST`. |
There was a problem hiding this comment.
Looks like it'll be PIL.Image.BILINEAR if omitted?
|
You are right, of course. Fixed now. |
|
For the consistency with the broader API, it should be specified as a lowercase string, rather than as a constant from |
|
@fchollet I addressed your comment. |
keras/preprocessing/image.py
Outdated
| @@ -302,14 +314,19 @@ def img_to_array(img, data_format=None): | |||
| return x | |||
|
|
|||
|
|
|||
| def load_img(path, grayscale=False, target_size=None): | |||
| def load_img(path, grayscale=False, target_size=None, | |||
| interpolation="bilinear"): | |||
There was a problem hiding this comment.
Use ' as quote char for consistency
keras/preprocessing/image.py
Outdated
| """Loads an image into PIL format. | ||
|
|
||
| # Arguments | ||
| path: Path to image file | ||
| grayscale: Boolean, whether to load the image as grayscale. | ||
| target_size: Either `None` (default to original size) | ||
| or tuple of ints `(img_height, img_width)`. | ||
| interpolation: Interpolation method used to resample the image if the | ||
| target size does not equal the loaded image. Supported methods are |
There was a problem hiding this comment.
"is different from that of the loaded image"
keras/preprocessing/image.py
Outdated
| @@ -330,7 +347,10 @@ def load_img(path, grayscale=False, target_size=None): | |||
| if target_size: | |||
| hw_tuple = (target_size[1], target_size[0]) | |||
| if img.size != hw_tuple: | |||
| img = img.resize(hw_tuple) | |||
| if interpolation not in _PIL_INTERPOLATION_METHODS: | |||
| raise ValueError('Invalid interpolation method specified') | |||
There was a problem hiding this comment.
The ValueError should be listed in the docstring. The error message should specify what was passed, and the list of values expected instead.
|
@fchollet I addressed your comments. In addition, I added a more fine-grain check for the available interpolation methods. Only newer pillow versions support "box" and "hamming". |
keras/preprocessing/image.py
Outdated
| 'bicubic': pil_image.BICUBIC, | ||
| } | ||
| # These methods were only introduced in version 3.4.0 (2016). | ||
| if hasattr(PIL.Image, "HAMMING"): |
There was a problem hiding this comment.
Please use consistent quote chars. If you know at which PIL version these attributes were introduce, prefer doing a version number check instead of repeated hasattr calls.
There was a problem hiding this comment.
Due to the various forks of PIL and hence different versioning schemes, I'd say it is safer to use this explicit test in this case.
keras/preprocessing/image.py
Outdated
| img = img.resize(hw_tuple) | ||
| width_height_tuple = (target_size[1], target_size[0]) | ||
| if img.size != width_height_tuple: | ||
| if interpolation not in _PIL_INTERPOLATION_METHODS: |
There was a problem hiding this comment.
Please add a unit test for this feature, include a check for the ValueError.
There was a problem hiding this comment.
Are there any test images shipped with keras that could be used for this test? Otherwise, it is not possible to get to this point without a failure before getting to this point.
There was a problem hiding this comment.
We haves test code that creates test images and write them in a temporary folder. You can check out existing tests for preprocessing/image.
|
Re-running tests. |
|
Re-running tests, again. |
|
@fchollet Added unit tests to test the correct behavior of the interpolation and the |
Otherwise, nearest neighbor interpolation is used, which produces awful results. I consider this as a bug. In any case, the old behavior can be recovered using
load_img(..., resample=PIL.Image.NEAREST).