Change from skimage.transform.resize to cv2.resize

To make the process of cropping and resizing images go faster, switch
from the resize function in skimage.transform to the resize function
in cv2.

Author: krikru

README.md

@@ -79,7 +79,7 @@ The performance of each model variant using the pre-trained weights converted fr
 ### Requirements
 
 * `keras >= 2.2.0` + `tensorflow`
-* `scikit-image`
+* `opencv >= 3.4.2`
 
 ### Installing from the source
 

efficientnet/preprocessing.py

@@ -13,16 +13,16 @@
 # limitations under the License.
 # ==============================================================================
 import numpy as np
-from skimage.transform import resize
+import cv2
 
 MEAN_RGB = [0.485 * 255, 0.456 * 255, 0.406 * 255]
 STDDEV_RGB = [0.229 * 255, 0.224 * 255, 0.225 * 255]
 
 MAP_INTERPOLATION_TO_ORDER = {
-    "nearest": 0,
-    "bilinear": 1,
-    "biquadratic": 2,
-    "bicubic": 3,
+    "nearest": cv2.INTER_NEAREST,
+    "bilinear": cv2.INTER_LINEAR,
+    "bicubic": cv2.INTER_CUBIC,
+    "laconzos": cv2.INTER_LANCZOS4
 }
 
 
@@ -59,11 +59,10 @@ def center_crop_and_resize(image, image_size, crop_padding=32, interpolation="bi
         offset_w : unpadded_center_crop_size_pre_scaling[1] + offset_w,
     ]
 
-    resized_image = resize(
+    resized_image = cv2.resize(
         image_crop,
-        (out_h, out_w),
-        order=MAP_INTERPOLATION_TO_ORDER[interpolation],
-        preserve_range=True,
+        (out_w, out_h),
+        interpolation=MAP_INTERPOLATION_TO_ORDER[interpolation] if inv_scale < 1 else cv2.INTER_AREA
     )
 
     return resized_image