Speed up image cropping and resizing

README.md

@@ -92,7 +92,7 @@ The performance of each model variant using the pre-trained weights converted fr
 
 * `Keras >= 2.2.0` / `TensorFlow >= 1.12.0`
 * `keras_applications >= 1.0.7`
-* `scikit-image`
+* `opencv >= 3.4.2` or `scikit-image` (for resizing images; `opencv` seems to be faster)
 
 ### Installing from the source
 

efficientnet/preprocessing.py

@@ -13,37 +13,75 @@
 # limitations under the License.
 # ==============================================================================
 import numpy as np
-from skimage.transform import resize
+try:
+    import cv2
+except ModuleNotFoundError:
+    import skimage.transform
 
-MAP_INTERPOLATION_TO_ORDER = {
-    "nearest": 0,
-    "bilinear": 1,
-    "biquadratic": 2,
-    "bicubic": 3,
-}
+
+try:
+    # OpenCV: Map interpolation string to OpenCV interpolation enum value
+    INTERPOLATION_DICT = {
+        "nearest": cv2.INTER_NEAREST,
+        "bilinear": cv2.INTER_LINEAR,
+        "bicubic": cv2.INTER_CUBIC,
+        "laconzos": cv2.INTER_LANCZOS4,
+    }
+except NameError:
+    # scikit-image: Map interpolation string to interpolation order
+    INTERPOLATION_DICT = {
+        "nearest": 0,
+        "bilinear": 1,
+        "biquadratic": 2,
+        "bicubic": 3,
+    }
 
 
 def center_crop_and_resize(image, image_size, crop_padding=32, interpolation="bicubic"):
     assert image.ndim in {2, 3}
-    assert interpolation in MAP_INTERPOLATION_TO_ORDER.keys()
+    assert interpolation in INTERPOLATION_DICT.keys()
+
+    in_h, in_w = image.shape[:2]
+
+    if isinstance(image_size, (int, float)):
+        out_h = out_w = image_size
+    else:
+        out_h, out_w = image_size
+
+    if isinstance(crop_padding, (int, float)):
+        crop_padding_h = crop_padding_w = crop_padding
+    else:
+        crop_padding_h, crop_padding_w = crop_padding
 
-    h, w = image.shape[:2]
+    padded_center_crop_shape_post_scaling = (out_h + crop_padding_h,
+                                             out_w + crop_padding_w)
 
-    padded_center_crop_size = int(
-        (image_size / (image_size + crop_padding)) * min(h, w)
-    )
-    offset_height = ((h - padded_center_crop_size) + 1) // 2
-    offset_width = ((w - padded_center_crop_size) + 1) // 2
+    inv_scale = min(in_h / padded_center_crop_shape_post_scaling[0],
+                in_w / padded_center_crop_shape_post_scaling[1])
+
+    unpadded_center_crop_size_pre_scaling = (round(out_h * inv_scale),
+                                             round(out_w * inv_scale))
+
+    offset_h = ((in_h - unpadded_center_crop_size_pre_scaling[0]) + 1) // 2
+    offset_w = ((in_w - unpadded_center_crop_size_pre_scaling[1]) + 1) // 2
 
     image_crop = image[
-                 offset_height: padded_center_crop_size + offset_height,
-                 offset_width: padded_center_crop_size + offset_width,
+                 offset_h : unpadded_center_crop_size_pre_scaling[0] + offset_h,
+                 offset_w : unpadded_center_crop_size_pre_scaling[1] + offset_w,
                  ]
-    resized_image = resize(
-        image_crop,
-        (image_size, image_size),
-        order=MAP_INTERPOLATION_TO_ORDER[interpolation],
-        preserve_range=True,
-    )
+
+    try:
+        resized_image = cv2.resize(
+            image_crop,
+            (out_w, out_h),
+            interpolation=INTERPOLATION_DICT[interpolation] if inv_scale < 1 else cv2.INTER_AREA,
+        )
+    except NameError:
+        resized_image = skimage.transform.resize(
+            image_crop,
+            (out_h, out_w),
+            order=INTERPOLATION_DICT[interpolation],
+            preserve_range=True,
+        )
 
     return resized_image