Refactor preprocess_input (#49)

* Use keras-applications preprocessing

Author: qubvel

efficientnet/keras.py

@@ -1,7 +1,7 @@
 from . import inject_keras_modules, init_keras_custom_objects
 from . import model
 
-from .preprocessing import center_crop_and_resize, preprocess_input
+from .preprocessing import center_crop_and_resize
 
 EfficientNetB0 = inject_keras_modules(model.EfficientNetB0)
 EfficientNetB1 = inject_keras_modules(model.EfficientNetB1)
@@ -12,4 +12,6 @@
 EfficientNetB6 = inject_keras_modules(model.EfficientNetB6)
 EfficientNetB7 = inject_keras_modules(model.EfficientNetB7)
 
+preprocess_input = inject_keras_modules(model.preprocess_input)
+
 init_keras_custom_objects()

efficientnet/model.py

@@ -36,9 +36,9 @@
 from six.moves import xrange
 from keras_applications.imagenet_utils import _obtain_input_shape
 from keras_applications.imagenet_utils import decode_predictions
+from keras_applications.imagenet_utils import preprocess_input as _preprocess_input
 
 from . import get_submodules_from_kwargs
-from .preprocessing import preprocess_input
 
 backend = None
 layers = None
@@ -76,9 +76,6 @@
                         'da76b3a29e1c011635376e191c2c2d54')
 }
 
-MEAN_RGB = [0.485 * 255, 0.456 * 255, 0.406 * 255]
-STDDEV_RGB = [0.229 * 255, 0.224 * 255, 0.225 * 255]
-
 BlockArgs = collections.namedtuple('BlockArgs', [
     'kernel_size', 'num_repeat', 'input_filters', 'output_filters',
     'expand_ratio', 'id_skip', 'strides', 'se_ratio'
@@ -127,6 +124,10 @@
 }
 
 
+def preprocess_input(x, **kwargs):
+    return _preprocess_input(x, mode='torch', **kwargs)
+
+
 def swish(x):
     """Swish activation function: x * sigmoid(x).
     Reference: [Searching for Activation Functions](https://arxiv.org/abs/1710.05941)
@@ -227,7 +228,7 @@ def mb_conv_block(inputs, block_args, drop_rate=None, relu_fn=swish, prefix='',
             block_args.input_filters * block_args.se_ratio
         ))
         se_tensor = layers.GlobalAveragePooling2D(name=prefix + 'se_squeeze')(x)
-        
+
         target_shape = (1, 1, filters) if backend.image_data_format() == 'channels_last' else (filters, 1, 1)
         se_tensor = layers.Reshape(target_shape, name=prefix + 'se_reshape')(se_tensor)
         se_tensor = layers.Conv2D(num_reduced_filters, 1,

efficientnet/preprocessing.py

@@ -15,9 +15,6 @@
 import numpy as np
 from skimage.transform import resize
 
-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,
@@ -39,9 +36,9 @@ def center_crop_and_resize(image, image_size, crop_padding=32, interpolation="bi
     offset_width = ((w - padded_center_crop_size) + 1) // 2
 
     image_crop = image[
-        offset_height : padded_center_crop_size + offset_height,
-        offset_width : padded_center_crop_size + offset_width,
-    ]
+                 offset_height: padded_center_crop_size + offset_height,
+                 offset_width: padded_center_crop_size + offset_width,
+                 ]
     resized_image = resize(
         image_crop,
         (image_size, image_size),
@@ -50,13 +47,3 @@ def center_crop_and_resize(image, image_size, crop_padding=32, interpolation="bi
     )
 
     return resized_image
-
-
-def preprocess_input(x, **kwargs):
-    assert x.ndim in (3, 4)
-    assert x.shape[-1] == 3
-
-    x = x - np.array(MEAN_RGB)
-    x = x / np.array(STDDEV_RGB)
-
-    return x

efficientnet/tfkeras.py

@@ -1,7 +1,7 @@
 from . import inject_tfkeras_modules, init_tfkeras_custom_objects
 from . import model
 
-from .preprocessing import center_crop_and_resize, preprocess_input
+from .preprocessing import center_crop_and_resize
 
 EfficientNetB0 = inject_tfkeras_modules(model.EfficientNetB0)
 EfficientNetB1 = inject_tfkeras_modules(model.EfficientNetB1)
@@ -12,4 +12,6 @@
 EfficientNetB6 = inject_tfkeras_modules(model.EfficientNetB6)
 EfficientNetB7 = inject_tfkeras_modules(model.EfficientNetB7)
 
+preprocess_input = inject_tfkeras_modules(model.preprocess_input)
+
 init_tfkeras_custom_objects()