def process_batch(images, models):
def zoom(x, z):
original_size = x.size[0]
target_size = int(x.size[0] * z)
x = x.resize((target_size, target_size), resample=Image.BILINEAR)
step = abs(target_size - original_size) // 2
if z > 1:
return x.crop(
(step, step, step + original_size, step + original_size))
elif z < 1:
new_image = Image.fromarray(np.zeros(
(original_size, original_size, 3), dtype=np.uint8),
mode="RGB")
new_image.paste(x, (step, step))
return new_image
else:
return x
angles = np.random.random(N_SAMPLES) * 1
zooms = np.random.random(N_SAMPLES) * 0.1 + 1
contrast = np.random.uniform(0.7, 1.3, size=N_SAMPLES)
color = np.random.uniform(0.7, 1.3, N_SAMPLES)
sharpness = np.random.uniform(0.6, 1.0, N_SAMPLES)
preds = []
for angle, zm, cont, colr, shrp in zip(angles, zooms, contrast, color,
sharpness):
batch = images.copy()
batch = [zoom(x, zm) for x in batch]
batch = [
x.rotate(angle, resample=Image.BILINEAR, expand=False)
for x in batch
]
batch = [ImageEnhance.Contrast(x).enhance(cont) for x in batch]
batch = [ImageEnhance.Color(x).enhance(colr) for x in batch]
batch = [ImageEnhance.Sharpness(x).enhance(shrp) for x in batch]
for model in models:
batch_ = batch.copy()
if model.name in {"resnet50", "vgg16"}:
batch_ = [
x.resize((SIDE_VGG, SIDE_VGG), resample=Image.BILINEAR)
for x in batch_
]
batch_ = np.stack([np.array(x) for x in batch_]).astype(K.floatx())
if model.name in {"resnet50", "vgg16"}:
batch_ = preprocess_vgg_resnet(batch_)
else:
batch_ = preprocess_inception(batch_)
if model.name == "inception_v3":
preds.append(model.predict(batch_)
[:, 1:]) # background class in adv_inception
else:
preds.append(model.predict(batch_))
preds = np.stack(preds).mean(axis=0)
labels = np.argmax(preds, axis=1) + 1 # add background class
return labels
for filename in classes:
# load image
path = ROOT_DIR + 'fruit-recognition_reduced/' + filename
list_images = os.listdir(path)[:100]
print(f'{filename}: {len(list_images)}')
for img in list_images:
# img_data = image.imread(ROOT_DIR + 'fruit-recognition_reduced/' + filename + '/' + img)
img_data = load_img(ROOT_DIR + 'fruit-recognition_reduced/' + filename + '/' + img, target_size=(224,224))
# img_data = img_data.resize((224, 224), Image.ANTIALIAS)
img_data = img_to_array(img_data)
# Scale images to the range [0,1]
# img_data = img_data.astype(np.float32)/255.0
# img_data = img_data.astype(np.float32) - [123.68, 116.779, 103.939]
# img_data = preprocess_vgg(img_data.astype(np.float32))
img_data = preprocess_inception(img_data.astype(np.float32))
# img_data = preprocess_resnet(img_data.astype(np.float32))
# Remove noise
# img_data = gaussian(img_data, sigma=1, multichannel=True)
# store loaded image
X.append(img_data)
y.append(c)
c += 1
X = np.array(X)
y = np.array(y)
print(set(y))
# Compute class weights
from sklearn.utils import class_weight
class_weights = class_weight.compute_class_weight('balanced', np.unique(y), y)
def predict(self, x):
if self.data_format == "channels_first":
x = x.transpose(0, 3, 1, 2)
x = preprocess_inception(x.astype(K.floatx()))
return self.model.predict(x, batch_size=self.batch_size)
x2 = GlobalAveragePooling2D()(model.get_layer("block2_conv2").output) # 128
x3 = GlobalAveragePooling2D()(model.get_layer("block3_conv3").output) # 256
x4 = GlobalAveragePooling2D()(model.get_layer("block4_conv3").output) # 512
x5 = GlobalAveragePooling2D()(model.get_layer("block5_conv3").output) # 512
x = Concatenate()([x2, x3, x4, x5])
self.model = Model(inputs=model.input, outputs=x)
self.batch_size = batch_size
self.data_format = K.image_data_format()
def predict(self, x):
if self.data_format == "channels_first":
x = x.transpose(0, 3, 1, 2)
x = preprocess_vgg(x.astype(K.floatx()))
return self.model.predict(x, batch_size=self.batch_size)
if __name__ == "__main__":
img_path = "data/elephant.jpg"
img = image.load_img(img_path, target_size=(299, 299))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_inception(x)
model = InceptionV3()
preds = model.predict(x)
# decode the results into a list of tuples (class, description, probability)
# (one such list for each sample in the batch)
print("Predicted:", decode_predictions(preds, top=3)[0])
# Predicted: [('n02504013', 'Indian_elephant', 0.78864819), ('n01871265', 'tusker', 0.029346621), ('n02504458', 'African_elephant', 0.01768155)]