def predict(p):
img_file = "../../imagenetv2-matched-frequency/{}/{}.jpeg".format(
p.cls, p.im)
# print("Predict {}".format(img_file))
image = imread(img_file)
x = center_crop_and_resize(image, image_size=p.image_size)
x = preprocess_input(x)
x = np.expand_dims(x, 0)
# make prediction and decode
y = model.predict(x)
pred_class = np.argmax(y)
# print("Predicted class: {}".format(pred_class))
# v = decode_predictions(y)
# Evaluate accuracy
p.total_imgs += 1.0
p.top_1_correct += (pred_class == p.cls)
# Loop through folders and images
p.im += 1
if p.im == p.NUM_IMAGES:
p.cls = p.cls + 1 if p.cls < p.NUM_CLASSES - 1 else 0
p.im = 0
print("Entering class: {}".format(p.cls))
def extract_feature(model, image_path):
try:
img = pil_loader(image_path)
img = img.resize((256, 256))
img_data = image.img_to_array(img)
except:
img_data = np.zeros((256, 256, 3))
img_data = np.expand_dims(img_data, axis=0)
img_data = preprocess_input(img_data)
feature = model.predict(img_data)
return feature
def generate_dataset_main(n=10000,
save_path=None,
seed=None,
model_res=1024,
image_size=256,
minibatch_size=32,
truncation=0.7):
"""
Generates a dataset of 'n' images of shape ('size', 'size', 3) with random seed 'seed'
along with their dlatent vectors W of shape ('n', 512)
These datasets can serve to train an inverse mapping from X to W as well as explore the latent space
More variation added to latents; also, negative truncation added to balance these examples.
"""
n = n // 2 # this gets doubled because of negative truncation below
model_scale = int(2 *
(math.log(model_res, 2) - 1)) # For example, 1024 -> 18
Gs = load_Gs()
if (model_scale % 3 == 0):
mod_l = 3
else:
mod_l = 2
if seed is not None:
b = bool(np.random.RandomState(seed).randint(2))
Z = np.random.RandomState(seed).randn(n * mod_l, Gs.input_shape[1])
else:
b = bool(np.random.randint(2))
Z = np.random.randn(n * mod_l, Gs.input_shape[1])
if b:
mod_l = model_scale // 2
mod_r = model_scale // mod_l
if seed is not None:
Z = np.random.RandomState(seed).randn(n * mod_l, Gs.input_shape[1])
else:
Z = np.random.randn(n * mod_l, Gs.input_shape[1])
W = Gs.components.mapping.run(
Z, None, minibatch_size=minibatch_size
) # Use mapping network to get unique dlatents for more variation.
dlatent_avg = Gs.get_var('dlatent_avg') # [component]
W = (W[np.newaxis] - dlatent_avg) * np.reshape([truncation, -truncation], [
-1, 1, 1, 1
]) + dlatent_avg # truncation trick and add negative image pair
W = np.append(W[0], W[1], axis=0)
W = W[:, :mod_r]
W = W.reshape((n * 2, model_scale, 512))
X = Gs.components.synthesis.run(W,
randomize_noise=False,
minibatch_size=minibatch_size,
print_progress=True,
output_transform=dict(
func=tflib.convert_images_to_uint8,
nchw_to_nhwc=True))
X = np.array([
cv2.resize(x, (image_size, image_size), interpolation=cv2.INTER_AREA)
for x in X
])
X = preprocess_input(X)
return W, X
def batch_preprocess_input(x_batch, network):
if network == 'Xception':
x_batch = xception.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'VGG16':
x_batch = vgg16.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'VGG19':
x_batch = vgg19.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'ResNet50' or network == 'ResNet101' or network == 'ResNet152':
x_batch = resnet.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'ResNet50V2' or network == 'ResNet101V2' or network == 'ResNet152V2':
x_batch = resnet_v2.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'ResNeXt50' or network == 'ResNeXt101':
x_batch = resnext.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'InceptionV3':
x_batch = inception_v3.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'InceptionResNetV2':
x_batch = inception_resnet_v2.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'MobileNet':
x_batch = mobilenet.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'MobileNetV2':
x_batch = mobilenet_v2.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'DenseNet121' or network == 'DenseNet169' or network == 'DenseNet201':
x_batch = densenet.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif network == 'NASNetMobile' or network == 'NASNetLarge':
x_batch = nasnet.preprocess_input(x_batch,
backend=keras.backend,
layers=keras.layers,
models=keras.models,
utils=keras.utils)
elif 'EfficientNet' in network:
x_batch = efficientnet.preprocess_input(x_batch)
else:
return None
return x_batch
def randaug(imgs):
# for i in range(imgs.shape[0]):
# imgs = distort_image_with_randaugment(imgs.astype('uint8'), 2, 18)
imgs = preprocess_input(imgs)
return imgs
def _get_panda_input(input_shape):
image = imread(PANDA_PATH)
image = efn.center_crop_and_resize(image, input_shape[1])
image = efn.preprocess_input(image)
image = np.expand_dims(image, 0)
return image