def start():
model = keras.applications.vgg19.VGG19(include_top=False, weights='imagenet', pooling='avg')
root_dir = "/media/rishabh/dump_bin/Animals_with_Attributes2/JPEGImages/"
for root, subdirs, files in os.walk(root_dir):
list_file_path = os.path.join(root, 'list_of_files.txt')
with open(list_file_path, 'wb') as list_file:
for filename in files:
if filename.endswith("jpg"):
file_path = os.path.join(root, filename)
img = image.load_img(file_path, target_size=(224,224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
features = model.predict(x)
np_name = filename[0:-4]
np_name = np_name+".npy"
np.save(os.path.join(root,np_name), features)
# npy = open(os.path.join(root,np_name),"w+")
print('file %s (full path: %s)' % (filename, file_path))
list_file.write(('%s\n' % filename).encode('utf-8'))
def preprocess_xray_flipped(xray_path):
xray = image.load_img(xray_path, color_mode="grayscale", vertical_flip=True,
target_size=(img_dims[0], img_dims[1], 1))
xray = image.img_to_array(xray)
xray = np.dstack([xray, xray, xray])
xray = preprocess_input(xray)
return xray
def classifyImage(fname):
img = image.load_img(fname, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
pred = decode_predictions(model.predict(x), top=1)[0][0]
return (pred[1], pred[2])
def predict(model, img, target_size, top_n=3): """Run model prediction on image """ if img.size != target_size: img = img.resize(target_size) x = image.img_to_array(img) x = np.expand_dims(x, axis=0) x = preprocess_input(x) preds = model.predict(x) return decode_predictions(preds, top=top_n)[0]
def predict(img):
print('{}:'.format(img))
# Load the image file, resizing it to 224x224 pixels (required by this model)
img_df = image.load_img(img, target_size=(224, 224))
x = image.img_to_array(img_df) # [row[columns[r, g, b]]]
x = np.expand_dims(x, axis=0) # array of images
# Scale and predict
predictions = model.predict(resnet50.preprocess_input(x))
predicted_classes = resnet50.decode_predictions(predictions, top=5)
for imagenet_id, name, likelihood in predicted_classes[0]:
print(' - {}: {:2f} likelihood'.format(name, likelihood))
print('\n')
def __read_all_images(src):
files = listdir(src)
images = {}
i = 0
total = len(files)
for f in files:
if not (f.endswith(".jpg")):
continue
im = Image.open(src + f)
im = img_to_array(im)
im = preprocess_input(im)
images[f[:-4]] = im
if i % 100 == 0:
print(str(i) + " / " + str(total))
i += 1
return images
def read_image(i):
rPath = image_filenames[i]
dPath = depth_filenames[i]
print(i)
image = misc.imread(rPath)/1.
image_resized = resize(image, output_shape=(224, 224))
depth = misc.imread(dPath).astype(np.uint16)/1000.0
depth_resized = resize(depth, output_shape=(224, 224)) # (480,640) -> Model Output (224, 224)
print(image_resized.shape)
print(np.expand_dims(depth_resized,-1).shape)
input("aki")
return preprocess_input(image_resized), np.expand_dims(depth_resized,-1)
def _f():
start = 0
end = start + batch_size
n = data.shape[0]
while True:
X_batch = session.run(resize_op, {img_placeholder: data[start:end]})
X_batch = preprocess_input(X_batch)
y_batch = labels[start:end]
start += batch_size
end += batch_size
if start >= n:
start = 0
end = batch_size
print(start, end)
yield (X_batch, y_batch)
def extract_feature(dir_path, net):
features = []
infos = []
num = 0
for image_name in os.listdir(dir_path):
arr = image_name.split('_')
person = int(arr[0])
camera = int(arr[1][1])
image_path = os.path.join(dir_path, image_name)
img = image.load_img(image_path, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
feature = net.predict(x)
features.append(np.squeeze(feature))
infos.append((person, camera))
return features, infos
def predict(model, img, target_size, top_n=3):
"""Run model prediction on image
Args:
model: keras model
img: PIL format image
target_size: (w,h) tuple
top_n: # of top predictions to return
Returns:
list of predicted labels and their probabilities
"""
if img.size != target_size:
img = img.resize(target_size)
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
return decode_predictions(preds, top=top_n)[0]
def pix2depth(path, model):
model_name = 'p2d'
originalImage = cv2.imread(path)
loaded_model = model_list['pix2depth'][model]
file_name = model+'_'+path.split('/')[-1]
output_file = os.path.join(output_path,file_name)
if model =='CNN':
originalImage = cv2.resize(originalImage,(img_dim,img_dim))
x = preprocess_input(originalImage/1.)
elif model == 'CycleGAN':
test(path)
os.system('cp gautam/inf_results/imgs/fakeA_0_0.jpg %s' % output_file)
else:
originalImage = cv2.resize(originalImage,(256,256))
x = originalImage/255.
if not model == 'CycleGAN':
p1 = get_depth_map(x, loaded_model)
cv2.imwrite(output_file,p1)
return output_file
def load_image(img_path):
data = image.load_img(img_path, target_size=(224, 224))
x = image.img_to_array(data)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
return x
######################################################################
# Load a test image
# ------------------
# A single cat dominates the examples!
from PIL import Image
from matplotlib import pyplot as plt
from keras.applications.resnet50 import preprocess_input
img_url = 'https://github.com/dmlc/mxnet.js/blob/master/data/cat.png?raw=true'
download(img_url, 'cat.png')
img = Image.open('cat.png').resize((224, 224))
plt.imshow(img)
plt.show()
# input preprocess
data = np.array(img)[np.newaxis, :].astype('float32')
data = preprocess_input(data).transpose([0, 3, 1, 2])
print('input_1', data.shape)
######################################################################
# Compile the model on NNVM
# --------------------------
# We should be familiar with the process now.
# convert the keras model(NHWC layout) to NNVM format(NCHW layout).
sym, params = nnvm.frontend.from_keras(keras_resnet50)
# compile the model
target = 'cuda'
shape_dict = {'input_1': data.shape}
with nnvm.compiler.build_config(opt_level=3):
graph, lib, params = nnvm.compiler.build(sym, target, shape_dict, params=params)
bounds=(0, 255),
preprocessing=preprocessing)
img_folderpath = "./sample_images/"
img_path = img_folderpath + 'sample_image_2.jpg' # An image of a yellow cab/taxi
x = image.load_img(img_path,
color_mode='rgb',
target_size=(img_rows, img_cols))
img = image.img_to_array(x)
img = np.expand_dims(img, axis=0)
img = img.reshape(img_shape)
label = 468 # For the class of taxi and cab
# Note that proprocess_input is an in-place operation.
prediction = kmodel.predict(
preprocess_input(np.copy(img)).reshape(
(1, img_rows, img_cols, nb_channels)))
# decode the results into a list of tuples (class, description, probability)
# (one such list for each sample in the batch)
print('Prediction on the original example:',
decode_predictions(prediction, top=3)[0])
# The original image is correctly classified as a cab with the confidence of 0.999.
# apply attack on source image
# ::-1 reverses the color channels, because Keras ResNet50 expects BGR instead of RGB
attack = foolbox.attacks.FGSM(fmodel)
adversarial = attack(img[:, :, ::-1], label)[:, :, ::-1]
# if the attack fails, adversarial will be None and a warning will be printed
adversarial_prediction = kmodel.predict(
adversarial.reshape((1, img_rows, img_cols, nb_channels)))
print('Prediction on the adversarial example:',
def preprocess_image(image_path):
img = load_img(image_path, target_size=(img_nrows, img_ncols))
img = img_to_array(img)
img = np.expand_dims(img, axis=0)
img = resnet50.preprocess_input(img)
return img
def extract_InceptionV3(tensor):
from keras.applications.inception_v3 import InceptionV3, preprocess_input
return InceptionV3(weights='imagenet',
include_top=False).predict(preprocess_input(tensor))
#%% # prepare the image for the VGG model processed_image = vgg16.preprocess_input(image_batch.copy()) # get the predicted probabilities for each class predictions = vgg_model.predict(processed_image) #print(predictions) # convert the probabilities to class labels # We will get top 5 predictions which is the default label = decode_predictions(predictions) print(label) #%% # prepare the image for the Resnet model processed_image = resnet50.preprocess_input(image_batch.copy()) # get the predicted probabilities for each class predictions = resnet_model.predict(processed_image) #print(predictions) # convert the probabilities to class labels # We will get top 5 predictions which is the default label = decode_predictions(predictions) print(label) #%% # prepare the image for the Resnet model processed_image = mobilenet.preprocess_input(image_batch.copy()) # get the predicted probabilities for each class
import numpy as np
from keras.preprocessing import image
from keras.applications import resnet50
model = resnet50.ResNet50()
img = image.load_img("image.jpg", target_size = (224,224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = resnet50.preprocess_input(x)
predictions = model.predict(x)
predicted_classes = resnet50.decode_predictions(predictions, top=9)
print("This is an image of : ")
for imagenet_id, name, likelihood in predicted_classes[0]:
print(" - {}: {:2f} likelihood".format(name, likelihood))
from keras.preprocessing.image import load_img from keras.preprocessing.image import img_to_array from keras.applications import resnet50 from keras.applications.imagenet_utils import decode_predictions import keras as keras import numpy as np model = keras.applications.resnet50.ResNet50(weights="imagenet") path = "../input/starfish/asd.jpg" # load an image in PIL format original_image = load_img(path, target_size=(224, 224)) numpy_image = img_to_array(original_image) # Convert the image into 4D Tensor (samples, height, width, channels) by adding an extra dimension to the axis 0. input_image = np.expand_dims(numpy_image, axis=0) # preprocess for resnet50 processed_image_resnet50 = resnet50.preprocess_input(input_image.copy()) # resnet50 predictions_resnet50 = model.predict(processed_image_resnet50) label_resnet50 = decode_predictions(predictions_resnet50) print(label_resnet50)
def predict_dog_label():
"""Predicts the dog breed and recommends similar dogs based on its characterictics"""
url = input("Enter img url :")
image_to_tensor = np.zeros(((1), 224, 224, 3), dtype='float32')
# converts image to centralized format for ResNet model
image_to_tensor[0] = preprocess_input(
np.expand_dims(read_img(url).copy(), axis=0))
# predicts bottleneck features using ResNet model
get_image_features = model_RN50.predict(image_to_tensor, verbose=1)
# predicts image label based on features from ResNet model
pred = ((model_2cr.predict(get_image_features) +
model_2dr.predict(get_image_features) +
model_2gr.predict(get_image_features) +
model_2jr.predict(get_image_features) +
model_2lr.predict(get_image_features)) / 5)
# returns top 5 predicted labels
top_5 = [
i[0] for i in sorted(list(zip(list(y_table.columns),
pred.tolist()[0])),
key=lambda x: x[1],
reverse=True)[:6]
]
label = list(y_table.columns)[int(pred.argmax(axis=1))]
show_image_pred(url, pred.argmax(axis=1))
# finds most similar picture in local database based on cosine similarity of image features
similar_pic = sorted(list(
zip(list(train_breeds_names[0]),
(cosine_similarity(get_image_features,
pd.DataFrame(train_breeds_features))).reshape(
-1, 1))),
key=lambda x: x[1],
reverse=True)[1][0]
time.sleep(0.5)
sim = input('\nDo you want the most similar pic in database: (y/n) ')
if sim == 'y':
print('\n Most similar pic in database: \n')
show_image(similar_pic)
time.sleep(0.5)
print('\nTop 5 visually most similar dogs:', top_5[1:])
show = input('\nDo you want to see those dogs?: (y/n) ')
if show == 'y':
show_rand_image(top_5[1:])
feat = input("\nDo you want to see the features of your dog? (y,n) ")
if feat == 'y':
if dog_features.loc[dog_features['label_name'] ==
label].values[0][1:-1][0] == 0:
print('\nThis dog is not a pet or no information available\n')
feat_vis = input(
'\nDo you want to see the features of visually similar dogs? (y/n) '
)
if feat_vis == 'y':
find_features_table(top_5[1:])
return label
find_features_table([label])
feat_sim = input(
'\nDo you want to find similar dogs based on its features? (y/n) ')
if feat_sim == 'y':
print('\n')
# finds most similar dog based on cosine similarity of dog characteristcs
top_5_feat = [
x[0] for x in sorted(list(
zip(list(dog_features.label_name), (
cosine_similarity(dog_features[dog_features.columns[1:-1]])
[int(pred.argmax(axis=1))]))),
key=lambda x: x[1],
reverse=True)[1:6]
]
find_features_table(top_5_feat)
feat_sim2 = input("\nDo you want to see them? (y,n) ")
if feat_sim2 == 'y':
show_rand_image(top_5_feat)
return label
def load_image(img_path):
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_input(x)
return x
resnet.summary()
# make a model to get output before flatten
activation_layer = resnet.get_layer('activation_49')
# create a model object
model = Model(inputs=resnet.input, outputs=activation_layer.output)
# get the feature map weights
final_dense = resnet.get_layer('fc1000')
W = final_dense.get_weights()[0]
while True:
img = image.load_img(np.random.choice(image_files), target_size=(224, 224))
x = preprocess_input(np.expand_dims(img, 0))
fmaps = model.predict(x)[0] # 7 x 7 x 2048
# get predicted class
probs = resnet.predict(x)
classnames = decode_predictions(probs)[0]
print(classnames)
classname = classnames[0][1]
pred = np.argmax(probs[0])
# get the 2048 weights for the relevant class
w = W[:, pred]
# "dot" w with fmaps
cam = fmaps.dot(w)
import random
from sklearn.model_selection import train_test_split
from keras.preprocessing import image
from keras.applications.resnet50 import preprocess_input
from sklearn.preprocessing import OneHotEncoder
random.shuffle(data_pair)
X = [item[0] for item in data_pair]
Y = [item[1] for item in data_pair]
image_data_list = []
for img_path in X:
img = image.load_img(img_path, target_size=(224, 224))
img_data = image.img_to_array(img)
img_data = preprocess_input(img_data)
image_data_np = np.array(img_data)
image_data_list.append(image_data_np)
image_data_list = np.array(image_data_list)
#print(len(image_data_list))
Y = np.array(Y)
onehot_encoder = OneHotEncoder(sparse=False)
Y = Y.reshape(len(Y), 1)
onehot_encoded = onehot_encoder.fit_transform(Y)
X_train, X_valid, Y_train, Y_valid = train_test_split(image_data_list,
onehot_encoded,
test_size=0.30,
random_state=42)
encoder = preprocessing.LabelEncoder()
encoder.fit(Classes['Bird Class'].values)
print(encoder.classes_)
y_train = onehot(encoder.transform(y_train1))
y_valid = onehot(encoder.transform(y_valid1))
y_test = onehot(encoder.transform(y_test1))
print(np.sum(np.argmax(y_train, axis=1) == encoder.transform(y_train1)))
print(np.sum(np.argmax(y_valid, axis=1) == encoder.transform(y_valid1)))
print(np.sum(np.argmax(y_test, axis=1) == encoder.transform(y_test1)))
x_train, y_train = shuffle(x_train, y_train)
x_valid, y_valid = shuffle(x_valid, y_valid)
x_test, y_test = shuffle(x_test, y_test)
x_train = preprocess_input(x_train)
x_valid = preprocess_input(x_valid)
x_test = preprocess_input(x_test)
# In[113]:
# trainning process
nb_epoch = int(sys.argv[1])
batch_size = int(sys.argv[2])
print("Epochs: {}".format(nb_epoch), "Batch Size: {}".format(batch_size))
checkpointer = ModelCheckpoint(filepath=sys.argv[3],
verbose=1,
monitor='val_categorical_accuracy',
save_best_only=True)
model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size),
steps_per_epoch=(x_train.shape[0] / batch_size) + 1,
# print(imgs.shape)
ag_imgs = []
for k in range(imgs.shape[0]):
n = 0
for ag in datagen.flow(imgs[k], batch_size=1):
ag_imgs.append(ag)
new_labels.append(label[k])
n += 1
if n > 5: #增强数量5张
break
# for ag in datagen.flow(,batch_size=1):
ag_imgs = np.asarray(ag_imgs)
ag_imgs = ag_imgs.reshape(ag_imgs.shape[0], ag_imgs.shape[2],
ag_imgs.shape[3], ag_imgs.shape[4])
if i == 0:
train_feature = model.predict(preprocess_input(ag_imgs))
else:
print(train_feature.shape,
model.predict(preprocess_input(ag_imgs)).shape)
train_feature = np.vstack(
(train_feature, model.predict(preprocess_input(ag_imgs))))
print(train_feature.shape, len(new_labels))
new_labels = np.asarray(new_labels)
list_2 = [test_data_path + list for list in os.listdir(test_data_path)]
for i in range(int(len(list_2) / BATCHSIZE) + 1):
imgs = []
for j in range(BATCHSIZE):
if i * BATCHSIZE + j < len(list_2):
imgs.append(
img_to_array(
load_img(list_2[i * BATCHSIZE + j],
######################################################################
# Load a test image
# ------------------
# A single cat dominates the examples!
from PIL import Image
from matplotlib import pyplot as plt
from keras.applications.resnet50 import preprocess_input
img_url = 'https://github.com/dmlc/mxnet.js/blob/master/data/cat.png?raw=true'
img_path = download_testdata(img_url, 'cat.png', module='data')
img = Image.open(img_path).resize((224, 224))
plt.imshow(img)
plt.show()
# input preprocess
data = np.array(img)[np.newaxis, :].astype('float32')
data = preprocess_input(data).transpose([0, 3, 1, 2])
######################################################################
# Compile the model with Relay
# ----------------------------
# convert the keras model(NHWC layout) to Relay format(NCHW layout).
shape_dict = {'input_1': data.shape}
mod, params = relay.frontend.from_keras(keras_resnet50, shape_dict)
# compile the model
target = "llvm -mcpu=core-avx2"
with relay.build_config(opt_level=3):
graph, lib, params = relay.build(mod, target, params=params)
from tvm.contrib import graph_runtime as runtime
def main():
parser = argparse.ArgumentParser(
description=
'Find latent representation of reference images using perceptual loss')
parser.add_argument('src_dir', help='Directory with images for encoding')
parser.add_argument('generated_images_dir',
help='Directory for storing generated images')
parser.add_argument('dlatent_dir',
help='Directory for storing dlatent representations')
parser.add_argument('--data_dir',
default='data',
help='Directory for storing optional models')
parser.add_argument('--mask_dir',
default='masks',
help='Directory for storing optional masks')
parser.add_argument('--load_last',
default='',
help='Start with embeddings from directory')
parser.add_argument(
'--dlatent_avg',
default='',
help=
'Use dlatent from file specified here for truncation instead of dlatent_avg from Gs'
)
parser.add_argument(
'--model_url',
default='gdrive:networks/stylegan2-ffhq-config-f.pkl',
help='Fetch a StyleGAN model to train on from this URL')
parser.add_argument('--model_res',
default=1024,
help='The dimension of images in the StyleGAN model',
type=int)
parser.add_argument('--batch_size',
default=1,
help='Batch size for generator and perceptual model',
type=int)
parser.add_argument(
'--optimizer',
default='ggt',
help='Optimization algorithm used for optimizing dlatents')
# Perceptual model params
parser.add_argument(
'--vgg_url',
default=
'https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2',
help='Fetch VGG model on from this URL')
parser.add_argument('--image_size',
default=256,
help='Size of images for perceptual model',
type=int)
parser.add_argument('--resnet_image_size',
default=256,
help='Size of images for the Resnet model',
type=int)
parser.add_argument('--lr',
default=0.25,
help='Learning rate for perceptual model',
type=float)
parser.add_argument('--decay_rate',
default=0.9,
help='Decay rate for learning rate',
type=float)
parser.add_argument('--iterations',
default=100,
help='Number of optimization steps for each batch',
type=int)
parser.add_argument(
'--decay_steps',
default=4,
help='Decay steps for learning rate decay (as a percent of iterations)',
type=float)
parser.add_argument('--early_stopping',
default=True,
help='Stop early once training stabilizes',
type=str2bool,
nargs='?',
const=True)
parser.add_argument('--early_stopping_threshold',
default=0.5,
help='Stop after this threshold has been reached',
type=float)
parser.add_argument('--early_stopping_patience',
default=10,
help='Number of iterations to wait below threshold',
type=int)
parser.add_argument(
'--load_effnet',
default='data/finetuned_effnet.h5',
help='Model to load for EfficientNet approximation of dlatents')
parser.add_argument(
'--load_resnet',
default='data/finetuned_resnet.h5',
help='Model to load for ResNet approximation of dlatents')
parser.add_argument(
'--use_preprocess_input',
default=True,
help='Call process_input() first before using feed forward net',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--use_best_loss',
default=True,
help='Output the lowest loss value found as the solution',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--average_best_loss',
default=0.25,
help=
'Do a running weighted average with the previous best dlatents found',
type=float)
parser.add_argument('--sharpen_input',
default=True,
help='Sharpen the input images',
type=str2bool,
nargs='?',
const=True)
# Loss function options
parser.add_argument(
'--use_vgg_loss',
default=0.4,
help='Use VGG perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_vgg_layer',
default=9,
help='Pick which VGG layer to use.',
type=int)
parser.add_argument(
'--use_pixel_loss',
default=1.5,
help='Use logcosh image pixel loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_mssim_loss',
default=200,
help='Use MS-SIM perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_lpips_loss',
default=100,
help='Use LPIPS perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_l1_penalty',
default=0.5,
help='Use L1 penalty on latents; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_discriminator_loss',
default=0.5,
help='Use trained discriminator to evaluate realism.',
type=float)
parser.add_argument(
'--use_adaptive_loss',
default=False,
help=
'Use the adaptive robust loss function from Google Research for pixel and VGG feature loss.',
type=str2bool,
nargs='?',
const=True)
# Generator params
parser.add_argument('--randomize_noise',
default=False,
help='Add noise to dlatents during optimization',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--tile_dlatents',
default=False,
help='Tile dlatents to use a single vector at each scale',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--clipping_threshold',
default=2.0,
help='Stochastic clipping of gradient values outside of this threshold',
type=float)
# Masking params
parser.add_argument('--load_mask',
default=False,
help='Load segmentation masks',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--face_mask',
default=True,
help='Generate a mask for predicting only the face area',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--use_grabcut',
default=True,
help=
'Use grabcut algorithm on the face mask to better segment the foreground',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--scale_mask',
default=1.4,
help='Look over a wider section of foreground for grabcut',
type=float)
parser.add_argument(
'--composite_mask',
default=True,
help='Merge the unmasked area back into the generated image',
type=str2bool,
nargs='?',
const=True)
parser.add_argument(
'--composite_blur',
default=8,
help='Size of blur filter to smoothly composite the images',
type=int)
# Video params
parser.add_argument('--video_dir',
default='videos',
help='Directory for storing training videos')
parser.add_argument('--output_video',
default=False,
help='Generate videos of the optimization process',
type=bool)
parser.add_argument('--video_codec',
default='MJPG',
help='FOURCC-supported video codec name')
parser.add_argument('--video_frame_rate',
default=24,
help='Video frames per second',
type=int)
parser.add_argument('--video_size',
default=512,
help='Video size in pixels',
type=int)
parser.add_argument(
'--video_skip',
default=1,
help='Only write every n frames (1 = write every frame)',
type=int)
args, other_args = parser.parse_known_args()
args.decay_steps *= 0.01 * args.iterations # Calculate steps as a percent of total iterations
if args.output_video:
import cv2
synthesis_kwargs = dict(output_transform=dict(
func=tflib.convert_images_to_uint8, nchw_to_nhwc=False),
minibatch_size=args.batch_size)
ref_images = [
os.path.join(args.src_dir, x) for x in os.listdir(args.src_dir)
if x[0] not in '._'
]
ref_images = list(filter(os.path.isfile, ref_images))
if len(ref_images) == 0:
raise Exception('%s is empty' % args.src_dir)
os.makedirs(args.data_dir, exist_ok=True)
os.makedirs(args.mask_dir, exist_ok=True)
os.makedirs(args.generated_images_dir, exist_ok=True)
os.makedirs(args.dlatent_dir, exist_ok=True)
os.makedirs(args.video_dir, exist_ok=True)
# Initialize generator and perceptual model
tflib.init_tf()
generator_network, discriminator_network, Gs_network = pretrained_networks.load_networks(
args.model_url)
generator = Generator(Gs_network,
args.batch_size,
randomize_noise=args.randomize_noise)
if (args.dlatent_avg != ''):
generator.set_dlatent_avg(np.load(args.dlatent_avg))
perc_model = None
if (args.use_lpips_loss > 0.00000001):
with dnnlib.util.open_url(args.vgg_url,
cache_dir='.stylegan2-cache') as f:
perc_model = pickle.load(f)
perceptual_model = PerceptualModel(args,
perc_model=perc_model,
batch_size=args.batch_size)
perceptual_model.build_perceptual_model(generator, discriminator_network)
ff_model = None
# Optimize (only) dlatents by minimizing perceptual loss between reference and generated images in feature space
for images_batch in tqdm(split_to_batches(ref_images, args.batch_size),
total=len(ref_images) // args.batch_size):
names = [
os.path.splitext(os.path.basename(x))[0] for x in images_batch
]
if args.output_video:
video_out = {}
for name in names:
video_out[name] = cv2.VideoWriter(
os.path.join(args.video_dir, f'{name}.avi'),
cv2.VideoWriter_fourcc(*args.video_codec),
args.video_frame_rate, (args.video_size, args.video_size))
perceptual_model.set_reference_images(images_batch)
perceptual_model.set_reference_images(images_batch)
dlatents = None
if (args.load_last != ''): # load previous dlatents for initialization
for name in names:
dl = np.expand_dims(np.load(
os.path.join(args.load_last, f'{name}.npy')),
axis=0)
if (dlatents is None):
dlatents = dl
else:
dlatents = np.vstack((dlatents, dl))
else:
if (ff_model is None):
if os.path.exists(args.load_resnet):
from keras.applications.resnet50 import preprocess_input
print("Loading ResNet Model:")
ff_model = load_model(args.load_resnet)
if (ff_model is None):
if os.path.exists(args.load_effnet):
import efficientnet
from efficientnet import preprocess_input
print("Loading EfficientNet Model:")
ff_model = load_model(args.load_effnet)
if (ff_model
is not None): # predict initial dlatents with ResNet model
if (args.use_preprocess_input):
dlatents = ff_model.predict(
preprocess_input(
load_images(images_batch,
image_size=args.resnet_image_size)))
else:
dlatents = ff_model.predict(
load_images(images_batch,
image_size=args.resnet_image_size))
if dlatents is not None:
generator.set_dlatents(dlatents)
op = perceptual_model.optimize(generator.dlatent_variable,
iterations=args.iterations,
use_optimizer=args.optimizer)
pbar = tqdm(op, leave=False, total=args.iterations)
vid_count = 0
best_loss = None
best_dlatent = None
avg_loss_count = 0
if args.early_stopping:
avg_loss = prev_loss = None
for loss_dict in pbar:
if args.early_stopping: # early stopping feature
if prev_loss is not None:
if avg_loss is not None:
avg_loss = 0.5 * avg_loss + (prev_loss -
loss_dict["loss"])
if avg_loss < args.early_stopping_threshold: # count while under threshold; else reset
avg_loss_count += 1
else:
avg_loss_count = 0
if avg_loss_count > args.early_stopping_patience: # stop once threshold is reached
print("")
break
else:
avg_loss = prev_loss - loss_dict["loss"]
pbar.set_description(" ".join(names) + ": " + "; ".join(
["{} {:.4f}".format(k, v) for k, v in loss_dict.items()]))
if best_loss is None or loss_dict["loss"] < best_loss:
if best_dlatent is None or args.average_best_loss <= 0.00000001:
best_dlatent = generator.get_dlatents()
else:
best_dlatent = 0.25 * best_dlatent + 0.75 * generator.get_dlatents(
)
if args.use_best_loss:
generator.set_dlatents(best_dlatent)
best_loss = loss_dict["loss"]
if args.output_video and (vid_count % args.video_skip == 0):
batch_frames = generator.generate_images()
for i, name in enumerate(names):
video_frame = PIL.Image.fromarray(
batch_frames[i], 'RGB').resize(
(args.video_size, args.video_size),
PIL.Image.LANCZOS)
video_out[name].write(
cv2.cvtColor(
np.array(video_frame).astype('uint8'),
cv2.COLOR_RGB2BGR))
generator.stochastic_clip_dlatents()
prev_loss = loss_dict["loss"]
if not args.use_best_loss:
best_loss = prev_loss
print(" ".join(names), " Loss {:.4f}".format(best_loss))
if args.output_video:
for name in names:
video_out[name].release()
# Generate images from found dlatents and save them
if args.use_best_loss:
generator.set_dlatents(best_dlatent)
generated_images = generator.generate_images()
generated_dlatents = generator.get_dlatents()
for img_array, dlatent, img_path, img_name in zip(
generated_images, generated_dlatents, images_batch, names):
mask_img = None
if args.composite_mask and (args.load_mask or args.face_mask):
_, im_name = os.path.split(img_path)
mask_img = os.path.join(args.mask_dir, f'{im_name}')
if args.composite_mask and mask_img is not None and os.path.isfile(
mask_img):
orig_img = PIL.Image.open(img_path).convert('RGB')
width, height = orig_img.size
imask = PIL.Image.open(mask_img).convert('L').resize(
(width, height))
imask = imask.filter(
ImageFilter.GaussianBlur(args.composite_blur))
mask = np.array(imask) / 255
mask = np.expand_dims(mask, axis=-1)
img_array = mask * np.array(img_array) + (
1.0 - mask) * np.array(orig_img)
img_array = img_array.astype(np.uint8)
#img_array = np.where(mask, np.array(img_array), orig_img)
img = PIL.Image.fromarray(img_array, 'RGB')
img.save(
os.path.join(args.generated_images_dir, f'{img_name}.png'),
'PNG')
np.save(os.path.join(args.dlatent_dir, f'{img_name}.npy'), dlatent)
generator.reset_dlatents()
#-*- coding: utf-8 -*-
from keras.applications.resnet50 import preprocess_input
import numpy as np
import cv2
from src.model_builder import CamModelBuilder
from src.utils import plot_img, list_files
if __name__ == "__main__":
detector = CamModelBuilder().get_cam_model()
detector.load_weights("weights.h5", by_name=True)
detector.summary()
imgs = list_files("dataset//train//text")
for i, img_path in enumerate(imgs):
original_img = cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2RGB)
img = cv2.resize(original_img, (224, 224))
img = np.expand_dims(img, 0).astype(np.float64)
cam_map = detector.predict(preprocess_input(img))
cam_map = cam_map[0, :, :, 1]
cam_map = cv2.resize(cam_map, (original_img.shape[1], original_img.shape[0]))
plot_img(original_img, cam_map, show=False, save_filename="{}.png".format(i+1))
print('the class', j)
path = 'D:\\sns_images\\class\\' + str(j)
for root, dir, files in os.walk(path):
imgbatch = zeros((batchsize, 224, 224, 3))
wordindex = 0
num = 0
classdata = []
for file in files:
try:
img = image.load_img(path + '\\' + file, target_size=(224, 224))
if wordindex == batchsize:
imgbatch = zeros((batchsize, 224, 224, 3))
wordindex = 0
img = image.img_to_array(img)
img = expand_dims(img, axis=0)
img = preprocess_input(img)
imgbatch[wordindex] = img
wordindex += 1
if wordindex == batchsize:
num += 1
print(num)
batchresult = model.predict(imgbatch)
batchmean = list(mean(batchresult, axis=0))
classdata.append(batchmean)
except:
print("iii")
classdata = mat(classdata)
classmean = mean(classdata, axis=0)
allclassmean[j] = classmean[0, :]
save('D:\\tencent\\miniproject\\allclassmean', allclassmean)
test_feature = np.array([
0,
])
BATCHSIZE = 1024
target_size = (224, 224)
list_2 = [
'./image2/image/' + list for list in os.listdir("./image2/image")
]
for i in range(int(len(list_2) / BATCHSIZE) + 1):
imgs = []
for j in range(BATCHSIZE):
imgs.append(
img_to_array(
load_img(list_2[i * BATCHSIZE + j],
target_size=target_size)))
if i * BATCHSIZE + j == len(list_2) - 1:
break
imgs = np.asarray(imgs)
if i == 0:
test_feature = model.predict(preprocess_input(imgs))
else:
test_feature = np.vstack(
[test_feature,
model.predict(preprocess_input(imgs))])
print(test_feature.shape)
# y = np.asarray(y, dtype=np.uint8)
with h5py.File("./dense161_final_feature.h5") as h:
# h.create_dataset("train", data=train_feature)
h.create_dataset("test", data=test_feature)
# h.create_dataset("label", data=y)
def main():
parser = argparse.ArgumentParser(
description='Find latent representation of reference images using perceptual losses',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('src_dir', help='Directory with images for encoding')
parser.add_argument('generated_images_dir', help='Directory for storing generated images')
parser.add_argument('dlatent_dir', help='Directory for storing dlatent representations')
parser.add_argument('--data_dir', default='data', help='Directory for storing optional models')
parser.add_argument('--mask_dir', default='masks', help='Directory for storing optional masks')
parser.add_argument('--load_last', default='', help='Start with embeddings from directory')
parser.add_argument('--dlatent_avg', default='',
help='Use dlatent from file specified here for truncation instead of dlatent_avg from Gs')
parser.add_argument('--model_url', default=config.Model,
help='Fetch a StyleGAN model to train on from this URL') # karras2019stylegan-ffhq-1024x1024.pkl
parser.add_argument('--model_res', default=1024, help='The dimension of images in the StyleGAN model', type=int)
parser.add_argument('--batch_size', default=1, help='Batch size for generator and perceptual model', type=int)
# Perceptual model params
parser.add_argument('--image_size', default=256, help='Size of images for perceptual model', type=int)
parser.add_argument('--resnet_image_size', default=224, help='Size of images for the Resnet model', type=int)
parser.add_argument('--lr', default=0.02, help='Learning rate for perceptual model', type=float)
parser.add_argument('--decay_rate', default=0.9, help='Decay rate for learning rate', type=float)
parser.add_argument('--iterations', default=100, help='Number of optimization steps for each batch', type=int)
parser.add_argument('--decay_steps', default=10,
help='Decay steps for learning rate decay (as a percent of iterations)', type=float)
parser.add_argument('--load_effnet', default='data/finetuned_effnet.h5',
help='Model to load for EfficientNet approximation of dlatents')
parser.add_argument('--load_resnet', default='data/resnet_18.h5',
help='Model to load for ResNet approximation of dlatents')
# Loss function options
parser.add_argument('--use_vgg_loss', default=0.4, help='Use VGG perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_vgg_layer', default=9, help='Pick which VGG layer to use.', type=int)
parser.add_argument('--use_pixel_loss', default=1.5,
help='Use logcosh image pixel loss; 0 to disable, > 0 to scale.', type=float)
parser.add_argument('--use_mssim_loss', default=100, help='Use MS-SIM perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_lpips_loss', default=100, help='Use LPIPS perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_l1_penalty', default=1, help='Use L1 penalty on latents; 0 to disable, > 0 to scale.',
type=float)
# Generator params
parser.add_argument('--randomize_noise', default=False, help='Add noise to dlatents during optimization', type=bool)
parser.add_argument('--tile_dlatents', default=False, help='Tile dlatents to use a single vector at each scale',
type=bool)
parser.add_argument('--clipping_threshold', default=2.0,
help='Stochastic clipping of gradient values outside of this threshold', type=float)
# Masking params
parser.add_argument('--load_mask', default=False, help='Load segmentation masks', type=bool)
parser.add_argument('--face_mask', default=False, help='Generate a mask for predicting only the face area',
type=bool)
parser.add_argument('--use_grabcut', default=True,
help='Use grabcut algorithm on the face mask to better segment the foreground', type=bool)
parser.add_argument('--scale_mask', default=1.5, help='Look over a wider section of foreground for grabcut',
type=float)
# Video params
parser.add_argument('--video_dir', default='videos', help='Directory for storing training videos')
parser.add_argument('--output_video', default=False, help='Generate videos of the optimization process', type=bool)
parser.add_argument('--video_codec', default='MJPG', help='FOURCC-supported video codec name')
parser.add_argument('--video_frame_rate', default=24, help='Video frames per second', type=int)
parser.add_argument('--video_size', default=512, help='Video size in pixels', type=int)
parser.add_argument('--video_skip', default=1, help='Only write every n frames (1 = write every frame)', type=int)
# 获取到基本设置时，如果运行命令中传入了之后才会获取到的其他配置，不会报错；而是将多出来的部分保存起来，留到后面使用
args, other_args = parser.parse_known_args()
# learning rate衰减的steps
args.decay_steps *= 0.01 * args.iterations # Calculate steps as a percent of total iterations
if args.output_video:
import cv2
synthesis_kwargs = dict(output_transform=dict(func=tflib.convert_images_to_uint8, nchw_to_nhwc=False),
minibatch_size=args.batch_size)
# 找到src_dir下所有图片文件，加入ref_images列表（即：源图的列表；只有一个图片也可以）
ref_images = [os.path.join(args.src_dir, x) for x in os.listdir(args.src_dir)]
ref_images = list(filter(os.path.isfile, ref_images))
if len(ref_images) == 0:
raise Exception('%s is empty' % args.src_dir)
# 创建工作目录
os.makedirs(args.data_dir, exist_ok=True)
os.makedirs(args.mask_dir, exist_ok=True)
os.makedirs(args.generated_images_dir, exist_ok=True)
os.makedirs(args.dlatent_dir, exist_ok=True)
os.makedirs(args.video_dir, exist_ok=True)
# Initialize generator and perceptual model
tflib.init_tf()
# 加载StyleGAN模型
model_file = glob.glob(args.model_url)
if len(model_file) == 1:
model_file = open(model_file[0], "rb")
else:
raise Exception('Failed to find the model')
generator_network, discriminator_network, Gs_network = pickle.load(model_file)
# 加载Generator类，参与构建VGG16 perceptual model，用于调用（说是生成，更好理解）generated_image
# generated_image通过perceptual_model转化为generated_img_features，参与计算loss
generator = Generator(Gs_network, args.batch_size, clipping_threshold=args.clipping_threshold,
tiled_dlatent=args.tile_dlatents, model_res=args.model_res,
randomize_noise=args.randomize_noise)
if (args.dlatent_avg != ''):
generator.set_dlatent_avg(np.load(args.dlatent_avg))
perc_model = None
if (args.use_lpips_loss > 0.00000001): # '--use_lpips_loss', default = 100
# 加载VGG16 perceptual模型
model_file = glob.glob('./models/vgg16_zhang_perceptual.pkl')
if len(model_file) == 1:
model_file = open(model_file[0], "rb")
else:
raise Exception('Failed to find the model')
perc_model = pickle.load(model_file)
# 创建VGG16 perceptual模型
perceptual_model = PerceptualModel(args, perc_model=perc_model, batch_size=args.batch_size)
perceptual_model.build_perceptual_model(generator)
ff_model = None
# Optimize (only) dlatents by minimizing perceptual loss between reference and generated images in feature space
# tqdm 是一个快速，可扩展的Python进度条，可以在 Python 长循环中添加一个进度提示信息
# 把ref_images分割为若干批次，每个批次的大小为args.batch_size，分批使用perceptual_model.optimize()求解每个源图的dlatents的最优解
# 对每一个源图，优化迭代的过程是从一个初始dlatents开始，在某个空间内，按正态分布取值，使用Adam优化器，逐步寻找使loss最小的dlatents，即：stochastic clipping方法
for images_batch in tqdm(split_to_batches(ref_images, args.batch_size), total=len(ref_images) // args.batch_size):
print('开始计时')
starttime = time.time()
# 读取每个批次中的文件名
names = [os.path.splitext(os.path.basename(x))[0] for x in images_batch]
if args.output_video:
video_out = {}
for name in names:
video_out[name] = cv2.VideoWriter(os.path.join(args.video_dir, f'{name}.avi'),
cv2.VideoWriter_fourcc(*args.video_codec), args.video_frame_rate,
(args.video_size, args.video_size))
# 给源图及源图用VGG16生成的features赋值(这是计算loss的基准)
perceptual_model.set_reference_images(images_batch)
dlatents = None
if (args.load_last != ''): # load previous dlatents for initialization
for name in names:
dl = np.expand_dims(np.load(os.path.join(args.load_last, f'{name}.npy')), axis=0)
if (dlatents is None):
dlatents = dl
else:
dlatents = np.vstack((dlatents, dl))
else:
if (ff_model is None):
if os.path.exists(args.load_resnet):
print("Loading ResNet Model:")
ff_model = load_model(args.load_resnet)
from keras.applications.resnet50 import preprocess_input
if (ff_model is None):
if os.path.exists(args.load_effnet):
import efficientnet
print("Loading EfficientNet Model:")
ff_model = load_model(args.load_effnet)
from efficientnet import preprocess_input
if (ff_model is not None): # predict initial dlatents with ResNet model
dlatents = ff_model.predict(
preprocess_input(load_images(images_batch, image_size=args.resnet_image_size)))
# 设置用于perceptual_model优化迭代的初始值dlatents，它是用resnet50或者efficientnet从源图预测得到的
if dlatents is not None:
generator.set_dlatents(dlatents)
# 对每一个源图，用tqdm构造进度条，显示优化迭代的过程
op = perceptual_model.optimize(generator.dlatent_variable, iterations=args.iterations)
pbar = tqdm(op, leave=False, total=args.iterations)
vid_count = 0
best_loss = None
best_dlatent = None
# 用stochastic clipping方法，使用VGG16 perceptual_model进行优化迭代，迭代次数为iterations=args.iterations
endtime = time.time()
print('开始迭代时间为:', round(endtime - starttime, 2), 'secs')
for loss_dict in pbar:
pbar.set_description(" ".join(names) + ": " + "; ".join(["{} {:.4f}".format(k, v)
for k, v in loss_dict.items()]))
if best_loss is None or loss_dict["loss"] < best_loss:
best_loss = loss_dict["loss"]
best_dlatent = generator.get_dlatents()
if args.output_video and (vid_count % args.video_skip == 0):
batch_frames = generator.generate_images()
for i, name in enumerate(names):
video_frame = PIL.Image.fromarray(batch_frames[i], 'RGB').resize((args.video_size, args.video_size),
PIL.Image.LANCZOS)
video_out[name].write(cv2.cvtColor(np.array(video_frame).astype('uint8'), cv2.COLOR_RGB2BGR))
# 用stochastic clip方法更新dlatent_variable
generator.stochastic_clip_dlatents()
print(" ".join(names), " Loss {:.4f}".format(best_loss))
if args.output_video:
for name in names:
video_out[name].release()
# Generate images from found dlatents and save them
generator.set_dlatents(best_dlatent)
generated_images = generator.generate_images()
generated_dlatents = generator.get_dlatents()
endtime = time.time()
print('计算完成时间为:', round(endtime - starttime, 2), 'secs')
for img_array, dlatent, img_name in zip(generated_images, generated_dlatents, names):
img = PIL.Image.fromarray(img_array, 'RGB')
img.save(os.path.join(args.generated_images_dir, f'{img_name}.png'), 'PNG')
np.save(os.path.join(args.dlatent_dir, f'{img_name}.npy'), dlatent)
generator.reset_dlatents()
endtime = time.time()
print('写入完成时间为:', round(endtime - starttime, 2), 'secs')
def preprocess_image(img):
img = image.load_img(img, target_size=(224,224))
img = image.img_to_array(img)
img = np.expand_dims(img, axis=0)
img = preprocess_input(img)
return img
def extract_Resnet50(tensor):
return ResNet50(weights='imagenet',
include_top=False).predict(preprocess_input(tensor))
def ResNet50_predict_labels(img_path):
# 返回img_path路径的图像的预测向量
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
if __name__ == '__main__':
#Output dim for your dataset
output_dim = 257 #For Caltech256
images_list = []
images_names = []
val_images_list = []
val_images_names = []
examples_per_class = 3
validation_per_class = 1
images_list, images_names, val_images_list, val_images_names = load_all_images(
'C:/Users/Chetan/Documents/CSE253/PA3/256_ObjectCategories',
'256_ObjectCategories', images_list, images_names, val_images_list,
val_images_names, examples_per_class, validation_per_class)
# Normalization
images_list = preprocess_input(np.array(images_list))
val_images_list = preprocess_input(np.array(val_images_list))
images_list = images_list / 255.0
val_images_list = val_images_list / 255.0
# Get one hot representation
image_category = get_one_hot(images_names)
val_image_category = get_one_hot(val_images_names)
# Shuffle
X_train, y_train = shuffle(images_list, image_category)
tl_model = getModel(output_dim)
tl_model.summary()
#Train the model
tl_model.compile(loss='categorical_crossentropy',
optimizer=RMSprop(lr=0.001,
decay=1e-2,
# More image pre-processing.
im_224 = image.img_to_array(im_224)
im_224 = np.expand_dims(
im_224, axis=0
) # This gives the image 4 dimensions and is necessary for future steps.
im_299 = image.img_to_array(im_299)
im_299 = np.expand_dims(im_299, axis=0)
im_331 = image.img_to_array(im_331)
im_331 = np.expand_dims(im_331, axis=0)
# Generate predictions for each model.
resnet_pred = imagenet_utils.decode_predictions(
resnet.predict(resnet50.preprocess_input(im_224)), 5)
xc_pred = imagenet_utils.decode_predictions(
xc.predict(xception.preprocess_input(im_299)), 5)
v19_pred = imagenet_utils.decode_predictions(
v19.predict(vgg19.preprocess_input(im_224)), 5)
ic3_pred = imagenet_utils.decode_predictions(
ic3.predict(inception_v3.preprocess_input(im_299)), 5)
ic_resnet_pred = imagenet_utils.decode_predictions(
ic_resnet.predict(inception_resnet_v2.preprocess_input(im_299)), 5)
mobile_pred = imagenet_utils.decode_predictions(
mobile.predict(mobilenet.preprocess_input(im_224)), 5)
nn_large_pred = imagenet_utils.decode_predictions(
nn_large.predict(nasnet.preprocess_input(im_331)), 5)
# Update result dictionaries based on tags.
resnet_results = update_dict(preds=resnet_pred,
def do_POST(s):
length = int(s.headers['Content-Length'])
body = s.rfile.read(length).decode('utf-8')
if s.headers['Content-type'] == 'application/json':
post_data = json.loads(body)
else:
post_data = urllib.parse.parse_qs(body)
modelid = post_data['model']
try:
model = model_impls[modelid]['class'](**model_impls[modelid]['params'])
except Exception as e:
logger.error("Unable to load model: {reason}".format(reason=e.message))
s.send_response(300)
s.send_header("Content-type", "application/json")
s.end_headers()
json.dump({
"status": 300,
"message": e.message,
}, s.wfile)
return
target_size = (dict([(m["id"],m['image_size']) for m in models]))[modelid]
concepts = []
for annotation in post_data['annotations']:
aid = annotation['annotationid']
begin = annotation['begin']
begin = annotation['end']
batch_x = np.zeros((len(annotation['frames']),target_size,target_size,3), dtype=np.float32)
for i,frame in enumerate(annotation['frames']):
# Load image to PIL format
img = Image.open(BytesIO(base64.b64decode(frame['screenshot'])))
# cache frame - FIXME: currently there is no mean to identify the video - same timstamp will overwrite an old frame (hash?)
img.save(os.path.join(CACHE_DIR,'{0}.png'.format(frame['timecode'])))
if img.mode != 'RGB':
img = img.convert('RGB')
hw_tuple = (target_size, target_size)
if img.size != hw_tuple:
logger.warn("Scaling image to model size - this should be done in advene!")
img = img.resize(hw_tuple)
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
batch_x[i] = x[0,:,:,:]
preds = model.predict_on_batch(np.asarray(batch_x))
# decode the results into a list of tuples (class, description, probability)
# (one such list for each sample in the batch)
decoded = decode_predictions(preds, top=top_n_preds)
confidences = dict()
for t in itertools.chain.from_iterable(decoded):
if t[1] in confidences:
confidences[t[1]].append(float(t[2]))
else:
confidences[t[1]] = [float(t[2])]
logger.debug(confidences)
concepts.extend([
{
'annotationid': aid,
'confidence': max(confidences[l]),
#FIXME: set correct timecode - set timecode of frame with max confidence?
'timecode': annotation['begin'], #timestamp_in_ms,
'label': l,
'uri': 'http://concept.org/%s' % l
} for l in confidences]
)
logger.debug(concepts)
s.send_response(200)
s.send_header("Content-type", "application/json")
s.end_headers()
response=json.dumps({
"status": 200,
"message": "OK",
"data": {
'media_filename': post_data["media_filename"],
'media_uri': post_data["media_uri"],
'concepts': concepts
}
})
s.wfile.write(response.encode())
resnet.summary()
# make a model to get output before flatten
activation_layer = resnet.get_layer('activation_49')
# create a model object
model = Model(inputs=resnet.input, outputs=activation_layer.output)
# get the feature map weights
final_dense = resnet.get_layer('fc1000')
W = final_dense.get_weights()[0]
while True:
img = image.load_img(np.random.choice(image_files), target_size=(224, 224))
x = preprocess_input(np.expand_dims(img, 0))
fmaps = model.predict(x)[0] # 7 x 7 x 2048
# get predicted class
probs = resnet.predict(x)
classnames = decode_predictions(probs)[0]
print(classnames)
classname = classnames[0][1]
pred = np.argmax(probs[0])
# get the 2048 weights for the relevant class
w = W[:, pred]
# "dot" w with fmaps
cam = fmaps.dot(w)
def ResNet50_predict_labels(img_path):
img = preprocess_input(to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
def ResNet50_predict_labels(img_path):
# 返回img_path路径的图像的预测向量
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
from keras.applications.resnet50 import ResNet50
from keras.preprocessing import image
from keras.applications.resnet50 import preprocess_input, decode_predictions
import numpy as np
height = 224
width = 224
channels = 3
top = 5
def load_image(path):
img = image.load_img(path, target_size=(224, 224))
x = image.img_to_array(img)
return x
if __name__ == '__main__':
args = docopt(__doc__)
paths = args['IMAGE_PATH']
model = ResNet50(weights='imagenet')
x = np.zeros((len(paths), height, width, channels))
for i, path in enumerate(paths):
x[i] = load_image(path)
x = preprocess_input(x)
predictions = decode_predictions(model.predict(x), top=top)
for prediction in predictions:
print(prediction)
def _preprocess(self, image_tensor):
"""Preprocess image data by modifying it directly"""
resnet50.preprocess_input(image_tensor)
def main():
parser = argparse.ArgumentParser(
description=
'Find latent representation of reference images using perceptual losses',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('src_dir', help='Directory with images for encoding')
parser.add_argument('generated_images_dir',
help='Directory for storing generated images')
parser.add_argument('dlatent_dir',
help='Directory for storing dlatent representations')
parser.add_argument('--data_dir',
default='data',
help='Directory for storing optional models')
parser.add_argument('--mask_dir',
default='masks',
help='Directory for storing optional masks')
parser.add_argument('--load_last',
default='',
help='Start with embeddings from directory')
parser.add_argument(
'--dlatent_avg',
default='',
help=
'Use dlatent from file specified here for truncation instead of dlatent_avg from Gs'
)
# parser.add_argument('--model_url', default='https://drive.google.com/uc?id=1MEGjdvVpUsu1jB4zrXZN7Y4kBBOzizDQ', help='Fetch a StyleGAN model to train on from this URL') # karras2019stylegan-ffhq-1024x1024.pkl
parser.add_argument(
'--model_url',
default='./cache/karras2019stylegan-ffhq-1024x1024.pkl',
help='load local model -- Jack12'
) # karras2019stylegan-ffhq-1024x1024.pkl
parser.add_argument('--model_res',
default=1024,
help='The dimension of images in the StyleGAN model',
type=int)
parser.add_argument('--batch_size',
default=1,
help='Batch size for generator and perceptual model',
type=int)
# Perceptual model params
parser.add_argument('--image_size',
default=256,
help='Size of images for perceptual model',
type=int)
parser.add_argument('--resnet_image_size',
default=256,
help='Size of images for the Resnet model',
type=int)
parser.add_argument('--lr',
default=0.02,
help='Learning rate for perceptual model',
type=float)
parser.add_argument('--decay_rate',
default=0.9,
help='Decay rate for learning rate',
type=float)
parser.add_argument('--iterations',
default=200,
help='Number of optimization steps for each batch',
type=int)
parser.add_argument(
'--decay_steps',
default=10,
help='Decay steps for learning rate decay (as a percent of iterations)',
type=float)
parser.add_argument(
'--load_effnet',
default='data/finetuned_effnet.h5',
help='Model to load for EfficientNet approximation of dlatents')
parser.add_argument(
'--load_resnet',
default='data/finetuned_resnet.h5',
help='Model to load for ResNet approximation of dlatents')
# Loss function options
parser.add_argument(
'--use_vgg_loss',
default=0.4,
help='Use VGG perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_vgg_layer',
default=9,
help='Pick which VGG layer to use.',
type=int)
parser.add_argument(
'--use_pixel_loss',
default=1.5,
help='Use logcosh image pixel loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_mssim_loss',
default=100,
help='Use MS-SIM perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_lpips_loss',
default=100,
help='Use LPIPS perceptual loss; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument(
'--use_l1_penalty',
default=1,
help='Use L1 penalty on latents; 0 to disable, > 0 to scale.',
type=float)
parser.add_argument('--use_tex_loss',
default=3.0,
help='Use texture space loss.',
type=float)
# Generator params
parser.add_argument('--randomize_noise',
default=False,
help='Add noise to dlatents during optimization',
type=bool)
parser.add_argument(
'--tile_dlatents',
default=False,
help='Tile dlatents to use a single vector at each scale',
type=bool)
parser.add_argument(
'--clipping_threshold',
default=2.0,
help='Stochastic clipping of gradient values outside of this threshold',
type=float)
# Face Alignment Model
parser.add_argument('--crop_face',
default=False,
help='Crop face image use face alignment model',
type=bool)
# Masking params
parser.add_argument('--load_mask',
default=False,
help='Load segmentation masks',
type=bool)
parser.add_argument(
'--face_mask',
default=False,
help='Generate a mask for predicting only the face area',
type=bool)
parser.add_argument(
'--use_grabcut',
default=True,
help=
'Use grabcut algorithm on the face mask to better segment the foreground',
type=bool)
parser.add_argument(
'--scale_mask',
default=1.5,
help='Look over a wider section of foreground for grabcut',
type=float)
# Video params
parser.add_argument('--video_dir',
default='videos',
help='Directory for storing training videos')
parser.add_argument('--output_video',
default=False,
help='Generate videos of the optimization process',
type=bool)
parser.add_argument('--video_codec',
default='MJPG',
help='FOURCC-supported video codec name')
parser.add_argument('--video_frame_rate',
default=24,
help='Video frames per second',
type=int)
parser.add_argument('--video_size',
default=512,
help='Video size in pixels',
type=int)
parser.add_argument(
'--video_skip',
default=1,
help='Only write every n frames (1 = write every frame)',
type=int)
args, other_args = parser.parse_known_args()
args.decay_steps *= 0.01 * args.iterations # Calculate steps as a percent of total iterations
if args.output_video:
import cv2
synthesis_kwargs = dict(output_transform=dict(
func=tflib.convert_images_to_uint8, nchw_to_nhwc=False),
minibatch_size=args.batch_size)
ref_images = [
os.path.join(args.src_dir, x) for x in os.listdir(args.src_dir)
]
ref_images = list(filter(os.path.isfile, ref_images))
if len(ref_images) == 0:
raise Exception('%s is empty' % args.src_dir)
os.makedirs(args.data_dir, exist_ok=True)
os.makedirs(args.mask_dir, exist_ok=True)
os.makedirs(args.generated_images_dir, exist_ok=True)
os.makedirs(args.dlatent_dir, exist_ok=True)
os.makedirs(args.video_dir, exist_ok=True)
# Initialize generator and perceptual model
tflib.init_tf()
with dnnlib.util.open_url(args.model_url, cache_dir=config.cache_dir) as f:
print('Load model from %s' % (args.model_url))
generator_network, discriminator_network, Gs_network = pickle.load(f)
generator = Generator(Gs_network,
args.batch_size,
clipping_threshold=args.clipping_threshold,
tiled_dlatent=args.tile_dlatents,
model_res=args.model_res,
randomize_noise=args.randomize_noise)
if (args.dlatent_avg != ''):
generator.set_dlatent_avg(np.load(args.dlatent_avg))
perc_model = None
if (args.use_lpips_loss > 0.00000001):
# with dnnlib.util.open_url('https://drive.google.com/uc?id=1N2-m9qszOeVC9Tq77WxsLnuWwOedQiD2', cache_dir=config.cache_dir) as f:
# perc_model = pickle.load(f)
with dnnlib.util.open_url('./cache/vgg16_zhang_perceptual.pkl',
cache_dir=config.cache_dir) as f:
perc_model = pickle.load(f)
# Jack12 load local model
perceptual_model = PerceptualModel(args,
perc_model=perc_model,
batch_size=args.batch_size)
perceptual_model.build_perceptual_model(generator)
ff_model = None
cur_batch_id = 0
# Optimize (only) dlatents by minimizing perceptual loss between reference and generated images in feature space
for images_batch in tqdm(split_to_batches(ref_images, args.batch_size),
total=len(ref_images) // args.batch_size):
cur_batch_id += 1
names = [
os.path.splitext(os.path.basename(x))[0] for x in images_batch
]
if args.output_video:
video_out = {}
for name in names:
video_out[name] = cv2.VideoWriter(
os.path.join(args.video_dir, f'{name}.avi'),
cv2.VideoWriter_fourcc(*args.video_codec),
args.video_frame_rate, (args.video_size, args.video_size))
perceptual_model.set_reference_images(images_batch, cur_batch_id)
dlatents = None
if (args.load_last != ''): # load previous dlatents for initialization
for name in names:
dl = np.expand_dims(np.load(
os.path.join(args.load_last, f'{name}.npy')),
axis=0)
if (dlatents is None):
dlatents = dl
else:
dlatents = np.vstack((dlatents, dl))
else:
if (ff_model is None):
if os.path.exists(args.load_resnet):
print("Loading ResNet Model:")
ff_model = load_model(args.load_resnet)
from keras.applications.resnet50 import preprocess_input
if (ff_model is None):
if os.path.exists(args.load_effnet):
import efficientnet
print("Loading EfficientNet Model:")
ff_model = load_model(args.load_effnet)
from efficientnet import preprocess_input
if (ff_model
is not None): # predict initial dlatents with ResNet model
dlatents = ff_model.predict(
preprocess_input(
load_images(images_batch,
image_size=args.resnet_image_size)))
if dlatents is not None:
generator.set_dlatents(dlatents)
op = perceptual_model.optimize(generator.dlatent_variable,
iterations=args.iterations)
pbar = tqdm(op, leave=False, total=args.iterations)
vid_count = 0
best_loss = None
best_dlatent = None
for loss_dict in pbar:
pbar.set_description(" ".join(names) + ": " + "; ".join(
["{} {:.4f}".format(k, v) for k, v in loss_dict.items()]))
if best_loss is None or loss_dict["loss"] < best_loss:
best_loss = loss_dict["loss"]
best_dlatent = generator.get_dlatents()
if args.output_video and (vid_count % args.video_skip == 0):
batch_frames = generator.generate_images()
for i, name in enumerate(names):
video_frame = PIL.Image.fromarray(
batch_frames[i], 'RGB').resize(
(args.video_size, args.video_size),
PIL.Image.LANCZOS)
video_out[name].write(
cv2.cvtColor(
np.array(video_frame).astype('uint8'),
cv2.COLOR_RGB2BGR))
generator.stochastic_clip_dlatents()
print(" ".join(names), " Loss {:.4f}".format(best_loss))
if args.output_video:
for name in names:
video_out[name].release()
# Generate images from found dlatents and save them
generator.set_dlatents(best_dlatent)
generated_images = generator.generate_images()
generated_dlatents = generator.get_dlatents()
for img_array, dlatent, img_name in zip(generated_images,
generated_dlatents, names):
img = PIL.Image.fromarray(img_array, 'RGB')
img.save(
os.path.join(args.generated_images_dir, f'{img_name}.png'),
'PNG')
np.save(os.path.join(args.dlatent_dir, f'{img_name}.npy'), dlatent)
generator.reset_dlatents()
x = Dense(702, activation='softmax', name='fc8', kernel_initializer=RandomNormal(mean=0.0, stddev=0.001))(x)
net = Model(input=base_model.input, output=x)
for layer in net.layers:
layer.trainable = True
# load data
images, labels = [], []
with open(LIST, 'r') as f:
for line in f:
line = line.strip()
img, lbl = line.split()
img = image.load_img(os.path.join(TRAIN, img), target_size=[224, 224])
img = image.img_to_array(img)
img = np.expand_dims(img, axis=0)
img = preprocess_input(img)
images.append(img[0])
labels.append(int(lbl))
images = np.array(images)
labels = to_categorical(labels)
# train
batch_size = 16
datagen = ImageDataGenerator(featurewise_center=False,
samplewise_center=False,
featurewise_std_normalization=False,
samplewise_std_normalization=False,
zca_whitening=False,
rotation_range=20, # 0.
for img in images:
x = Image.open( img_path + '/' + img)
x = x.resize( ( 224 , 224 ))
x = np.array(x , dtype = np.float32)
x_.append(x)
y_.append(dict_labels[img_folder])
X = np.array(x_)
Y = np.array(y_)
print(X.shape)
print(Y.shape)
print(type(X[3]))
X = preprocess_input(X)
print(type(X[3]))
Y = np_utils.to_categorical(Y)
print(Y.shape)
X, Y = shuffle(X, Y)
X_train, X_test, Y_train, Y_test = train_test_split( X, Y, test_size = 0.2 )
print(X_train.shape, X_test.shape, Y_train.shape, Y_test.shape)
resnet_model = ResNet50( input_shape = (224,224,3) , include_top = False, weights = 'imagenet' )
model = ResNet50(
weights="../weight/resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5",
include_top=False,
pooling='avg')
for flower in flower_type:
data = []
dir = "../train/" + flower
print("Feature Extraction (" + flower + ") begin")
for img in os.listdir(dir):
path = os.path.join(dir, img)
img = image.load_img(path, target_size=(224, 224))
data.append(np.array(img))
X = np.array(data)
X = preprocess_input(X)
feature = model.predict(X)
np.save("../feature/" + flower + "_resnet50", feature)
print("Feature Extraction (" + flower + ") done")
dir = "../test"
data = []
print("Feature Extraction (test) begin")
for i in range(424):
path = "../test/" + str(i) + ".jpg"
img = image.load_img(path, target_size=(224, 224))
data.append(np.array(img))
test = np.array(data)
test = preprocess_input(test)
def generator(batch_size, image_list, image_shape, coco_instance, id_to_index,
is_training):
"""Generator
Generate the images for models to train
Args:
- batch_size: batch_size
- image_list: the list of file name of images
- image_shape: the target image shape
- coco_instance: the ground truth of COCO dataset
- id_to_index: dictionary project id to index
- is_training: open or close data augmentation
Returns:
- all_img: shape: (batch_size, image_shape[0], image_shape[1], 3)
- label: (batch_size, image_shape[0], image_shape[1], classes)
"""
aug, mask_hook = utilities.img_aug()
def f(id):
if id != 0:
return id_to_index[id]
else:
return id_to_index[183]
# Lambda function to convert id to index
vfunc = np.vectorize(f)
while True:
# Random Shuffling
random.shuffle(image_list)
index = 0
while index + batch_size < len(image_list):
all_img = np.zeros((batch_size, image_shape[0], image_shape[1], 3),
dtype=np.float32)
label = np.zeros(
(batch_size, image_shape[0], image_shape[1], len(id_to_index)))
i = 0
while i < batch_size:
image = image_list[index]
im = cv2.imread(image)
# Mode 4 stands for any image between bgr and gray scale
# Convert back to RGB
# If the width or height is smaller than indicated size
# or if the image is grayscale
#if (im.shape[0] < image_shape[0] or im.shape[1] < image_shape[1]) or im.ndim == 2:
# i -= 1
# continue
if (im.shape[0] < image_shape[0]
or im.shape[1] < image_shape[1]):
index += 1
continue
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
# Get the id containing 12 numbers (000000XXXXXX)
lbl_id = int(image.replace(".jpg", '')[-12:])
lbl = cocoSegmentationToSegmentationMap(coco_instance, lbl_id)
lbl = vfunc(lbl.astype(np.uint8))
# Resize
#im = transform.rescale(im, 0.5)
# Random Crop
rnd_x = random.randint(0, im.shape[0] - image_shape[0])
rnd_y = random.randint(0, im.shape[1] - image_shape[1])
crop_im = im[rnd_x:rnd_x + image_shape[0],
rnd_y:rnd_y + image_shape[1], :]
crop_lbl = lbl[rnd_x:rnd_x + image_shape[0],
rnd_y:rnd_y + image_shape[1]]
# Convert to one hot
crop_lbl = keras.utils.to_categorical(
crop_lbl, num_classes=len(id_to_index))
# Save data
all_img[i] = crop_im
label[i] = crop_lbl
index += 1
i += 1
if is_training:
aug_det = aug.to_deterministic()
all_img = aug.augment_images(all_img)
label = aug.augment_images(label, hooks=mask_hook)
#all_img = preprocess_input(all_img, mode = "torch")
all_img = preprocess_input(all_img)
yield all_img, label
def ResNet50_predict_labels(img_path):
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))
img_path = sys.argv[1]
try:
# load YAML and create model
yaml_file = open('model.yaml', 'r')
loaded_model_yaml = yaml_file.read()
yaml_file.close()
model = model_from_yaml(loaded_model_yaml)
# load weights into new model
model.load_weights("model.h5")
print("Loaded model from disk")
except:
model = ResNet50(weights='imagenet')
# serialize model to YAML
model_yaml = model.to_yaml()
with open("model.yaml", "w") as yaml_file:
yaml_file.write(model_yaml)
# serialize weights to HDF5
model.save_weights("model.h5")
print("Saved model to disk")
img = image.load_img(img_path, target_size=(224, 224))
x = image.img_to_array(img)
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
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])
def get_features(img, model):
img_data = img_to_array(img.resize((224, 224)))
img_data = expand_dims(img_data, axis=0)
img_data = preprocess_input(img_data)
return model.predict(img_data)
def ResNet50_predict_labels(img_path):
# returns prediction vector for image located at img_path
img = preprocess_input(path_to_tensor(img_path))
return np.argmax(ResNet50_model.predict(img))