def preprocessing_image(img_path):
im = image.load_img(img_path, target_size=(224, 224, 3))
im = image.img_to_array(im)
im = np.expand_dims(im, axis=0)
im = preprocess_input(im)
return im
def generate_dataset_main(n=10000,
save_path=None,
seed=None,
model_res=1024,
image_size=256,
minibatch_size=16,
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, backend = keras.backend, layers = keras.layers, models = keras.models, utils = keras.utils)
X = preprocess_input(X)
return W, X
def prepare_image(raw_image):
img = cv2.resize(raw_image, dsize=(224, 224))
img = image.img_to_array(img)
img = np.expand_dims(img, axis=0)
img = preprocess_input(img)
return img
def rgb_to_gray(img):
x = preprocess_input(img)
gray_pic = np.dot(x[..., :3], [0.299, 0.587, 0.114])
final_pic = np.repeat(gray_pic[:, :, np.newaxis], 3, axis=2)
return final_pic
metadata[r[0]] = {"coord": [r[2], r[3]], "score": trueskill.Rating()}
# Load and preprocess images to be scored
# In[8]:
print("loading images")
images = {}
for imName in metadata.keys():
im = cv2.imread(data_path + img_folder + imName + ".jpg")
if im is not None:
im = cv2.resize(im, (224, 224), interpolation=cv2.INTER_CUBIC)
im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
images[imName] = preprocess_input(im)
# Keep only the "mu" value of the trueskill scores
# In[9]:
# Experimental: compare the analyzed images
print("Loading left and right")
i = 0
endi = len(images) / length
for l in range(0, len(images), length):
leftIds = [
tId for tId in random.sample(images.keys(), nComp)
for u in range(0, length)
]
t0 = time()
trained_model_path = "new_trained_from_resnet50_jun1"
model = tf.keras.models.load_model(trained_model_path)
t = gmtime(time() - t0)
print("Model Load Done in", strftime("%H:%M:%S", t))
while True:
try:
frame = get_frame_from_UDP()
image = frame.reshape(-1, 224, 224, 3)
x = resnet50.preprocess_input(image)
y = model.predict(x)
print(y, y.argmax())
if len(y) == 1:
prob_blocked = y[0][0]
if prob_blocked < 0.5:
cv.putText(frame, "Blocked", (10, 20), cv.FONT_HERSHEY_COMPLEX, 0.5, (0, 0, 255), 1)
else:
cv.putText(frame, "Free", (10, 20), cv.FONT_HERSHEY_COMPLEX, 0.5, (0, 255, 0), 1)
cv.imshow("jetbot", frame)
if cv.waitKey(1)==ord('q') :
break
except Exception as e:
print(e.args[0])
idx for idx in glob.glob(
'/home/VinBigData_ChestXray/data_classify/4_10_11_12_single_class_ver2_stage1_test_cropped/*.png'
)
]
classes_lst = []
class2id = {0: '4', 1: '10', 2: '11', 3: '12'}
RESULTS_PATH = '/home/VinBigData_ChestXray/data_classify/results_4_10_11_12_single_class_ver2_stage1_test_cropped/'
create_folder(RESULTS_PATH)
create_folder(os.path.join(RESULTS_PATH, '4'))
create_folder(os.path.join(RESULTS_PATH, '10'))
create_folder(os.path.join(RESULTS_PATH, '11'))
create_folder(os.path.join(RESULTS_PATH, '12'))
for image_idx in test_images:
img = image.load_img(image_idx, target_size=(1024, 1024))
img_array = image.img_to_array(img)
img_batch = np.expand_dims(img_array, axis=0)
img_preprocessed = preprocess_input(img_batch)
prediction = model.predict(img_preprocessed)
classes = class2id[np.argmax(prediction[0])]
print('Saving {} as {}'.format(
image_idx.split('/')[-1],
classes + '_' + image_idx.split('/')[-1]))
shutil.copyfile(
image_idx,
os.path.join(RESULTS_PATH, classes,
classes + '_' + image_idx.split('/')[-1]))
print('Length of test images', len(classes_lst))
# One-hot編碼
Y_train = to_categorical(Y_train, 10)
Y_test = to_categorical(Y_test, 10)
# 載入 ResNet50 模型
resnet_model = ResNet50(weights="imagenet",
include_top=False,
input_shape=(200, 200, 3))
# 調整X_train的圖片尺寸
print("調整X_train的圖片尺寸...")
X_train_new = np.array([
np.asarray(Image.fromarray(X_train[i]).resize((200, 200)))
for i in range(0, len(X_train))
])
X_train_new = X_train_new.astype("float32")
# 訓練資料的資料前處理
train_input = preprocess_input(X_train_new)
# 使用 ResNet50 模型預測訓練資料的特徵資料
print("使用 ResNet50 模型預測訓練資料的特徵資料...")
train_features = resnet_model.predict(train_input)
# 調整X_test的圖片尺寸
print("調整X_test的圖片尺寸...")
X_test_new = np.array([
np.asarray(Image.fromarray(X_test[i]).resize((200, 200)))
for i in range(0, len(X_test))
])
X_test_new = X_test_new.astype("float32")
# 測試資料的資料前處理
test_input = preprocess_input(X_test_new)
# 使用 ResNet50 模型預測測試資料的特徵資料
print("使用 ResNet50 模型預測測試資料的特徵資料...")
test_features = resnet_model.predict(test_input)
from tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions
import base64
tf.compat.v1.disable_v2_behavior() # 以静态图的方式运行
# 输入是string类型
input_imgs = tf.compat.v1.placeholder(shape=None, dtype=tf.string)
# 把base64字符串图像解码成jpeg格式
decoded = tf.image.decode_jpeg(tf.compat.v1.decode_base64(input_imgs), channels=3)
# 用最近邻法调整图像大小到[224,224],因为ResNet50需要输入图像大小是[224,224]
decoded = tf.compat.v1.image.resize_images(decoded, [224, 224], tf.image.ResizeMethod.NEAREST_NEIGHBOR)
# 在0位置增加一个值是1的维度，使其成为一个图像
tensorimg = tf.expand_dims(tf.cast(decoded, dtype=tf.float32), 0)
tensorimg = preprocess_input(tensorimg) # 图像预处理
with tf.compat.v1.Session() as sess: # 构建一个会话
sess.run(tf.compat.v1.global_variables_initializer())
# 加载ResNet50模型
Reslayer = ResNet50(weights='resnet50_weights_tf_dim_ordering_tf_kernels.h5')
logits = Reslayer(tensorimg) # 获取模型的输出节点
# 得到该图片的每个类别的概率值
prediction = tf.squeeze(tf.cast(tf.argmax(logits, 1), dtype=tf.int32), [0])
img_path = './dog.jpg' # 定义测试图片路径
with open(img_path, "rb") as image_file:
# 把图像编码成base64字符串格式
encoded_string = str(base64.urlsafe_b64encode(image_file.read()), "utf-8")
'''
# In[9]:
'''
img_generator = image_dataset_from_directory('C:\\Users\\calvom\\ThesisMaster\\notebooks\\downloaded_data\\training_top_23_0.1_resize_672_672',
batch_size=32,
image_size=(224, 224),
shuffle=True,
label_mode='categorical')
'''
base_model = ResNet50(weights='imagenet', include_top=False)
#base_model = MobileNetV2(weights='imagenet', include_top=False)
input_img = Input((224, 224, 3))
x = preprocess_input(input_img)
#x = tf.keras.applications.mobilenet_v2.preprocess_input(input_img)
encoded_features = base_model(x, training=False)
x3 = GlobalAveragePooling2D()(encoded_features)
x4 = Dense(1024, activation='relu')(x3)
x5 = Dropout(0.4)(x4)
predictions = Dense(
n_classes,
activation='softmax',
name='softmax',
kernel_regularizer=l2(0.01),
bias_regularizer=l2(0.01),
image_path =IMAGES_DIR / 'lena.jpg'
image_path =IMAGES_DIR / 'military-raptor.jpg' # warplane
image_path =IMAGES_DIR / 'baboon.png' # baboon
image_path =IMAGES_DIR / 'fighter_jet.jpg' # warplane
image_path =IMAGES_DIR / 'watch.png' # stopwatch
image_path =IMAGES_DIR / 'bear.tif' # brown_bear
image_path =IMAGES_DIR / 'wild_flowers.tif' # greenhouse
print (image_path)
# read image in RGB format
image = imageio.imread(str(image_path))
print(image.shape)
image = vision.resize(image, 224, 224)
batch = expand_to_batch(image)
batch = preprocess_input(batch)
b_model = resnet.model_resnet50(weights="imagenet")
b_y = b_model.predict(batch)
b_label = decode_predictions(b_y)
b_label = b_label[0][0]
# print the classification
print('OUR: %s (%.2f%%)' % (b_label[1], b_label[2]*100))
a_model = ResNet50(weights="imagenet")
a_y = a_model.predict(batch)
a_label = decode_predictions(a_y)
a_label = a_label[0][0]
# print the classification
print('KERAS: %s (%.2f%%)' % (a_label[1], a_label[2]*100))
def ImageEncode(img_path):
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)
return x
def process_image(img_path):
img = image.load_img(img_path, target_size=(512, 512))
x = image.img_to_array(img)
return preprocess_input(x)
# model.fit(x=x_train, y=y_train, epochs=3, callbacks=[checkpoint], verbose=0)
# print(get_test_accuracy(model, x_test, y_test))
#
# new_model = load_model(checkpoint_path)
# print(get_test_accuracy(new_model, x_test, y_test))
#
# model.save('my_model.h5')
# new_model1 = load_model('my_model.h5')
# print(get_test_accuracy(new_model1, x_test, y_test))
model = ResNet50(weights='imagenet', include_top=True)
from tensorflow.keras.preprocessing import image
img_input = image.load_img('my_picture.jpg', target_size=(224, 224))
img_input = image.img_to_array(img_input)
from tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions
preprocess_input(img_input[np.newaxis, ...])
preds = model.predict(img_input)
decode_predictions = decode_predictions(preds, top=3)[0]
from tensorflow.keras.applications import ResNet50
model = ResNet50(weights='imagenet')
from tensorflow.keras.preprocessing.image import load_img
lemon_img = load_img('lemon.jpg', target_size=(224, 224))
viaduct_img = load_img('viaduct.jpg', target_size=(224, 224))
water_tower_img = load_img('water_tower.jpg', target_size=(224, 224))
# Useful function: presents top 5 predictions and probabilities
from tensorflow.keras.preprocessing.image import img_to_array
def _read_image(self, image):
full_img_path = path.join(self._basepath, "all_images", image)
img = pil_image.open(full_img_path).convert("RGB")
img = img.resize((224, 224), pil_image.BILINEAR)
return preprocess_input(np.array(img, dtype=np.float32))
end = time.time()
time_taken = end - start
print('ResNet50 loading time: ', time_taken, " seconds")
all_images = np.zeros((len(names), 224, 224, 3))
start = time.time()
for i, img_path in enumerate(names):
img = image.load_img(img_path, target_size=(224, 224))
img2 = np.expand_dims(image.img_to_array(img), axis=0)
all_images[i, :, :, :] = img2
end = time.time()
time_taken = end - start
print('Image loading time: ', time_taken, " seconds")
start = time.time()
all_images = preprocess_input(all_images)
time_taken = end - start
print('Preprocess time: ', time_taken, " seconds")
start = time.time()
predictions = model.predict(all_images)
del all_images
end = time.time()
time_taken = end - start
print('Feature extraction time: ', time_taken, " seconds")
predictions = predictions.reshape((predictions.shape[0], -1))
pca_comp = 62
start = time.time()
pca = PCA(n_components=pca_comp)
pca_predictions = pca.fit_transform(predictions)
end = time.time()
def _read_image(self, image_path):
img = pil_image.open(image_path).convert("RGB")
img = img.resize((224, 224), pil_image.BILINEAR)
return preprocess_input(np.array(img, dtype=np.float32))
def build_model(self, data_sources: Dict[str, BaseDataSource], mode: str):
"""Build model."""
data_source = next(iter(data_sources.values()))
input_tensors = data_source.output_tensors
# stack left-eye and right-eye to get a 6 channels tensors, placeholder
# le = input_tensors['left-eye']
# re = input_tensors['right-eye']
le = input_tensors['left-eye'] # shape (60 224 3)
re = input_tensors['right-eye'] # shape (60 224 3)
face = input_tensors['face'] #shape (224, 224, 3)
le = preprocess_input(le) # add one dimension for batch
re = preprocess_input(re) # add one dimension for batch
face = preprocess_input(face) # add one dimension for batch
# RN = resNet50.ResNet50
# le_conv = RN(weights='imagenet', include_top=False, input_tensor = le)
le_conv = RN(weights='imagenet', include_top=False, input_tensor = le)
# re_conv = RN(weights='imagenet', include_top=False, input_tensor = re)
re_conv = RN(weights='imagenet', include_top=False, input_tensor = re)
face_conv = RN(weights='imagenet', include_top=False, input_tensor = face)
le_flatten = Flatten()(le_conv.output)
re_flatten = Flatten()(re_conv.output)
face_flatten = Flatten()(face_conv.output)
# for face
# face_flatten = Dense(2048, activation='relu', name='face_1')(face_flatten)
face_flatten = Dense(1000, activation='relu', name='face_3')(face_flatten)
# for eye
# le_flatten = Dense(2048, activation='relu', name='le_3')(le_flatten)
le_flatten = Dense(1000, activation='relu', name='le_4')(le_flatten)
# re_flatten = Dense(2048, activation='relu', name='le_3')(re_flatten)
re_flatten = Dense(1000, activation='relu', name='le_4')(re_flatten)
# for face landmark
# add face landmarks(coordinates)
# add head pose into the vector
# x = tf.concat((le_flatten, re_flatten, face_flatten, lm, input_tensors['head']), axis = 1)
x = tf.concat((le_flatten, re_flatten, face_flatten, input_tensors['head']), axis = 1)
x = BatchNormalization()(x)
# x = tf.concat((le_flatten, re_flatten, face_flatten), axis = 1)
x = Dense(1024, activation='relu', name='fc_3')(x)
# x = Dense(1024, activation='relu', name='fc_4')(x)
x = Dense(512, activation='relu', name='fc_6')(x)
x = Dense(128, activation='relu', name='fc_7')(x)
preds = Dense(2,name='preds')(x)
# eye_model = Model(lre, preds)
# Define outputs
loss_terms = {}
metrics = {}
if 'gaze' in input_tensors:
y = input_tensors['gaze']
with tf.variable_scope('mse'): # To optimize
# NOTE: You are allowed to change the optimized loss
loss_terms['gaze_mse'] = tf.reduce_mean(tf.squared_difference(preds, y))
# loss_terms['gaze_mse'] = util.gaze.tensorflow_angular_error_from_pitchyaw(preds, y)
with tf.variable_scope('ang'): # To evaluate in addition to loss terms
metrics['gaze_angular'] = util.gaze.tensorflow_angular_error_from_pitchyaw(preds, y)
return {'gaze': preds}, loss_terms, metrics # are graphs to be executed
from sklearn.preprocessing import LabelBinarizer
from random import shuffle
# select which GPU to use for training
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
### Params ###
# Set here the basic params to choose which data to use for the training
path = "images-atmos/" # The folder where the images are located
cspath = "crowdsourcing/atmosphere-final.json" # The file where the crowdsourcing results are located
output_path = "models/full_model_atmos.h5" # The file where the trained model will be saved
# Load all images
tempImages = {}
for f in os.listdir(path):
tempImages[f[:-4]] = preprocess_input(cv2.imread(path + f))
# Shuffle dataset, to have it more balanced
def shuffle_data(left, right, lab):
shuffled = list(zip(left, right, lab))
shuffle(shuffled)
return zip(*shuffled)
lb = LabelBinarizer()
imgsLeft = []
imgsRight = []
labels = []
batchPaths = imagePaths[i:i + bs]
batchLabels = labels[i:i + bs]
batchImages = []
# loop over the images and labels in the current batch
for imagePath in batchPaths:
# load the input image using the Keras helper utility while
# ensuring the image is resized to 224x224 pixels
image = load_img(imagePath, target_size=(224, 224))
image = img_to_array(image)
# preprocess the image by (1) expanding the dimensions and
# (2) subtracting the mean RGB pixel intensity from the
# ImageNet dataset
image = np.expand_dims(image, axis=0)
image = preprocess_input(image)
# add the image to the batch
batchImages.append(image)
# pass the images through the network and use the outputs as our
# actual features, then reshape the features into a flattened
# volume
batchImages = np.vstack(batchImages)
features = model.predict(batchImages, batch_size=bs)
features = features.reshape((features.shape[0], 7 * 7 * 2048))
# loop over the class labels and extracted features
for (label, vec) in zip(batchLabels, features):
# construct a row that exists of the class label and extracted
# features
def preprocessing(image):
# Pre-processing
input_tensor = preprocess_input(np.expand_dims(image, axis=0))
return input_tensor
# -*- coding: utf-8 -*-
"""
Created on Thu Sep 17 15:11:45 2020
@author: ASELSAN
title: RESNET50 MİMARİSİ İLE HAYVAN SINIFLANDIRMA
STAJDAN BAĞIMSIZ BİR ÖRNEK OLDU AMA KALABİLİR :)
"""
from tensorflow.keras.applications.resnet50 import ResNet50
from tensorflow.keras.preprocessing import image
from tensorflow.keras.applications.resnet50 import preprocess_input, decode_predictions
import numpy as np
model = ResNet50(weights='imagenet')
img_path = 'tiger.jpg'
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)
#print(preds)
# 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: [(u'n02504013', u'Indian_elephant', 0.82658225), (u'n01871265', u'tusker', 0.1122357), (u'n02504458', u'African_elephant', 0.061040461)]
NUM_LOOPS_PER_THREAD = 100
COMPILED_MODEL_DIR = "./rn50_fp16_compiled_b" + str(args.batch_size) + "_nc" + str(args.num_neuroncores) + "/1"
# Ensure there's enough buffer capacity to hold in-flight requests in runtime
NUM_INFERS_IN_FLIGHT = args.num_neuroncores + 1
os.environ['NEURON_MAX_NUM_INFERS'] = str(NUM_INFERS_IN_FLIGHT)
num_groups = avail_neuroncores // args.num_neuroncores
group_sizes = [str(args.num_neuroncores)] * num_groups
os.environ['NEURONCORE_GROUP_SIZES'] = ','.join(group_sizes)
# Create input from image
img_sgl = image.load_img('kitten_small.jpg', target_size=(224, 224))
img_arr = image.img_to_array(img_sgl)
img_arr2 = np.expand_dims(img_arr, axis=0)
img_arr3 = resnet50.preprocess_input(np.repeat(img_arr2, USER_BATCH_SIZE, axis=0))
# Load model
NUM_THREADS_PER_PREDICTOR = args.num_neuroncores + 1
pred_list = [tf.contrib.predictor.from_saved_model(COMPILED_MODEL_DIR) for _ in range(num_groups)]
pred_list = pred_list * NUM_THREADS_PER_PREDICTOR
num_threads = len(pred_list)
num_infer_per_thread = []
tot_latency_per_thread = []
thread_active = []
for i in range(num_threads):
num_infer_per_thread.append(0)
tot_latency_per_thread.append(0)
thread_active.append(0)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("-d", "--dataset", required=True, help="Path to dataset")
parser.add_argument("--pca", default=None, help="Path to PCA model")
parser.add_argument("--scaler", default=None, help="Path to Scaler model")
parser.add_argument("-o", "--output", required=True, help="Output HDF5 file")
parser.add_argument("-b", "--buffer", type=int, default=1000, help="Buffer size for feature extraction")
args = parser.parse_args()
# Get labels
img_paths = sorted(glob.glob(f"{args.dataset}/*.jpg"))
labels = [int(os.path.basename(img_path)[0]) for img_path in img_paths]
labels = np.array(labels)
batch_size = 32
# Feature extraction model
model = ResNet50(weights="imagenet", include_top=False)
features = []
for i in tqdm.tqdm(np.arange(0, len(img_paths), batch_size)):
# Get batch
batch_img_paths = img_paths[i:i+batch_size]
# Process each path
batch_images = []
for img_path in batch_img_paths:
image = load_img(img_path, target_size=(224, 224))
image = img_to_array(image)
image = np.expand_dims(image, axis=0)
image = preprocess_input(image)
batch_images.append(image)
batch_images = np.vstack(batch_images)
# Extract features
batch_features = model.predict(batch_images, batch_size=batch_size)
batch_features = batch_features.reshape((batch_features.shape[0], -1))
features.append(batch_features)
features = np.vstack(features)
print(f"[INFO] Extracted features: {features.shape}")
# Load models
scaler = joblib.load(args.scaler) if args.scaler else StandardScaler().fit(features)
features = scaler.transform(features)
pca = joblib.load(args.pca) if args.pca else PCA(n_components=1024, whiten=True).fit(features)
features = pca.transform(features)
print(f"[INFO] PCA features: {features.shape}")
print(f"[INFO] PCA Explained variance %: {sum(pca.explained_variance_ratio_)}")
# Save models
os.makedirs("models", exist_ok=True)
if not args.scaler:
print("[INFO] Saving Scaler model ...")
joblib.dump(scaler, "models/scaler.pkl")
if not args.pca:
print("[INFO] Saving PCA model ...")
joblib.dump(pca, "models/pca.pkl")
# Write to hdf5
dims = (len(img_paths), 1024)
dataset = HDF5Writer(args.output, dims, buffer_size=args.buffer)
dataset.store_class_names(["Not Food", "Food"])
for i in tqdm.tqdm(np.arange(0, len(img_paths), batch_size)):
# Add to dataset
batch_features = features[i:i+batch_size]
batch_labels = labels[i:i+batch_size]
dataset.add(batch_features, batch_labels)
dataset.close()
name: sess.graph.get_tensor_by_name(ts_name)
for name, ts_name in input_names.items()
}
outputs = {
name: sess.graph.get_tensor_by_name(ts_name)
for name, ts_name in output_names.items()
}
tf.saved_model.simple_save(sess, model_dir, inputs, outputs)
SAVED_MODEL_DIR = './rn50_fp16'
shutil.rmtree(SAVED_MODEL_DIR, ignore_errors=True)
input_tname = "{}:0".format(args.input)
output_tname = "{}:0".format(args.output)
pb_to_saved_model(args.graph, {input_tname: input_tname},
{output_tname: output_tname}, SAVED_MODEL_DIR)
# Create input from image
img_sgl = image.load_img('kitten_small.jpg', target_size=(224, 224))
img_arr = image.img_to_array(img_sgl)
img_arr2 = np.expand_dims(img_arr, axis=0)
img_arr3 = resnet50.preprocess_input(np.repeat(img_arr2, 1, axis=0))
# Load model
predictor_host = tf.contrib.predictor.from_saved_model(SAVED_MODEL_DIR)
# Run inference
model_feed_dict = {'input_1:0': img_arr3}
infa_rslts = predictor_host(model_feed_dict)
print(resnet50.decode_predictions(infa_rslts[output_tname], top=5)[0])
# looping over region proposal bounding boxes coordinates generated by running selective search
for (x, y, w, h) in rects:
# filtering the unusable boxes
# filtering the boxes with size less than 10% of the size of the image
if w / float(W) < 0.1 or h / float(H) < 0.1:
continue
# extracting regions of interest from the input image and converting it from BGR to RGB
# resizing it 224x224 shape, required by the classifier
roi = image[y:y + h, x:x + w]
roi = cv2.cvtColor(roi, cv2.COLOR_BGR2RGB)
roi = cv2.resize(roi, (224, 224))
# further preprocessing required
roi = img_to_array(roi)
roi = preprocess_input(roi)
#updating list of bounding boxes and proposals
proposals.append(roi)
boxes.append((x, y, w, h))
# converting the list of proposals to a numpy array
proposals = np.array(proposals)
print("[INFO] proposals shape {}".format(proposals.shape))
# classifying each proposal ROIs using ResNet and decoding the predictions
print("[INFO] classifying proposals...")
preds = model.predict(proposals)
preds = imagenet_utils.decode_predictions(preds, top=1)
# initializing a dictionary to map class labels to bounding box associated with it
labels = {}
for item in (PATH_DATA / 'validation').glob('**/*'):
if not item.is_dir():
images_validation_path.append(item)
n_images_validation = 3
images_validation_path = random.sample(images_validation_path,
n_images_validation)
images_validation = [
Image.open(image_validation_path)
for image_validation_path in images_validation_path
]
# Remember to preprocess the input!
image_batch = list()
image_batch = np.stack([
preprocess_input(np.array(image_validation.resize((224, 224))))
for image_validation in images_validation
])
prediction_probabilities = model_tl.predict(image_batch)
fig, axes = plt.subplots(1, n_images_validation, figsize=(15, 5))
for ii, image in enumerate(images_validation):
axes[ii].imshow(image)
axes[ii].set_title("Thanos {:.1f}%, Grimace {:.1f}%".format(
100 * prediction_probabilities[ii, 0],
100 * (1 - prediction_probabilities[ii, 0])))
axes[ii].axis('off')
# %% [markdown]
# ### Bonus: Grimos
# Lastly, I want to try the model with images that I consider specially
def read_and_prep_image(img_path, img_height = image_size, img_width = image_size):
img = load_img(img_path, target_size = (img_height, img_width))
img_array = np.array([img_to_array(img)])
output = preprocess_input(img_array)
return(output)
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 dogornot(img):
img = preprocess_input(img)
img = np.array([img])
return DOgOrNot[np.argsort(-(model.predict(img)))[0][0]]