def predict(image_file):
image = preprocess_image(image_file, [256, 256, 3])
# Training
model = tflearn.DNN(resnet())
model.load(pjoin(MODEL_PATH, MODEL_NAME))
y_pred = model.predict([image])
label = np.argmax(y_pred[0])
return 'Cat' if label == 0 else 'Dog'
def predict(self, img):
img = image_utils.preprocess_image(img)
input_data = img[
np.newaxis, :, :, :] # Append batch dimension for keras
prediction = self.model.predict(input_data)[0]
top5 = self.get_top5(prediction)
return top5
def classify(self, data):
image = self.convertImg(data)
image = preprocess_image(image, dbg=False)
#if not self.hasSaved:
# self.hasSaved = True
# np.save(self.encoding,image)
# Run classifier
shouldTurn = False
pub.publish(shouldTurn)
def predict_cropped(self, img):
img = image_utils.preprocess_image(img)
cropped = image_utils.crop_split(img)
cropped.append(img)
input_data = np.stack(cropped, axis=0)
predictions = self.model.predict(input_data)
avg_predictions = np.mean(predictions, axis=0)
top5 = self.get_top5(avg_predictions)
return top5
def recognize_face(self, img, conf=0.6):
with self.graph.as_default():
with tf.Session(graph=self.graph) as sess:
# Get facenet embeddings
emb_array = np.zeros((1, self.embedding_size))
image = np.array([preprocess_image(img, False, False, 160)])
feed_dict = {
self.images_placeholder: image,
self.phase_train_placeholder: False
}
emb_array[0, :] = sess.run(self.embeddings,
feed_dict=feed_dict)
print(emb_array)
def classify(self, data):
image = self.convertImg(data)
image = preprocess_image(image, dbg=False)
print(image.shape)
raw_input('image.shape')
shouldTurn, duration = self._classifier.predict(image, verbose=True)
#if not self.hasSaved:
# self.hasSaved = True
# np.save(self.encoding,image)
# Run classifier
# shouldTurn = False
pub.publish(shouldTurn)
def read_data():
X = []
Y = []
for f in glob.glob(TRAIN_DATA + '/*.jpg'):
fname = os.path.basename(f)
# 0 for cat, 1 for dog
label = 0 if fname.startswith('cat') else 1
image = preprocess_image(f, [256, 256, 3])
X.append(image)
Y.append(label)
# split training data and validation set data
X, X_test, y, y_test = train_test_split(X,
Y,
test_size=0.2,
random_state=42)
return (X, y), (X_test, y_test)
def encode_image(
self,
image_file: Optional[Union[PurePath, str]] = None,
image_array: Optional[np.ndarray] = None,
) -> np.ndarray:
"""
Generate CNN encoding for a single image.
Args:
image_file: Path to the image file.
image_array: Optional, used instead of image_file. Image typecast to numpy array.
Returns:
encoding: Encodings for the image in the form of numpy array.
Example:
```
from imagededup.methods import CNN
myencoder = CNN()
encoding = myencoder.encode_image(image_file='path/to/image.jpg')
OR
encoding = myencoder.encode_image(image_array=<numpy array of image>)
```
"""
if isinstance(image_file, str):
image_file = Path(image_file)
if isinstance(image_file,
(PurePath, urllib.request.http.client.HTTPResponse)):
if isinstance(image_file, PurePath) and not image_file.is_file():
raise ValueError(
"Please provide either image file path or image array!")
image_pp = load_image(image_file=image_file,
target_size=self.target_size,
grayscale=False)
elif isinstance(image_array, np.ndarray):
image_pp = preprocess_image(image=image_array,
target_size=self.target_size,
grayscale=False)
else:
raise ValueError(
"Please provide either image file path or image array!")
return (self._get_cnn_features_single(image_pp) if isinstance(
image_pp, np.ndarray) else None)
def batched_calibration_fn():
"""function to pass as argument to the builder function
it will be used to optimize the network based on given examples """
print("[CALIBRATION] Starting calibration process...")
images_found = sorted(os.listdir(IMAGES_PATH))
print("[CALIBRATION] Obtaining calibration images from {}"
.format(images_found))
print("[CALIBRATION] Done! Found {} images for calibration"
.format(len(images_found)))
print("[CALIBRATION] Starting image yielding...")
start_calibration = time.time()
batched_input = np.zeros((len(images_found), INPUT_SIZE, INPUT_SIZE, 3), dtype=np.uint8)
for input_value in range(len(images_found)):
# read and resize the image
input_image = imgutils.read_image_from_cv2(IMAGES_PATH + images_found[input_value])
image_data = imgutils.preprocess_image(input_image, (INPUT_SIZE, INPUT_SIZE))
# add a new axis to match requested shape
final_image = image_data[np.newaxis, ...].astype("uint8")
# add image to batched images
batched_input[input_value, :] = final_image
print("[CALIBRATION] Adding image {} from {}".format(input_value+1, images_found[input_value]))
start_adding = time.time()
print("[CALIBRATION] Image adding step Done, it took {} ms"
.format((time.time()-start_adding)*1000))
print("[CALIBRATION] Yielding batched input")
start_yielding = time.time()
yield (final_image,)
print("[CALIBRATION] Image yielding Done, it took {} ms"
.format((time.time()-start_yielding)*1000))
CALIBRATION_TIME = (time.time()-start_calibration)*1000
print("[CALIBRATION] Calibration procces finished, it took {} ms"
.format(CALIBRATION_TIME))
def DataGenerator(data_loader, batch_size=32, use_augment=True):
while True:
datas = []
labels = []
while len(datas) < batch_size:
im = data_loader.random_im()
img = data_loader.get_image(im)
label = data_loader.get_label(im)
img = image_utils.preprocess_image(img)
data = [img]
if use_augment:
augmentations = augment.get_augmentations(img)
data.extend(augmentations)
data = random.sample(data, 4)
datas.extend(data)
for i in xrange(len(data)):
labels.append(label)
data = np.stack(datas[:batch_size], axis=0)
label = np.stack(labels[:batch_size], axis=0)
yield (data, label)
def style_transfer(content_image,
style_image,
image_size,
style_size,
content_layer,
content_weight,
style_layers,
style_weights,
tv_weight,
init_random=False):
"""Run style transfer!
Inputs:
- content_image: filename of content image
- style_image: filename of style image
- image_size: size of smallest image dimension (used for content loss and generated image)
- style_size: size of smallest style image dimension
- content_layer: layer to use for content loss
- content_weight: weighting on content loss
- style_layers: list of layers to use for style loss
- style_weights: list of weights to use for each layer in style_layers
- tv_weight: weight of total variation regularization term
- init_random: initialize the starting image to uniform random noise
"""
# Extract features from the content image
content_img = preprocess_image(load_image(content_image, size=image_size))
feats = model.extract_features(model.image)
content_target = sess.run(feats[content_layer],
{model.image: content_img[None]})
# Extract features from the style image
style_img = preprocess_image(load_image(style_image, size=style_size))
style_feat_vars = [feats[idx] for idx in style_layers]
style_target_vars = []
# Compute list of TensorFlow Gram matrices
for style_feat_var in style_feat_vars:
style_target_vars.append(gram_matrix(style_feat_var))
# Compute list of NumPy Gram matrices by evaluating the TensorFlow graph on the style image
style_targets = sess.run(style_target_vars, {model.image: style_img[None]})
# Initialize generated image to content image
if init_random:
img_var = tf.Variable(tf.random_uniform(content_img[None].shape, 0, 1),
name="image")
else:
img_var = tf.Variable(content_img[None], name="image")
# Extract features on generated image
feats = model.extract_features(img_var)
# Compute loss
c_loss = content_loss(content_weight, feats[content_layer], content_target)
s_loss = style_loss(feats, style_layers, style_targets, style_weights)
t_loss = tv_loss(img_var, tv_weight)
loss = c_loss + s_loss + t_loss
# Set up optimization hyperparameters
initial_lr = 3.0
decayed_lr = 0.1
decay_lr_at = 180
max_iter = 200
# Create and initialize the Adam optimizer
lr_var = tf.Variable(initial_lr, name="lr")
# Create train_op that updates the generated image when run
with tf.variable_scope("optimizer") as opt_scope:
train_op = tf.train.AdamOptimizer(lr_var).minimize(loss,
var_list=[img_var])
# Initialize the generated image and optimization variables
opt_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=opt_scope.name)
sess.run(tf.variables_initializer([lr_var, img_var] + opt_vars))
# Create an op that will clamp the image values when run
clamp_image_op = tf.assign(img_var, tf.clip_by_value(img_var, -1.5, 1.5))
f, axarr = plt.subplots(1, 2)
axarr[0].axis('off')
axarr[1].axis('off')
axarr[0].set_title('Content Source Img.')
axarr[1].set_title('Style Source Img.')
axarr[0].imshow(deprocess_image(content_img))
axarr[1].imshow(deprocess_image(style_img))
plt.savefig('orig_imgs')
plt.show()
plt.figure()
# Hardcoded handcrafted
for t in range(max_iter):
# Take an optimization step to update img_var
sess.run(train_op)
if t < decay_lr_at:
sess.run(clamp_image_op)
if t == decay_lr_at:
sess.run(tf.assign(lr_var, decayed_lr))
if t % 100 == 0:
print('Iteration {}'.format(t))
img = sess.run(img_var)
plt.imshow(deprocess_image(img[0], rescale=True))
plt.axis('off')
plt.savefig('iteration{}'.format(t))
plt.show()
print('Iteration {}'.format(t))
img = sess.run(img_var)
plt.imshow(deprocess_image(img[0], rescale=True))
plt.axis('off')
plt.savefig('style_transfer.png')
plt.show()
#Shallow feature reconstruction
filename = 'kitten.jpg'
layer = 3 # layers start from 0 so these are features after 4 convolutions
img = imresize(imread(filename), (64, 64))
plt.imshow(img)
plt.gcf().set_size_inches(3, 3)
plt.title('Original image')
plt.axis('off')
plt.show()
# Preprocess the image before passing it to the network:
# subtract the mean, add a dimension, etc
img_pre = preprocess_image(img, data['mean_image'])
# Extract features from the image
feats, _ = model.forward(img_pre, end=layer)
# Invert the features
kwargs = {
'num_iterations': 400,
'learning_rate': 5000,
'l2_reg': 1e-8,
'show_every': 100,
'blur_every': 10,
}
X = invert_features(feats, layer, model, **kwargs)
#deep feature reconstruction
def style_transfer(content_img_path,
img_size,
style_img_path,
style_size,
content_layer,
content_weight,
style_layers,
style_weights,
tv_weight,
init_random=False):
"""Perform style transfer from style image to source content image
Args:
content_img_path (str): File location of the content image.
img_size (int): Size of the smallest content image dimension.
style_img_path (str): File location of the style image.
style_size (int): Size of the smallest style image dimension.
content_layer (int): Index of the layer to use for content loss.
content_weight (float): Scalar weight for content loss.
style_layers ([]int): Indices of layers to use for style loss.
style_weights ([]float): List of scalar weights to use for each layer in style_layers.
tv_weigh (float): Scalar weight of total variation regularization term.
init_random (boolean): Whether to initialize the starting image to uniform random noise.
"""
tf.reset_default_graph()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
try:
model = SqueezeNet(ckpt_path=CKPT_PATH, sess=sess)
except NotFoundError:
raise ValueError('checkpoint file is not found, please check %s' %
CKPT_PATH)
# Extract features from content image
content_img = preprocess_image(load_image(content_img_path, size=img_size))
content_feats = model.extract_features(model.image)
# Create content target
content_target = sess.run(content_feats[content_layer],
{model.image: content_img[None]})
# Extract features from style image
style_img = preprocess_image(load_image(style_img_path, size=style_size))
style_feats_by_layer = [content_feats[i] for i in style_layers]
# Create style targets
style_targets = []
for style_feats in style_feats_by_layer:
style_targets.append(gram_matrix(style_feats))
style_targets = sess.run(style_targets, {model.image: style_img[None]})
if init_random:
generated_img = tf.Variable(tf.random_uniform(content_img[None].shape,
0, 1),
name="image")
else:
generated_img = tf.Variable(content_img[None], name="image")
# Extract features from generated image
current_feats = model.extract_features(generated_img)
loss = content_loss(content_weight, current_feats[content_layer], content_target) + \
style_loss(current_feats, style_layers, style_targets, style_weights) + \
total_variation_loss(generated_img, tv_weight)
# Set up optimization parameters
init_learning_rate = 3.0
decayed_learning_rate = 0.1
max_iter = 200
learning_rate = tf.Variable(init_learning_rate, name="lr")
with tf.variable_scope("optimizer") as opt_scope:
train_op = tf.train.AdamOptimizer(learning_rate).minimize(
loss, var_list=[generated_img])
opt_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope=opt_scope.name)
sess.run(
tf.variables_initializer([learning_rate, generated_img] + opt_vars))
# Create an op that will clamp the image values when run
clamp_image_op = tf.assign(generated_img,
tf.clip_by_value(generated_img, -1.5, 1.5))
display_content_and_style(content_img, style_img)
for t in range(max_iter):
sess.run(train_op)
if t < int(0.90 * max_iter):
sess.run(clamp_image_op)
elif t == int(0.90 * max_iter):
sess.run(tf.assign(learning_rate, decayed_learning_rate))
if t % 20 == 0:
current_loss = sess.run(loss)
print 'Iteration %d: %f' % (t, current_loss)
img = sess.run(generated_img)
plt.imshow(deprocess_image(img[0], rescale=True))
plt.axis('off')
plt.show()
for patch_index, start in enumerate(starts):
print(
f"Processing patch {patch_index} of {len(starts)} for image {image_id}..."
)
end = start[0] + n0, start[1] + n1
patch_image = image[start[0]:end[0], start[1]:end[1], :]
downsampled_patch = resize(patch_image, (r0, r1),
anti_aliasing=config['image_antialiasing'],
order=3,
preserve_range=True)
downsampled_patch = preprocess_image(downsampled_patch, model_type)
downsampled_mask = np.squeeze(
model.predict(downsampled_patch[np.newaxis, :], batch_size=1))
# Upsample the predicted mask to match the original patch
patch_mask = resize(downsampled_mask, (n0, n1),
anti_aliasing=False,
order=3,
preserve_range=True)
# Plonk that back into the super mask, and update the count
mask_image[start[0]:end[0], start[1]:end[1]] += patch_mask
count_image[start[0]:end[0], start[1]:end[1]] += 1.0
mask_image = mask_image / count_image
def style_transfer(content_img,style_img,content_size,style_size,content_layer,style_layers,
content_weight,style_weights,tv_weight,init_random=False):
content_pre_img = preprocess_image(load_image(content_img,size=content_size))
feats = model.extract_features(model.image) #extract features of every layer from the input image
content_targets = sess.run(feats[content_layer],{model.image:content_pre_img[None]})
style_pre_img = preprocess_image(load_image(style_img,size = style_size))
style_feats = [feats[idx] for idx in style_layers]
#to transfer gram
style_target=[]
for style_feat_var in style_feats:
style_target.append(gram_matrix(style_feat_var))
style_targets = sess.run(style_target,{model.image:style_pre_img[None]})
if init_random:
img_var = tf.Variable(tf.random_uniform(content_pre_img[None].shape,0,1),name="image")
else:
img_var = tf.Variable(content_pre_img[None], name="image")
# to compute loss
#print(img_var[None].shape)
feat = model.extract_features(img_var)
conloss = content_loss(content_weight,feat[content_layer],content_targets)
styloss = style_loss(style_weights,feat,style_layers,style_targets)
tvloss = TV_loss(img_var,tv_weight)
loss = conloss+styloss+tvloss
#params
initial_lr = 3.0
decayed_lr = 0.1
decayed_lr_at = 180
max_iters = 200
lr_var = tf.Variable(initial_lr,name="lr")
# Create train_op that updates the generated image when run
with tf.variable_scope("optimizer") as opt_scope:
train_op = tf.train.AdamOptimizer(lr_var).minimize(loss,var_list=[img_var])
# Initialize the generated image and optimization variables
opt_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,scope=opt_scope.name)
sess.run(tf.variables_initializer([lr_var,img_var]+opt_vars))
# Create an op that will clamp the image values when run
clamp_image = tf.assign(img_var,tf.clip_by_value(img_var,-1.5,1.5))
#plot
f,s = plt.subplots(1,2)
s[0].axis('off')
s[1].axis('off')
s[0].set_title('content source img')
s[1].set_title('style source img')
s[0].imshow(deprocess_image(content_pre_img))
s[1].imshow(deprocess_image(style_pre_img))
plt.show()
plt.figure()
for i in range(max_iters):
#take a optimization step to update the img
sess.run(train_op)
if i < decayed_lr_at:
sess.run(clamp_image)
if i == decayed_lr_at:
sess.run(tf.assign(lr_var,decayed_lr))
if i % 100 ==0:
print('Iteration:{}'.format(i))
img = sess.run(img_var)
plt.imshow(deprocess_image(img[0],rescale=True))
plt.axis('off')
plt.show()
print('Iteration:{}'.format(i))
img = sess.run(img_var)
plt.imshow(deprocess_image(img[0],rescale=True))
plt.axis('off')
plt.show()
def check_scipy():
import scipy
vnum = int(scipy.__version__.split('.')[1])
assert vnum>=16,"You must install scipy >=0.16.0"
check_scipy()
from squeezenet import SqueezeNet
import tensorflow as tf
tf.reset_default_graph()
sess = get_session()
SAVE_PATH = 'datasets/squeezenet.ckpt'
print(SAVE_PATH)
#if not os.path.exists(SAVE_PATH):
#raise ValueError("You need to download SqueezeNet!")
model = SqueezeNet(save_path=SAVE_PATH, sess=sess)
#load data for testing
content_img_test = preprocess_image(load_image('tubingen.jpg',size=192))[None]
style_img_test = preprocess_image(load_image('starry_night.jpg',size=192))[None]
answers = np.load('style-transfer-checks.npz')
def content_loss(content_weight,content_curr,content_orig):
return content_weight*tf.reduce_sum(tf.squared_difference(content_curr,content_orig))
def gram_matrix(features,normalize=True):
"""Inputs: the shape of features is (1,H,W,C)"""
features = tf.transpose(features,[0,3,1,2])
shape = tf.shape(features)
features = tf.reshape(features,(shape[0],shape[1],-1))
transpose_features = tf.transpose(features,[0,2,1])
output = tf.matmul(features,transpose_features)
if normalize:
output = tf.div(output,tf.cast(shape[0]*shape[1]*shape[2]*shape[3],tf.float32))
def style_transfer(content_image, style_image, image_size, style_size, content_layer, content_weight,
style_layers, style_weights, tv_weight, model, sess, out_path, init_random = False,
max_iter = 100):
"""Run style transfer!
Inputs:
- content_image: filename of content image
- style_image: filename of style image
- image_size: size of smallest image dimension (used for content loss and generated image)
- style_size: size of smallest style image dimension
- content_layer: layer to use for content loss
- content_weight: weighting on content loss
- style_layers: list of layers to use for style loss
- style_weights: list of weights to use for each layer in style_layers
- tv_weight: weight of total variation regularization term
- init_random: initialize the starting image to uniform random noise
"""
# Extract features from the content image
content_img = preprocess_image(load_image(content_image, size=image_size))
feats = model.extract_features(model.image)
content_target = sess.run(feats[content_layer],
{model.image: content_img[None]})
# Extract features from the style image
style_img = preprocess_image(load_image(style_image, size=style_size))
style_feat_vars = [feats[idx] for idx in style_layers]
style_target_vars = []
# Compute list of TensorFlow Gram matrices
for style_feat_var in style_feat_vars:
style_target_vars.append(gram_matrix(style_feat_var))
# Compute list of NumPy Gram matrices by evaluating the TensorFlow graph on the style image
style_targets = sess.run(style_target_vars, {model.image: style_img[None]})
# Initialize generated image to content image
if init_random:
img_var = tf.Variable(tf.random_uniform(content_img[None].shape, 0, 1), name="image")
else:
img_var = tf.Variable(content_img[None], name="image")
# Extract features on generated image
feats = model.extract_features(img_var)
# Compute loss
c_loss = content_loss(content_weight, feats[content_layer], content_target)
s_loss = style_loss(feats, style_layers, style_targets, style_weights)
t_loss = tv_loss(img_var, tv_weight)
loss = c_loss + s_loss + t_loss
# Set up optimization hyperparameters
initial_lr = 3.0
decayed_lr = 0.1
decay_lr_at = 180
# Create and initialize the Adam optimizer
lr_var = tf.Variable(initial_lr, name="lr")
# Create train_op that updates the generated image when run
with tf.variable_scope("optimizer") as opt_scope:
train_op = tf.train.AdamOptimizer(lr_var).minimize(loss, var_list=[img_var])
# Initialize the generated image and optimization variables
opt_vars = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope=opt_scope.name)
sess.run(tf.variables_initializer([lr_var, img_var] + opt_vars))
# Create an op that will clamp the image values when run
clamp_image_op = tf.assign(img_var, tf.clip_by_value(img_var, -1.5, 1.5))
# Hardcoded handcrafted
for t in range(max_iter):
# Take an optimization step to update img_var
sess.run(train_op)
if t < decay_lr_at:
sess.run(clamp_image_op)
if t == decay_lr_at:
sess.run(tf.assign(lr_var, decayed_lr))
img = sess.run(img_var)
fig=plt.figure()
ax=fig.add_subplot(1,1,1)
plt.axis('off')
plt.imshow(deprocess_image(img[0], rescale=True))
extent = ax.get_window_extent().transformed(fig.dpi_scale_trans.inverted())
if out_path is not None:
plt.savefig(out_path, bbox_inches=extent, pad_inches=0)
else:
out_path = re.split("/|\.", content_image)[-2]
out_path = "output/" + out_path + "_out.jpg"
plt.savefig(out_path, bbox_inches=extent, pad_inches=0)
error = rel_error(correct, output)
print('Error is {}'.format(error))
def tv_loss_test(correct):
tv_weight = 2e-2
t_loss = tv_loss(model.image, tv_weight)
output = sess.run(t_loss, {model.image: content_img_test})
error = rel_error(correct, output)
print('Error is {}'.format(error))
if __name__ == '__main__':
tf.reset_default_graph() # remove all existing variables in the graph
sess = get_session() # start a new Session
model = load_model(sess)
content_img_test = preprocess_image(
load_image('styles/tubingen.jpg', size=192))[None]
style_img_test = preprocess_image(
load_image('styles/starry_night.jpg', size=192))[None]
answers = np.load('style-transfer-checks-tf.npz')
content_loss_test(answers['cl_out'])
gram_matrix_test(answers['gm_out'])
style_loss_test(answers['sl_out'])
tv_loss_test(answers['tv_out'])
