epochs = val_data.__len__() / hps['batch_size']
# epochs = 1
for i in range(epochs):
print("========== Running Epoch {} / {} =========".format(i, epochs))
test_data = val_data[i * hps['batch_size']:(i + 1) * hps['batch_size']]
image_ids = [sample['image_id'] for sample in test_data]
images_data = load_images(image_ids)
print("\t {} Images loaded".format(images_data.__len__()))
images_modified = [preprocess_image(im) for im in images_data]
features = [
CNN.extract_features(
input=tf.reshape(im, [1, 266, 400, 3]))[hps['feature_layer']]
for im in images_modified
]
features = [
tf.nn.max_pool(feat, [1, 4, 4, 1], [1, 4, 4, 1], 'SAME')
for feat in features
]
features = tf.stack(features, axis=0)
print("\t cnn features shape :::{}".format(features.shape))
features = tf.reshape(features, (features.shape[0], -1))
questions_vec = [
np.array(
question_to_vector(test_data[j]['question'], words_to_idx,
hps['enc_timesteps']))
for j in range(hps['batch_size'])
def build_graph(squeezenet_path,
cond='train',
param='initialize',
batch_size=1):
with tf.device('/cpu:0'):
with tf.Graph().as_default():
sess = get_session()
sess1 = tf.Session()
if cond == 'train':
SAVE_PATH = squeezenet_path #'C:/Users/Junaid/Desktop/style proj/datasets/squeezenet.ckpt'
model = SqueezeNet(save_path=SAVE_PATH, sess=sess1)
# print("kndflkn")
x = tf.placeholder(tf.float32, [batch_size, 256, 256, 3], name="x")
phase = tf.placeholder(tf.bool, name='phase')
style_image = tf.placeholder(tf.float32,
shape=[256, 256, 3],
name='style_image')
keep_prob = tf.placeholder(tf.float32, name="keep_prob")
padding = [[0, 0], [40, 40], [40, 40], [0, 0]]
x_new = tf.pad(x, padding, 'REFLECT')
########## MODEL ###############
content_image = model_arch(
x_new, phase, param, keep_prob, batch_size
) # put 1 for intilization and 2 for restoring saved weights in 3rd parametr
if cond == 'train':
# parameter initialzation
content_layer = 3
content_weight = tf.constant([5e-2])
style_layers = [1, 4, 6, 7]
style_weights = [2000.0, 500.0, 12.0, 1.0]
tv_weight = tf.constant([5e-2])
# CONTENT_WEIGHT = 7.5e0
# STYLE_WEIGHT = 1e2
# TV_WEIGHT = 2e2
content_original = x
# calculating_features
# fake_extract = sess.run(fake_feat[content_layer])
original_feat = model.extract_features(content_original)
original_extract = original_feat[content_layer]
fake_feat = model.extract_features(content_image)
fake_extract = fake_feat[content_layer]
style_target = [
gram_matrix(original_feat[idx]) for idx in style_layers
]
# fake_style=[fake_feat[idx] for idx in style_layers]
s_loss = style_loss(fake_feat, style_layers, style_target,
style_weights)
# con_loss==tf.py_func(content_loss,[fake_extract, original_extract],tf.float32)
con_loss = content_weight * tf.reduce_sum(
(fake_extract - original_extract)**2)
# tvariation_loss=tv_loss(content_image, tv_weight)#tf.py_func(tv_loss, [content_image, tv_weight], tf.float32)
w_variance = tf.reduce_sum(
tf.pow(
content_image[:, :, :-1, :] - content_image[:, :, 1:, :],
2))
h_variance = tf.reduce_sum(
tf.pow(
content_image[:, :-1, :, :] - content_image[:, 1:, :, :],
2))
tvariation_loss = tv_weight * (h_variance + w_variance)
total_loss = tf.add(s_loss, con_loss, name="total_loss")
final_loss = tf.add(total_loss, tvariation_loss, name="final_loss")
###trainig_initialization
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train = tf.train.AdamOptimizer(1e-2, name="adam1").minimize(
total_loss, name="training")
if cond == 'test':
return dict(x=x,
style_image=style_image,
keep_prob=keep_prob,
phase=phase,
content_image=content_image,
saver=tf.train.Saver())
else:
return dict(x=x,
style_image=style_image,
keep_prob=keep_prob,
phase=phase,
content_image=content_image,
final_loss=final_loss,
train=train,
saver=tf.train.Saver())