def find_image_mean(image_path, dataset_path = '/raid/mainak/datasets/nabirds/'):
image_paths = load_image_paths(dataset_path, path_prefix=image_path)
image_class_labels = load_image_labels(dataset_path)
# Load in the train / test split
train_images, test_images = load_train_test_split(dataset_path)
# Visualize the images and their annotations
image_identifiers = image_paths.keys()
random.shuffle(image_identifiers)
count = 0
for image_id in train_images:
image_path = image_paths[image_id]
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
im_resized = np.asarray(cv2.resize(image, IM_ORIG_SIZE), dtype = np.float64)
# x = randint(0,28)
# y = randint(0,28)
# im_cropped = im_resized[x:x+IM_SIZE[0], y:y+IM_SIZE[1]]
im_cropped = im_resized
im_cropped = im_cropped / 255
if (count == 0):
im_mean = im_cropped
else:
im_mean = (count * im_mean + im_cropped)/(count+1)
count += 1
return im_mean
def load_image_batch(im_mean, start_idx, batch_size, image_path, im_size, dataset_path = '/raid/mainak/datasets/nabirds/', is_train=1):
image_paths = load_image_paths(dataset_path, path_prefix=image_path)
image_class_labels = load_image_labels(dataset_path)
# Load in the train / test split
train_images, test_images = load_train_test_split(dataset_path)
# Visualize the images and their annotations
image_identifiers = image_paths.keys()
random.shuffle(image_identifiers)
if is_train ==1:
image_set = train_images
else:
image_set = test_images
if start_idx == 0:
random.shuffle(image_set)
batch_images = image_set[start_idx:start_idx+batch_size]
count = 0
out_images = []
image_labels = []
for image_id in batch_images:
image_path = image_paths[image_id]
# print image_path
image = cv2.imread(image_path, cv2.IMREAD_COLOR)
class_label = image_class_labels[image_id]
#if (class_label > 21):
# print "Error!: ", class_label
count += 1
processed_image = preprocessImage(image, im_mean, im_size, is_train)
out_images.append(processed_image)
image_labels.append(class_label)
#out_images_arr = np.transpose(np.asarray(out_images), (0,3,1,2)) # Make it n_images x n_ch x n_row x n_col
out_images_arr = np.asarray(out_images) # Make it n_images x n_row x n_col x n_ch
image_labels_arr = np.asarray(image_labels)
batch_x = out_images_arr
batch_y = image_labels_arr
return batch_x, batch_y
for i in range(predictions.shape[0]):
if np.argmax(predictions[i]) == np.argmax(labels[i]):
correct += 1
return 100.0 * (correct / predictions.shape[0])
if __name__ == "__main__":
args = parse_args()
datapath = Path(args.dataset_path)
samplepath = Path(args.sample_path)
summaries_dir = "summaries/"
num_patches = 10
num_epochs = 10
num_classes = 555
savepath = Path(args.save_path)
train_images, test_images = load_train_test_split()
image_paths = load_image_paths()
image_labels = load_image_labels()
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.4)
batch_size = 30
with tf.Session(config=tf.ConfigProto()) as sess:
vgg = vgg16.Vgg16(555, 80)
images = tf.placeholder("float", [batch_size, 224, 224, 3])
with tf.name_scope("content_vgg"):
vgg.build(images)
conv_1 = vgg.conv1_1
conv_6 = vgg.conv3_2
conv_9 = vgg.conv4_2
conv_12 = vgg.conv5_2
batch_cnt = tf.Variable(0, trainable=False)
starter_learning_rate = 0.0001