import os import numpy as np import tensorflow as tf from module import encoder, decoder from glob import glob import runway @runway.setup(options={'styleCheckpoint': runway.file(is_directory=True)}) def setup(opts): sess = tf.Session() init_op = tf.global_variables_initializer() sess.run(init_op) with tf.name_scope('placeholder'): input_photo = tf.placeholder(dtype=tf.float32, shape=[1, None, None, 3], name='photo') input_photo_features = encoder(image=input_photo, options={'gf_dim': 32}, reuse=False) output_photo = decoder(features=input_photo_features, options={'gf_dim': 32}, reuse=False) saver = tf.train.Saver() path = opts['styleCheckpoint'] model_name = [p for p in os.listdir(path) if os.path.isdir(os.path.join(path, p))][0] checkpoint_dir = os.path.join(path, model_name, 'checkpoint_long') ckpt = tf.train.get_checkpoint_state(checkpoint_dir) ckpt_name = os.path.basename(ckpt.model_checkpoint_path) saver.restore(sess, os.path.join(checkpoint_dir, ckpt_name)) return dict(sess=sess, input_photo=input_photo, output_photo=output_photo) @runway.command('stylize', inputs={'contentImage': runway.image}, outputs={'stylizedImage': runway.image}) def stylize(model, inp): img = inp['contentImage'] img = np.array(img) img = img / 127.5 - 1. img = np.expand_dims(img, axis=0) img = model['sess'].run(model['output_photo'], feed_dict={model['input_photo']: img}) img = (img + 1.) * 127.5 img = img.astype('uint8') img = img[0] return dict(stylizedImage=img) if __name__ == '__main__': runway.run()