def conceptcombine()

in ebm_combine.py [0:0]
69 lines of code
10 McCabe index (conditional complexity)

def conceptcombine(sess, kvs, data, latents, save_exp_dir):
    X_NOISE = kvs['X_NOISE']
    LABEL_SIZE = kvs['LABEL_SIZE']
    LABEL_SHAPE = kvs['LABEL_SHAPE']
    LABEL_POS = kvs['LABEL_POS']
    LABEL_ROT = kvs['LABEL_ROT']
    model_size = kvs['model_size']
    model_shape = kvs['model_shape']
    model_pos = kvs['model_pos']
    model_rot = kvs['model_rot']
    weight_size = kvs['weight_size']
    weight_shape = kvs['weight_shape']
    weight_pos = kvs['weight_pos']
    weight_rot = kvs['weight_rot']

    x_mod = X_NOISE
    for i in range(FLAGS.num_steps):

        if FLAGS.cond_scale:
            e_noise = model_size.forward(x_mod, weight_size, label=LABEL_SIZE)
            x_grad = tf.gradients(e_noise, [x_mod])[0]
            x_mod = x_mod + tf.random_normal(tf.shape(x_mod), mean=0.0, stddev=0.005)
            x_mod = x_mod - FLAGS.step_lr * x_grad
            x_mod = tf.clip_by_value(x_mod, 0, 1)

        if FLAGS.cond_shape:
            e_noise = model_shape.forward(x_mod, weight_shape, label=LABEL_SHAPE)
            x_grad = tf.gradients(e_noise, [x_mod])[0]
            x_mod = x_mod + tf.random_normal(tf.shape(x_mod), mean=0.0, stddev=0.005)
            x_mod = x_mod - FLAGS.step_lr * x_grad
            x_mod = tf.clip_by_value(x_mod, 0, 1)

        if FLAGS.cond_pos:
            e_noise = model_pos.forward(x_mod, weight_pos, label=LABEL_POS)
            x_grad = tf.gradients(e_noise, [x_mod])[0]
            x_mod = x_mod + tf.random_normal(tf.shape(x_mod), mean=0.0, stddev=0.005)
            x_mod = x_mod - FLAGS.step_lr * x_grad
            x_mod = tf.clip_by_value(x_mod, 0, 1)

        if FLAGS.cond_rot:
            e_noise = model_rot.forward(x_mod, weight_rot, label=LABEL_ROT)
            x_grad = tf.gradients(e_noise, [x_mod])[0]
            x_mod = x_mod + tf.random_normal(tf.shape(x_mod), mean=0.0, stddev=0.005)
            x_mod = x_mod - FLAGS.step_lr * x_grad
            x_mod = tf.clip_by_value(x_mod, 0, 1)

        print("Finished constructing loop {}".format(i))

    x_final = x_mod

    data_try = data[:10]
    data_init = 0.5 + 0.5 * np.random.randn(10, 64, 64)
    label_scale = latents[:10, 2:3]
    label_shape = np.eye(3)[(latents[:10, 1]-1).astype(np.uint8)]
    label_rot = latents[:10, 3:4]
    label_rot = np.concatenate([np.cos(label_rot), np.sin(label_rot)], axis=1)
    label_pos = latents[:10, 4:]

    feed_dict = {X_NOISE: data_init, LABEL_SIZE: label_scale, LABEL_SHAPE: label_shape, LABEL_POS: label_pos,
                 LABEL_ROT: label_rot}
    x_out = sess.run([x_final], feed_dict)[0]

    im_name = "im"

    if FLAGS.cond_scale:
        im_name += "_condscale"

    if FLAGS.cond_shape:
        im_name += "_condshape"

    if FLAGS.cond_pos:
        im_name += "_condpos"

    if FLAGS.cond_rot:
        im_name += "_condrot"

    im_name += ".png"

    x_out_pad, data_try_pad = np.ones((10, 66, 66)), np.ones((10, 66, 66))
    x_out_pad[:, 1:-1, 1:-1] = x_out
    data_try_pad[:, 1:-1, 1:-1] = data_try

    im_output = np.concatenate([x_out_pad, data_try_pad], axis=2).reshape(-1, 66*2)
    impath = osp.join(save_exp_dir, im_name)
    imsave(impath, im_output)
    print("Successfully saved images at {}".format(impath))