iGAN_predict.py

from __future__ import print_function

import theano
import theano.tensor as T
from time import time
from lib import HOGNet
from lib.rng import np_rng
from lib.theano_utils import floatX, sharedX
import numpy as np

from lib import AlexNet
import lasagne
from scipy import optimize
import argparse
from PIL import Image
from pydoc import locate
from lib import activations


def def_feature(layer='conv4', up_scale=4):
    print('COMPILING...')
    t = time()
    x = T.tensor4()
    x_t = AlexNet.transform_im(x)
    x_net = AlexNet.build_model(x_t, layer=layer, shape=(None, 3, 64, 64), up_scale=up_scale)
    AlexNet.load_model(x_net, layer=layer)
    x_f = lasagne.layers.get_output(x_net[layer], deterministic=True)
    _ftr = theano.function(inputs=[x], outputs=x_f)
    print('%.2f seconds to compile _feature function' % (time() - t))
    return _ftr


def def_bfgs(model_G, layer='conv4', npx=64, alpha=0.002):
    print('COMPILING...')
    t = time()

    x_f = T.tensor4()
    x = T.tensor4()
    z = T.matrix()
    tanh = activations.Tanh()
    gx = model_G(tanh(z))

    if layer is 'hog':
        gx_f = HOGNet.get_hog(gx, use_bin=True, BS=4)
    else:
        gx_t = AlexNet.transform_im(gx)
        gx_net = AlexNet.build_model(gx_t, layer=layer, shape=(None, 3, npx, npx))
        AlexNet.load_model(gx_net, layer=layer)
        gx_f = lasagne.layers.get_output(gx_net[layer], deterministic=True)

    f_rec = T.mean(T.sqr(x_f - gx_f), axis=(1, 2, 3)) * sharedX(alpha)
    x_rec = T.mean(T.sqr(x - gx), axis=(1, 2, 3))
    cost = T.sum(f_rec) + T.sum(x_rec)
    grad = T.grad(cost, z)
    output = [cost, grad, gx]
    _invert = theano.function(inputs=[z, x, x_f], outputs=output)

    print('%.2f seconds to compile _bfgs function' % (time() - t))
    return _invert, z


def def_predict(model_P):
    print('COMPILING...')
    t = time()
    x = T.tensor4()
    z = model_P(x)
    _predict = theano.function([x], [z])
    print('%.2f seconds to compile _predict function' % (time() - t))
    return _predict


def def_invert_models(gen_model, layer='conv4', alpha=0.002):
    bfgs_model = def_bfgs(gen_model.model_G, layer=layer, npx=gen_model.npx, alpha=alpha)
    ftr_model = def_feature(layer=layer)
    predict_model = def_predict(gen_model.model_P)
    return gen_model, bfgs_model, ftr_model, predict_model


def predict_z(gen_model, _predict, ims, batch_size=32):
    n = ims.shape[0]
    n_gen = 0
    zs = []
    n_batch = int(np.ceil(n / float(batch_size)))
    for i in range(n_batch):
        imb = gen_model.transform(ims[batch_size * i:min(n, batch_size * (i + 1)), :, :, :])
        zmb = _predict(imb)
        zs.append(zmb)
        n_gen += len(imb)
    zs = np.squeeze(np.concatenate(zs, axis=0))
    if np.ndim(zs) == 1:
        zs = zs[np.newaxis, :]

    return zs


def invert_bfgs_batch(gen_model, invert_model, ftr_model, ims, z_predict=None, npx=64):
    zs = []
    recs = []
    fs = []
    n_imgs = ims.shape[0]
    print('reconstruct %d images using bfgs' % n_imgs)

    for n in range(n_imgs):
        im_n = ims[[n], :, :, :]
        if z_predict is not None:
            z0_n = z_predict[[n], ...]
        else:
            z0_n = None
        gx, z_value, f_value = invert_bfgs(gen_model, invert_model, ftr_model, im=im_n, z_predict=z0_n, npx=npx)
        rec_im = (gx * 255).astype(np.uint8)
        fs.append(f_value[np.newaxis, ...])
        zs.append(z_value[np.newaxis, ...])
        recs.append(rec_im)
    recs = np.concatenate(recs, axis=0)
    zs = np.concatenate(zs, axis=0)
    fs = np.concatenate(fs, axis=0)
    return recs, zs, fs


def invert_bfgs(gen_model, invert_model, ftr_model, im, z_predict=None, npx=64):
    _f, z = invert_model
    nz = gen_model.nz
    if z_predict is None:
        z_predict = np_rng.uniform(-1., 1., size=(1, nz))
    else:
        z_predict = floatX(z_predict)
    z_predict = np.arctanh(z_predict)
    im_t = gen_model.transform(im)
    ftr = ftr_model(im_t)

    prob = optimize.minimize(f_bfgs, z_predict, args=(_f, im_t, ftr),
                             tol=1e-6, jac=True, method='L-BFGS-B', options={'maxiter': 200})
    print('n_iters = %3d, f = %.3f' % (prob.nit, prob.fun))
    z_opt = prob.x
    z_opt_n = floatX(z_opt[np.newaxis, :])
    [f_opt, g, gx] = _f(z_opt_n, im_t, ftr)
    gx = gen_model.inverse_transform(gx, npx=npx)
    z_opt = np.tanh(z_opt)
    return gx, z_opt, f_opt


def f_bfgs(z0, _f, x, x_f):
    z0_n = floatX(z0[np.newaxis, :])
    [f, g, gx] = _f(z0_n, x, x_f)
    f = f.astype(np.float64)
    g = g[0].astype(np.float64)
    return f, g


def invert_images_CNN_opt(invert_models, ims, solver='cnn'):
    gen_model, invert_model, ftr_model, predict_model = invert_models
    n_imgs = len(ims)
    print('process %d images' % n_imgs)
    # gen_samples(self, z0=None, n=32, batch_size=32, use_transform=True)
    if solver == 'cnn' or solver == 'cnn_opt':
        z_predict = predict_z(gen_model, predict_model, ims, batch_size=n_imgs)
    else:
        z_predict = None

    if solver == 'cnn':
        recs = gen_model.gen_samples(z0=z_predict, n=n_imgs, batch_size=n_imgs)
        zs = None

    if solver == 'cnn_opt' or solver == 'opt':
        recs, zs, loss = invert_bfgs_batch(gen_model, invert_model, ftr_model, ims, z_predict=z_predict, npx=npx)

    return recs, zs, z_predict


def parse_args():
    parser = argparse.ArgumentParser(description='iGAN: Interactive Visual Synthesis Powered by GAN')
    parser.add_argument('--model_name', dest='model_name', help='the model name', default='shoes_64', type=str)
    parser.add_argument('--model_type', dest='model_type', help='the generative models and its deep learning framework', default='dcgan_theano', type=str)
    parser.add_argument('--input_image', dest='input_image', help='input image', default='./pics/shoes_test.png', type=str)
    parser.add_argument('--output_image', dest='output_image', help='output reconstruction image', default=None, type=str)
    parser.add_argument('--model_file', dest='model_file', help='the file that stores the generative model', type=str, default=None)
    # cnn: feed-forward network; opt: optimization based; cnn_opt: hybrid of the two methods
    parser.add_argument('--solver', dest='solver', help='solver (cnn, opt, or cnn_opt)', type=str, default='cnn_opt')
    args = parser.parse_args()
    return args


if __name__ == "__main__":
    args = parse_args()
    if not args.model_file:  # if the model file is not specified
        args.model_file = './models/%s.%s' % (args.model_name, args.model_type)
    if not args.output_image:  # if the output image path is not specified
        args.output_image = args.input_image.replace('.png', '_%s.png' % args.solver)

    for arg in vars(args):
        print('[%s] =' % arg, getattr(args, arg))

    # read a single image
    im = Image.open(args.input_image)
    [h, w] = im.size
    print('read image: %s (%dx%d)' % (args.input_image, h, w))
    # define the theano models
    model_class = locate('model_def.%s' % args.model_type)
    gen_model = model_class.Model(model_name=args.model_name, model_file=args.model_file, use_predict=True)
    invert_models = def_invert_models(gen_model, layer='conv4', alpha=0.002)
    # pre-processing steps
    npx = gen_model.npx
    im = im.resize((npx, npx))
    im = np.array(im)
    im_pre = im[np.newaxis, :, :, :]
    # run the model
    rec, _, _ = invert_images_CNN_opt(invert_models, im_pre, solver=args.solver)
    rec = np.squeeze(rec)
    rec_im = Image.fromarray(rec)
    # resize the image (input aspect ratio)
    rec_im = rec_im.resize((h, w))
    print('write result to %s' % args.output_image)
    rec_im.save(args.output_image)