示例#1
0
def eval(csvspath, runs_path, model_path, dataset_path, output_path):

    with h5py.File(dataset_path, "r") as data:
        xtest = data["xtest"][:]
        ctest = data["ctest"][:]

    with tf.Session() as sess:
        fid = create_fid_func(model_path)
        eval_test(csvspath, runs_path, fid, xtest, ctest, output_path)
示例#2
0
def run(cfg, dataset_path, logs_dir, checkpoints_dir, checkpoints=True):
    with tf.Session() as sess:
        get_data = from_files.get_data_provider(dataset_path, cfg.batch_size)
        get_gen = from_files.get_genconst_provider(dataset_path, cfg.batch_size)

        if cfg.validation:
            get_valid_data = from_files.get_valid_provider(dataset_path, cfg.batch_size)
            get_valid_const = from_files.get_validconst_provider(dataset_path, cfg.batch_size)
        else:
            get_valid_data = None
            get_valid_const = None

        get_noise = cfg.noise_provider(**cfg.noise_provider_args)

        if cfg.fid_model is not None:
            fid = create_fid_func(cfg.fid_model, sess)
        else:
            fid = lambda x, y: 0

        # Building models
        Z = tf.placeholder(tf.float32, (None, cfg.zx, cfg.zx, cfg.nz,), name="Z")
        X = tf.placeholder(tf.float32, (None, cfg.npx, cfg.npx, cfg.channels,), name="X")
        C = tf.placeholder(tf.float32, (None, cfg.npx, cfg.npx, cfg.channels,), name="C")
        Cf = tf.placeholder(tf.float32, (None, cfg.npx, cfg.npx, cfg.channels,), name="Cf")

        D = cfg.discriminator(X, C, **cfg.disc_args)
        G = cfg.generator(Z, Cf, **cfg.gen_args)

        D_out = D.output
        G_out = G.output

        with tf.name_scope("DG"):
            DG = D([G_out, Cf])

        # Objectives
        with tf.name_scope("D_real_objective"):
            D_real_objective = tf.reduce_mean(-tf.log(D_out + 1e-8))

        with tf.name_scope("D_fake_objective"):
            D_fake_objective = tf.reduce_mean(-tf.log(1 - DG + 1e-8))

        with tf.name_scope("D_objective"):
            D_objective = D_real_objective + D_fake_objective

        with tf.name_scope("G_objective"):
            G_real_objective = tf.reduce_mean(-tf.log(DG + 1e-8))
            G_objective = G_real_objective

        # Optimizers
        D_optimizer = cfg.disc_optimizer(**cfg.disc_optimizer_args)
        G_optimizer = cfg.gen_optimizer(**cfg.gen_optimizer_args)

        D_cost = D_optimizer.minimize(D_objective, var_list=D.trainable_weights)
        G_cost = G_optimizer.minimize(G_objective, var_list=G.trainable_weights)

        # Logging costs
        drealcostsumm = tf.summary.scalar("D_real_cost", D_real_objective)
        dfakecostsumm = tf.summary.scalar("D_fake_cost", D_fake_objective)
        gcostsumm = tf.summary.scalar("G_cost", G_real_objective)

        # Logging images
        constimgpl = tf.placeholder(tf.float32, shape=(1, cfg.npx, cfg.npx, 3))
        consttrueimgpl = tf.placeholder(tf.float32, shape=(1, cfg.npx, cfg.npx, 3))
        imgpl = tf.placeholder(tf.float32, shape=(1, cfg.npx, cfg.npx, cfg.channels))
        trueimgpl = tf.placeholder(tf.float32, shape=(1, cfg.npx, cfg.npx, cfg.channels))

        imgsummaries = [
            tf.summary.image("Generated_const", constimgpl),
            tf.summary.image("Ground_truth_const", consttrueimgpl),
            tf.summary.image("Generated", imgpl),
            tf.summary.image("Ground_truth", trueimgpl)
        ]

        imgsumm = tf.summary.merge(imgsummaries)

        # Logging weights histograms
        weightsumms = []
        for layer in D.layers:
            i = 0
            for vect in layer.trainable_weights:
                weightsumms.append(tf.summary.histogram(
                    "D_" + layer.name + str(i), vect))
        for layer in G.layers:
            i = 0
            for vect in layer.trainable_weights:
                weightsumms.append(tf.summary.histogram(
                    "G_" + layer.name + str(i), vect))

        weightsum = tf.summary.merge(weightsumms)

        # Setting up the training
        sess.run(tf.global_variables_initializer())
        writer = tf.summary.FileWriter(
            logs_dir + os.sep + cfg.name, tf.get_default_graph())
        os.mkdir(checkpoints_dir)
        writer.flush()

        # Do the actual training
        for epoch in range(cfg.epochs):
            bar = progressbar.ProgressBar(maxvalue=cfg.epoch_iters, redirect_stdout=True)
            print("----------------------------------------------------Epoch " + str(epoch) + "----------------------------------------------------")

            for it in bar(range(int(cfg.dataset_size / cfg.batch_size))):
                # Training D
                x_real, c_real = get_data()
                c_fake = get_gen()
                noise = get_noise(cfg.batch_size)
                _, drealout, dfakeout = sess.run([D_cost, drealcostsumm, dfakecostsumm], feed_dict={X: x_real, Z: noise, C: c_real, Cf: c_fake})

                # Training G
                c_fake = get_gen()
                noise = get_noise(cfg.batch_size)
                _, gout = sess.run([G_cost, gcostsumm], feed_dict={Z: noise, Cf: c_fake})

                # Logging losses
                t = int(cfg.dataset_size / cfg.batch_size) * epoch + it
                writer.add_summary(drealout, t)
                writer.add_summary(dfakeout, t)
                writer.add_summary(gout, t)

            # Epoch ended

            # Logging metrics on validation set
            curmets = {}
            generated = []
            valid_data = []
            valid_consts = []
            bar = progressbar.ProgressBar(maxvalue=int(cfg.valid_size / cfg.batch_size), redirect_stdout=True)
            print("Generating on validation")
            for i in bar(range(int(cfg.valid_size / cfg.batch_size))):
                real_imgs = get_valid_data()
                consts = get_valid_const()
                if len(real_imgs) == cfg.batch_size:
                    noise = get_noise(cfg.batch_size)
                    generated.extend(list(sess.run(G_out, feed_dict={Z: noise, Cf: consts})))
                    valid_data.extend(list(real_imgs))
                    valid_consts.extend(list(consts))

            generated = np.asarray(generated)
            valid_data = np.asarray(valid_data)
            valid_consts = np.asarray(valid_consts)

            for m in cfg.metrics:
                if m not in curmets:
                    curmets[m] = []
                curmets[m].append(cfg.metrics[m](valid_data, generated))

            metricslist = [
                tf.Summary.Value(tag="MSE", simple_value=metrics.mse(generated, valid_consts)),
                tf.Summary.Value(tag="FID", simple_value=fid(valid_data, generated))
            ]

            for m in curmets.keys():
                metricslist.append(tf.Summary.Value(tag=m, simple_value=np.mean(curmets[m])))
            metricsout = tf.Summary(value=metricslist)

            # Logging weights histograms
            weightout = sess.run(weightsum)

            # Logging images
            print("Logging images")
            true_img = np.expand_dims(x_real[0], axis=0)
            const = np.expand_dims(c_real[0], axis=0)
            noise = get_noise(1)
            img = sess.run(G_out, feed_dict={Z: noise, Cf: const})

            imgout = sess.run(
                imgsumm, feed_dict={
                    trueimgpl: true_img,
                    imgpl: img,
                    constimgpl: log.constraints_image(img, const),
                    consttrueimgpl: log.constraints_image(true_img, const)
                })
            writer.flush()

            # Writing all logs as tensorboard
            writer.add_summary(metricsout, epoch)
            writer.add_summary(imgout, epoch)
            writer.add_summary(weightout, epoch)
            writer.flush()

            # Saving weights
            if checkpoints:
                G.save(checkpoints_dir + os.sep + "G_" + str(epoch) + ".hdf5", include_optimizer=False)
                D.save(checkpoints_dir + os.sep + "D_" + str(epoch) + ".hdf5", include_optimizer=False)

        # Run end
        writer.close()
    tf.reset_default_graph()