Ejemplo n.º 1
0
def pretrain_osm(lam_kld=0.0):
    # Initialize a source of randomness
    rng = np.random.RandomState(1234)

    # Load some data to train/validate/test with
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, zero_mean=False)
    Xtr = datasets[0][0]
    Xtr = Xtr.get_value(borrow=False)
    Xva = datasets[2][0]
    Xva = Xva.get_value(borrow=False)
    print("Xtr.shape: {0:s}, Xva.shape: {1:s}".format(str(Xtr.shape),str(Xva.shape)))

    # get and set some basic dataset information
    Xtr_mean = np.mean(Xtr, axis=0)
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 100
    batch_reps = 5

    # setup some symbolic variables and stuff
    Xd = T.matrix('Xd_base')
    Xc = T.matrix('Xc_base')
    Xm = T.matrix('Xm_base')
    data_dim = Xtr.shape[1]
    prior_sigma = 1.0

    ##########################
    # NETWORK CONFIGURATIONS #
    ##########################
    gn_params = {}
    shared_config = [PRIOR_DIM, 1000, 1000]
    top_config = [shared_config[-1], data_dim]
    gn_params['shared_config'] = shared_config
    gn_params['mu_config'] = top_config
    gn_params['sigma_config'] = top_config
    gn_params['activation'] = relu_actfun
    gn_params['init_scale'] = 1.4
    gn_params['lam_l2a'] = 0.0
    gn_params['vis_drop'] = 0.0
    gn_params['hid_drop'] = 0.0
    gn_params['bias_noise'] = 0.0
    gn_params['input_noise'] = 0.0
    # choose some parameters for the continuous inferencer
    in_params = {}
    shared_config = [data_dim, 1000, 1000]
    top_config = [shared_config[-1], PRIOR_DIM]
    in_params['shared_config'] = shared_config
    in_params['mu_config'] = top_config
    in_params['sigma_config'] = top_config
    in_params['activation'] = relu_actfun
    in_params['init_scale'] = 1.4
    in_params['lam_l2a'] = 0.0
    in_params['vis_drop'] = 0.0
    in_params['hid_drop'] = 0.0
    in_params['bias_noise'] = 0.0
    in_params['input_noise'] = 0.0
    # Initialize the base networks for this OneStageModel
    IN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=in_params, shared_param_dicts=None)
    GN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=gn_params, shared_param_dicts=None)
    # Initialize biases in IN and GN
    IN.init_biases(0.2)
    GN.init_biases(0.2)

    #########################
    # INITIALIZE THE GIPAIR #
    #########################
    osm_params = {}
    osm_params['x_type'] = 'bernoulli'
    osm_params['xt_transform'] = 'sigmoid'
    osm_params['logvar_bound'] = LOGVAR_BOUND
    OSM = OneStageModel(rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, \
            p_x_given_z=GN, q_z_given_x=IN, \
            x_dim=data_dim, z_dim=PRIOR_DIM, params=osm_params)
    OSM.set_lam_l2w(1e-5)
    safe_mean = (0.9 * Xtr_mean) + 0.05
    safe_mean_logit = np.log(safe_mean / (1.0 - safe_mean))
    OSM.set_output_bias(safe_mean_logit)
    OSM.set_input_bias(-Xtr_mean)

    ######################
    # BASIC VAE TRAINING #
    ######################
    out_file = open(RESULT_PATH+"pt_osm_results.txt", 'wb')
    # Set initial learning rate and basic SGD hyper parameters
    obs_costs = np.zeros((batch_size,))
    costs = [0. for i in range(10)]
    learn_rate = 0.0005
    for i in range(150000):
        scale = min(1.0, float(i) / 10000.0)
        if ((i > 1) and ((i % 20000) == 0)):
            learn_rate = learn_rate * 0.9
        # do a minibatch update of the model, and compute some costs
        tr_idx = npr.randint(low=0,high=tr_samples,size=(batch_size,))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xc_batch = 0.0 * Xd_batch
        Xm_batch = 0.0 * Xd_batch
        # do a minibatch update of the model, and compute some costs
        OSM.set_sgd_params(lr_1=(scale*learn_rate), mom_1=0.5, mom_2=0.98)
        OSM.set_lam_nll(1.0)
        OSM.set_lam_kld(lam_kld_1=(1.0 + (scale*(lam_kld-1.0))), lam_kld_2=0.0)
        result = OSM.train_joint(Xd_batch, Xc_batch, Xm_batch, batch_reps)
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 1000) == 0):
            # record and then reset the cost trackers
            costs = [(v / 1000.0) for v in costs]
            str_1 = "-- batch {0:d} --".format(i)
            str_2 = "    joint_cost: {0:.4f}".format(costs[0])
            str_3 = "    nll_cost  : {0:.4f}".format(costs[1])
            str_4 = "    kld_cost  : {0:.4f}".format(costs[2])
            str_5 = "    reg_cost  : {0:.4f}".format(costs[3])
            costs = [0.0 for v in costs]
            # print out some diagnostic information
            joint_str = "\n".join([str_1, str_2, str_3, str_4, str_5])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
        if ((i % 2000) == 0):
            Xva = row_shuffle(Xva)
            model_samps = OSM.sample_from_prior(500)
            file_name = RESULT_PATH+"pt_osm_samples_b{0:d}_XG.png".format(i)
            utils.visualize_samples(model_samps, file_name, num_rows=20)
            # compute information about free-energy on validation set
            file_name = RESULT_PATH+"pt_osm_free_energy_b{0:d}.png".format(i)
            fe_terms = OSM.compute_fe_terms(Xva[0:2500], 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            fe_str = "    nll_bound : {0:.4f}".format(fe_mean)
            print(fe_str)
            out_file.write(fe_str+"\n")
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
            # compute information about posterior KLds on validation set
            file_name = RESULT_PATH+"pt_osm_post_klds_b{0:d}.png".format(i)
            post_klds = OSM.compute_post_klds(Xva[0:2500])
            post_dim_klds = np.mean(post_klds, axis=0)
            utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, \
                    file_name)
        if ((i % 5000) == 0):
            IN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_b{0:d}_IN.pkl".format(i))
            GN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_b{0:d}_GN.pkl".format(i))
    IN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_IN.pkl")
    GN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_GN.pkl")
    return
Ejemplo n.º 2
0
def pretrain_osm(lam_kld=0.0):
    # Initialize a source of randomness
    rng = np.random.RandomState(1234)

    # Load some data to train/validate/test with
    data_file = 'data/tfd_data_48x48.pkl'
    dataset = load_tfd(tfd_pkl_name=data_file, which_set='unlabeled', fold='all')
    Xtr_unlabeled = dataset[0]
    dataset = load_tfd(tfd_pkl_name=data_file, which_set='train', fold='all')
    Xtr_train = dataset[0]
    Xtr = np.vstack([Xtr_unlabeled, Xtr_train])
    dataset = load_tfd(tfd_pkl_name=data_file, which_set='valid', fold='all')
    Xva = dataset[0]
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 400
    batch_reps = 6
    carry_frac = 0.25
    carry_size = int(batch_size * carry_frac)
    reset_prob = 0.04

    # setup some symbolic variables and stuff
    Xd = T.matrix('Xd_base')
    Xc = T.matrix('Xc_base')
    Xm = T.matrix('Xm_base')
    data_dim = Xtr.shape[1]
    prior_sigma = 1.0
    Xtr_mean = np.mean(Xtr, axis=0)

    ##########################
    # NETWORK CONFIGURATIONS #
    ##########################
    gn_params = {}
    shared_config = [PRIOR_DIM, 1500, 1500]
    top_config = [shared_config[-1], data_dim]
    gn_params['shared_config'] = shared_config
    gn_params['mu_config'] = top_config
    gn_params['sigma_config'] = top_config
    gn_params['activation'] = relu_actfun
    gn_params['init_scale'] = 1.4
    gn_params['lam_l2a'] = 0.0
    gn_params['vis_drop'] = 0.0
    gn_params['hid_drop'] = 0.0
    gn_params['bias_noise'] = 0.0
    gn_params['input_noise'] = 0.0
    # choose some parameters for the continuous inferencer
    in_params = {}
    shared_config = [data_dim, 1500, 1500]
    top_config = [shared_config[-1], PRIOR_DIM]
    in_params['shared_config'] = shared_config
    in_params['mu_config'] = top_config
    in_params['sigma_config'] = top_config
    in_params['activation'] = relu_actfun
    in_params['init_scale'] = 1.4
    in_params['lam_l2a'] = 0.0
    in_params['vis_drop'] = 0.0
    in_params['hid_drop'] = 0.0
    in_params['bias_noise'] = 0.0
    in_params['input_noise'] = 0.0
    # Initialize the base networks for this OneStageModel
    IN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=in_params, shared_param_dicts=None)
    GN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=gn_params, shared_param_dicts=None)
    # Initialize biases in IN and GN
    IN.init_biases(0.2)
    GN.init_biases(0.2)

    ######################################
    # LOAD AND RESTART FROM SAVED PARAMS #
    ######################################
    # gn_fname = RESULT_PATH+"pt_osm_params_b110000_GN.pkl"
    # in_fname = RESULT_PATH+"pt_osm_params_b110000_IN.pkl"
    # IN = load_infnet_from_file(f_name=in_fname, rng=rng, Xd=Xd, \
    #         new_params=None)
    # GN = load_infnet_from_file(f_name=gn_fname, rng=rng, Xd=Xd, \
    #         new_params=None)
    # in_params = IN.params
    # gn_params = GN.params

    #########################
    # INITIALIZE THE GIPAIR #
    #########################
    osm_params = {}
    osm_params['x_type'] = 'bernoulli'
    osm_params['xt_transform'] = 'sigmoid'
    osm_params['logvar_bound'] = LOGVAR_BOUND
    OSM = OneStageModel(rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, \
            p_x_given_z=GN, q_z_given_x=IN, \
            x_dim=data_dim, z_dim=PRIOR_DIM, params=osm_params)
    OSM.set_lam_l2w(1e-5)
    safe_mean = (0.9 * Xtr_mean) + 0.05
    safe_mean_logit = np.log(safe_mean / (1.0 - safe_mean))
    OSM.set_output_bias(safe_mean_logit)
    OSM.set_input_bias(-Xtr_mean)

    ######################
    # BASIC VAE TRAINING #
    ######################
    out_file = open(RESULT_PATH+"pt_osm_results.txt", 'wb')
    # Set initial learning rate and basic SGD hyper parameters
    obs_costs = np.zeros((batch_size,))
    costs = [0. for i in range(10)]
    learn_rate = 0.002
    for i in range(200000):
        scale = min(1.0, float(i) / 5000.0)
        if ((i > 1) and ((i % 20000) == 0)):
            learn_rate = learn_rate * 0.8
        if (i < 50000):
            momentum = 0.5
        elif (i < 10000):
            momentum = 0.7
        else:
            momentum = 0.9
        if ((i == 0) or (npr.rand() < reset_prob)):
            # sample a fully random batch
            batch_idx = npr.randint(low=0,high=tr_samples,size=(batch_size,))
        else:
            # sample a partially random batch, which retains some portion of
            # the worst scoring examples from the previous batch
            fresh_idx = npr.randint(low=0,high=tr_samples,size=(batch_size-carry_size,))
            batch_idx = np.concatenate((fresh_idx.ravel(), carry_idx.ravel()))
        # do a minibatch update of the model, and compute some costs
        tr_idx = npr.randint(low=0,high=tr_samples,size=(batch_size,))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xc_batch = 0.0 * Xd_batch
        Xm_batch = 0.0 * Xd_batch
        # do a minibatch update of the model, and compute some costs
        OSM.set_sgd_params(lr_1=(scale*learn_rate), \
                mom_1=(scale*momentum), mom_2=0.98)
        OSM.set_lam_nll(1.0)
        OSM.set_lam_kld(lam_kld_1=scale*lam_kld, lam_kld_2=0.0, lam_kld_c=50.0)
        result = OSM.train_joint(Xd_batch, Xc_batch, Xm_batch, batch_reps)
        batch_costs = result[4] + result[5]
        obs_costs = collect_obs_costs(batch_costs, batch_reps)
        carry_idx = batch_idx[np.argsort(-obs_costs)[0:carry_size]]
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 1000) == 0):
            # record and then reset the cost trackers
            costs = [(v / 1000.0) for v in costs]
            str_1 = "-- batch {0:d} --".format(i)
            str_2 = "    joint_cost: {0:.4f}".format(costs[0])
            str_3 = "    nll_cost  : {0:.4f}".format(costs[1])
            str_4 = "    kld_cost  : {0:.4f}".format(costs[2])
            str_5 = "    reg_cost  : {0:.4f}".format(costs[3])
            costs = [0.0 for v in costs]
            # print out some diagnostic information
            joint_str = "\n".join([str_1, str_2, str_3, str_4, str_5])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
        if ((i % 2000) == 0):
            Xva = row_shuffle(Xva)
            model_samps = OSM.sample_from_prior(500)
            file_name = RESULT_PATH+"pt_osm_samples_b{0:d}_XG.png".format(i)
            utils.visualize_samples(model_samps, file_name, num_rows=20)
            file_name = RESULT_PATH+"pt_osm_inf_weights_b{0:d}.png".format(i)
            utils.visualize_samples(OSM.inf_weights.get_value(borrow=False).T, \
                    file_name, num_rows=30)
            file_name = RESULT_PATH+"pt_osm_gen_weights_b{0:d}.png".format(i)
            utils.visualize_samples(OSM.gen_weights.get_value(borrow=False), \
                    file_name, num_rows=30)
            # compute information about free-energy on validation set
            file_name = RESULT_PATH+"pt_osm_free_energy_b{0:d}.png".format(i)
            fe_terms = OSM.compute_fe_terms(Xva[0:2500], 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            fe_str = "    nll_bound : {0:.4f}".format(fe_mean)
            print(fe_str)
            out_file.write(fe_str+"\n")
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
            # compute information about posterior KLds on validation set
            file_name = RESULT_PATH+"pt_osm_post_klds_b{0:d}.png".format(i)
            post_klds = OSM.compute_post_klds(Xva[0:2500])
            post_dim_klds = np.mean(post_klds, axis=0)
            utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, \
                    file_name)
        if ((i % 5000) == 0):
            IN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_b{0:d}_IN.pkl".format(i))
            GN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_b{0:d}_GN.pkl".format(i))
    IN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_IN.pkl")
    GN.save_to_file(f_name=RESULT_PATH+"pt_osm_params_GN.pkl")
    return
Ejemplo n.º 3
0
def pretrain_osm(lam_kld=0.0):
    # Initialize a source of randomness
    rng = np.random.RandomState(1234)

    # Load some data to train/validate/test with
    data_file = 'data/tfd_data_48x48.pkl'
    dataset = load_tfd(tfd_pkl_name=data_file,
                       which_set='unlabeled',
                       fold='all')
    Xtr_unlabeled = dataset[0]
    dataset = load_tfd(tfd_pkl_name=data_file, which_set='train', fold='all')
    Xtr_train = dataset[0]
    Xtr = np.vstack([Xtr_unlabeled, Xtr_train])
    dataset = load_tfd(tfd_pkl_name=data_file, which_set='valid', fold='all')
    Xva = dataset[0]
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 400
    batch_reps = 6
    carry_frac = 0.25
    carry_size = int(batch_size * carry_frac)
    reset_prob = 0.04

    # setup some symbolic variables and stuff
    Xd = T.matrix('Xd_base')
    Xc = T.matrix('Xc_base')
    Xm = T.matrix('Xm_base')
    data_dim = Xtr.shape[1]
    prior_sigma = 1.0
    Xtr_mean = np.mean(Xtr, axis=0)

    ##########################
    # NETWORK CONFIGURATIONS #
    ##########################
    gn_params = {}
    shared_config = [PRIOR_DIM, 1500, 1500]
    top_config = [shared_config[-1], data_dim]
    gn_params['shared_config'] = shared_config
    gn_params['mu_config'] = top_config
    gn_params['sigma_config'] = top_config
    gn_params['activation'] = relu_actfun
    gn_params['init_scale'] = 1.4
    gn_params['lam_l2a'] = 0.0
    gn_params['vis_drop'] = 0.0
    gn_params['hid_drop'] = 0.0
    gn_params['bias_noise'] = 0.0
    gn_params['input_noise'] = 0.0
    # choose some parameters for the continuous inferencer
    in_params = {}
    shared_config = [data_dim, 1500, 1500]
    top_config = [shared_config[-1], PRIOR_DIM]
    in_params['shared_config'] = shared_config
    in_params['mu_config'] = top_config
    in_params['sigma_config'] = top_config
    in_params['activation'] = relu_actfun
    in_params['init_scale'] = 1.4
    in_params['lam_l2a'] = 0.0
    in_params['vis_drop'] = 0.0
    in_params['hid_drop'] = 0.0
    in_params['bias_noise'] = 0.0
    in_params['input_noise'] = 0.0
    # Initialize the base networks for this OneStageModel
    IN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=in_params, shared_param_dicts=None)
    GN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=gn_params, shared_param_dicts=None)
    # Initialize biases in IN and GN
    IN.init_biases(0.2)
    GN.init_biases(0.2)

    ######################################
    # LOAD AND RESTART FROM SAVED PARAMS #
    ######################################
    # gn_fname = RESULT_PATH+"pt_osm_params_b110000_GN.pkl"
    # in_fname = RESULT_PATH+"pt_osm_params_b110000_IN.pkl"
    # IN = load_infnet_from_file(f_name=in_fname, rng=rng, Xd=Xd, \
    #         new_params=None)
    # GN = load_infnet_from_file(f_name=gn_fname, rng=rng, Xd=Xd, \
    #         new_params=None)
    # in_params = IN.params
    # gn_params = GN.params

    #########################
    # INITIALIZE THE GIPAIR #
    #########################
    osm_params = {}
    osm_params['x_type'] = 'bernoulli'
    osm_params['xt_transform'] = 'sigmoid'
    osm_params['logvar_bound'] = LOGVAR_BOUND
    OSM = OneStageModel(rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, \
            p_x_given_z=GN, q_z_given_x=IN, \
            x_dim=data_dim, z_dim=PRIOR_DIM, params=osm_params)
    OSM.set_lam_l2w(1e-5)
    safe_mean = (0.9 * Xtr_mean) + 0.05
    safe_mean_logit = np.log(safe_mean / (1.0 - safe_mean))
    OSM.set_output_bias(safe_mean_logit)
    OSM.set_input_bias(-Xtr_mean)

    ######################
    # BASIC VAE TRAINING #
    ######################
    out_file = open(RESULT_PATH + "pt_osm_results.txt", 'wb')
    # Set initial learning rate and basic SGD hyper parameters
    obs_costs = np.zeros((batch_size, ))
    costs = [0. for i in range(10)]
    learn_rate = 0.002
    for i in range(200000):
        scale = min(1.0, float(i) / 5000.0)
        if ((i > 1) and ((i % 20000) == 0)):
            learn_rate = learn_rate * 0.8
        if (i < 50000):
            momentum = 0.5
        elif (i < 10000):
            momentum = 0.7
        else:
            momentum = 0.9
        if ((i == 0) or (npr.rand() < reset_prob)):
            # sample a fully random batch
            batch_idx = npr.randint(low=0,
                                    high=tr_samples,
                                    size=(batch_size, ))
        else:
            # sample a partially random batch, which retains some portion of
            # the worst scoring examples from the previous batch
            fresh_idx = npr.randint(low=0,
                                    high=tr_samples,
                                    size=(batch_size - carry_size, ))
            batch_idx = np.concatenate((fresh_idx.ravel(), carry_idx.ravel()))
        # do a minibatch update of the model, and compute some costs
        tr_idx = npr.randint(low=0, high=tr_samples, size=(batch_size, ))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xc_batch = 0.0 * Xd_batch
        Xm_batch = 0.0 * Xd_batch
        # do a minibatch update of the model, and compute some costs
        OSM.set_sgd_params(lr_1=(scale*learn_rate), \
                mom_1=(scale*momentum), mom_2=0.98)
        OSM.set_lam_nll(1.0)
        OSM.set_lam_kld(lam_kld_1=scale * lam_kld,
                        lam_kld_2=0.0,
                        lam_kld_c=50.0)
        result = OSM.train_joint(Xd_batch, Xc_batch, Xm_batch, batch_reps)
        batch_costs = result[4] + result[5]
        obs_costs = collect_obs_costs(batch_costs, batch_reps)
        carry_idx = batch_idx[np.argsort(-obs_costs)[0:carry_size]]
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 1000) == 0):
            # record and then reset the cost trackers
            costs = [(v / 1000.0) for v in costs]
            str_1 = "-- batch {0:d} --".format(i)
            str_2 = "    joint_cost: {0:.4f}".format(costs[0])
            str_3 = "    nll_cost  : {0:.4f}".format(costs[1])
            str_4 = "    kld_cost  : {0:.4f}".format(costs[2])
            str_5 = "    reg_cost  : {0:.4f}".format(costs[3])
            costs = [0.0 for v in costs]
            # print out some diagnostic information
            joint_str = "\n".join([str_1, str_2, str_3, str_4, str_5])
            print(joint_str)
            out_file.write(joint_str + "\n")
            out_file.flush()
        if ((i % 2000) == 0):
            Xva = row_shuffle(Xva)
            model_samps = OSM.sample_from_prior(500)
            file_name = RESULT_PATH + "pt_osm_samples_b{0:d}_XG.png".format(i)
            utils.visualize_samples(model_samps, file_name, num_rows=20)
            file_name = RESULT_PATH + "pt_osm_inf_weights_b{0:d}.png".format(i)
            utils.visualize_samples(OSM.inf_weights.get_value(borrow=False).T, \
                    file_name, num_rows=30)
            file_name = RESULT_PATH + "pt_osm_gen_weights_b{0:d}.png".format(i)
            utils.visualize_samples(OSM.gen_weights.get_value(borrow=False), \
                    file_name, num_rows=30)
            # compute information about free-energy on validation set
            file_name = RESULT_PATH + "pt_osm_free_energy_b{0:d}.png".format(i)
            fe_terms = OSM.compute_fe_terms(Xva[0:2500], 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            fe_str = "    nll_bound : {0:.4f}".format(fe_mean)
            print(fe_str)
            out_file.write(fe_str + "\n")
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
            # compute information about posterior KLds on validation set
            file_name = RESULT_PATH + "pt_osm_post_klds_b{0:d}.png".format(i)
            post_klds = OSM.compute_post_klds(Xva[0:2500])
            post_dim_klds = np.mean(post_klds, axis=0)
            utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, \
                    file_name)
        if ((i % 5000) == 0):
            IN.save_to_file(f_name=RESULT_PATH +
                            "pt_osm_params_b{0:d}_IN.pkl".format(i))
            GN.save_to_file(f_name=RESULT_PATH +
                            "pt_osm_params_b{0:d}_GN.pkl".format(i))
    IN.save_to_file(f_name=RESULT_PATH + "pt_osm_params_IN.pkl")
    GN.save_to_file(f_name=RESULT_PATH + "pt_osm_params_GN.pkl")
    return
Ejemplo n.º 4
0
def test_with_model_init():
    ##########################
    # Get some training data #
    ##########################
    rng = np.random.RandomState(1234)
    Xtr, Xva, Xte = load_binarized_mnist(data_path='./data/')
    del Xte
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 200
    batch_reps = 1

    ############################################################
    # Setup some parameters for the Iterative Refinement Model #
    ############################################################
    obs_dim = Xtr.shape[1]
    z_dim = 20
    h_dim = 200
    ir_steps = 6
    init_scale = 1.0
    
    x_type = 'bernoulli'

    # some InfNet instances to build the TwoStageModel from
    x_in_sym = T.matrix('x_in_sym')
    x_out_sym = T.matrix('x_out_sym')

    #################
    # p_hi_given_si #
    #################
    params = {}
    shared_config = [obs_dim, 300, 300]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_hi_given_si = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_hi_given_si.init_biases(0.2)
    ######################
    # p_sip1_given_si_hi #
    ######################
    params = {}
    shared_config = [h_dim, 300, 300]
    output_config = [obs_dim, obs_dim, obs_dim]
    params['shared_config'] = shared_config
    params['output_config'] = output_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_sip1_given_si_hi = HydraNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_sip1_given_si_hi.init_biases(0.2)
    ################
    # p_s0_given_z #
    ################
    params = {}
    shared_config = [z_dim, 250, 250]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_s0_given_z = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_s0_given_z.init_biases(0.2)
    ###############
    # q_z_given_x #
    ###############
    params = {}
    shared_config = [obs_dim, 250, 250]
    top_config = [shared_config[-1], z_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_z_given_x = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    q_z_given_x.init_biases(0.2)
    ###################
    # q_hi_given_x_si #
    ###################
    params = {}
    shared_config = [(obs_dim + obs_dim), 500, 500]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_hi_given_x_si = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    q_hi_given_x_si.init_biases(0.2)


    ################################################################
    # Define parameters for the MultiStageModel, and initialize it #
    ################################################################
    print("Building the MultiStageModel...")
    msm_params = {}
    msm_params['x_type'] = x_type
    msm_params['obs_transform'] = 'sigmoid'
    MSM = MultiStageModel(rng=rng, x_in=x_in_sym, x_out=x_out_sym, \
            p_s0_given_z=p_s0_given_z, \
            p_hi_given_si=p_hi_given_si, \
            p_sip1_given_si_hi=p_sip1_given_si_hi, \
            q_z_given_x=q_z_given_x, \
            q_hi_given_x_si=q_hi_given_x_si, \
            obs_dim=obs_dim, z_dim=z_dim, h_dim=h_dim, \
            ir_steps=ir_steps, params=msm_params)

    ################################################################
    # Apply some updates, to check that they aren't totally broken #
    ################################################################
    out_file = open("MSM_A_RESULTS.txt", 'wb')
    costs = [0. for i in range(10)]
    learn_rate = 0.0003
    momentum = 0.9
    batch_idx = np.arange(batch_size) + tr_samples
    for i in range(250000):
        scale = min(1.0, ((i+1) / 3000.0))
        if (((i + 1) % 10000) == 0):
            learn_rate = learn_rate * 0.95
        # get the indices of training samples for this batch update
        batch_idx += batch_size
        if (np.max(batch_idx) >= tr_samples):
            # we finished an "epoch", so we rejumble the training set
            Xtr = row_shuffle(Xtr)
            batch_idx = np.arange(batch_size)
        # set sgd and objective function hyperparams for this update
        MSM.set_sgd_params(lr_1=scale*learn_rate, lr_2=scale*learn_rate, \
                mom_1=scale*momentum, mom_2=0.99)
        MSM.set_train_switch(1.0)
        MSM.set_lam_nll(lam_nll=1.0)
        MSM.set_lam_kld(lam_kld_z=1.0, lam_kld_q2p=0.8, lam_kld_p2q=0.2)
        MSM.set_lam_kld_l1l2(lam_kld_l1l2=1.0)
        MSM.set_lam_l2w(1e-4)
        MSM.set_drop_rate(0.0)
        MSM.q_hi_given_x_si.set_bias_noise(0.0)
        MSM.p_hi_given_si.set_bias_noise(0.0)
        MSM.p_sip1_given_si_hi.set_bias_noise(0.0)
        # perform a minibatch update and record the cost for this batch
        Xb_tr = to_fX( Xtr.take(batch_idx, axis=0) )
        result = MSM.train_joint(Xb_tr, Xb_tr, batch_reps)
        costs = [(costs[j] + result[j]) for j in range(len(result)-1)]
        if ((i % 500) == 0):
            costs = [(v / 500.0) for v in costs]
            str1 = "-- batch {0:d} --".format(i)
            str2 = "    joint_cost: {0:.4f}".format(costs[0])
            str3 = "    nll_cost  : {0:.4f}".format(costs[1])
            str4 = "    kld_cost  : {0:.4f}".format(costs[2])
            str5 = "    reg_cost  : {0:.4f}".format(costs[3])
            joint_str = "\n".join([str1, str2, str3, str4, str5])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            costs = [0.0 for v in costs]
        if (((i % 2000) == 0) or ((i < 10000) and ((i % 1000) == 0))):
            MSM.set_drop_rate(0.0)
            MSM.q_hi_given_x_si.set_bias_noise(0.0)
            MSM.p_hi_given_si.set_bias_noise(0.0)
            MSM.p_sip1_given_si_hi.set_bias_noise(0.0)
            # Get some validation samples for computing diagnostics
            Xva = row_shuffle(Xva)
            Xb_va = to_fX( Xva[0:2000] )
            # draw some independent random samples from the model
            samp_count = 200
            model_samps = MSM.sample_from_prior(samp_count)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count):
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MSM_A_SAMPLES_IND_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # draw some conditional random samples from the model
            samp_count = 200
            Xs = np.vstack((Xb_tr[0:(samp_count/4)], Xb_va[0:(samp_count/4)]))
            Xs = np.repeat(Xs, 2, axis=0)
            # draw some conditional random samples from the model
            model_samps = MSM.sample_from_input(Xs, guided_decoding=False)
            model_samps.append(Xs)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count): 
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MSM_A_SAMPLES_CND_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # compute information about posterior KLds on validation set
            raw_klds = MSM.compute_raw_klds(Xb_va, Xb_va)
            init_kld, q2p_kld, p2q_kld = raw_klds
            file_name = "MSM_A_H0_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(init_kld.shape[1]), \
                    np.mean(init_kld, axis=0), file_name)
            file_name = "MSM_A_HI_Q2P_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(q2p_kld.shape[1]), \
                    np.mean(q2p_kld, axis=0), file_name)
            file_name = "MSM_A_HI_P2Q_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(p2q_kld.shape[1]), \
                    np.mean(p2q_kld, axis=0), file_name)
            Xb_tr = to_fX( Xtr[0:2000] )
            fe_terms = MSM.compute_fe_terms(Xb_tr, Xb_tr, 30)
            fe_nll = np.mean(fe_terms[0])
            fe_kld = np.mean(fe_terms[1])
            fe_joint = fe_nll + fe_kld
            joint_str = "    vfe-tr: {0:.4f}, nll: ({1:.4f}, {2:.4f}, {3:.4f}), kld: ({4:.4f}, {5:.4f}, {6:.4f})".format( \
                    fe_joint, fe_nll, np.min(fe_terms[0]), np.max(fe_terms[0]), fe_kld, np.min(fe_terms[1]), np.max(fe_terms[1]))
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            fe_terms = MSM.compute_fe_terms(Xb_va, Xb_va, 30)
            fe_nll = np.mean(fe_terms[0])
            fe_kld = np.mean(fe_terms[1])
            fe_joint = fe_nll + fe_kld
            joint_str = "    vfe-va: {0:.4f}, nll: ({1:.4f}, {2:.4f}, {3:.4f}), kld: ({4:.4f}, {5:.4f}, {6:.4f})".format( \
                    fe_joint, fe_nll, np.min(fe_terms[0]), np.max(fe_terms[0]), fe_kld, np.min(fe_terms[1]), np.max(fe_terms[1]))
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
Ejemplo n.º 5
0
def test_with_model_init():
    ##########################
    # Get some training data #
    ##########################
    rng = np.random.RandomState(1234)
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, zero_mean=False)
    Xtr_shared = datasets[0][0]
    Xva_shared = datasets[1][0]
    Xtr = Xtr_shared.get_value(borrow=False).astype(theano.config.floatX)
    Xva = Xva_shared.get_value(borrow=False).astype(theano.config.floatX)
    tr_samples = Xtr.shape[0]
    batch_size = 200
    batch_reps = 1

    ############################################################
    # Setup some parameters for the Iterative Refinement Model #
    ############################################################
    obs_dim = Xtr.shape[1]
    z_dim = 20
    h_dim = 100
    x_type = 'bernoulli'

    # some InfNet instances to build the TwoStageModel from
    X_sym = T.matrix('X_sym')

    ########################
    # p_s0_obs_given_z_obs #
    ########################
    params = {}
    shared_config = [z_dim, 250, 250]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 1e-3
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_s0_obs_given_z_obs = InfNet(rng=rng, Xd=X_sym, \
            params=params, shared_param_dicts=None)
    p_s0_obs_given_z_obs.init_biases(0.2)
    #################
    # p_hi_given_si #
    #################
    params = {}
    shared_config = [obs_dim, 250, 250]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_hi_given_si = InfNet(rng=rng, Xd=X_sym, \
            params=params, shared_param_dicts=None)
    p_hi_given_si.init_biases(0.2)
    ######################
    # p_sip1_given_si_hi #
    ######################
    params = {}
    shared_config = [h_dim, 250, 250]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_sip1_given_si_hi = InfNet(rng=rng, Xd=X_sym, \
            params=params, shared_param_dicts=None)
    p_sip1_given_si_hi.init_biases(0.2)
    ###############
    # q_z_given_x #
    ###############
    params = {}
    shared_config = [obs_dim, 250, 250]
    top_config = [shared_config[-1], z_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_z_given_x = InfNet(rng=rng, Xd=X_sym, \
            params=params, shared_param_dicts=None)
    q_z_given_x.init_biases(0.2)
    ###################
    # q_hi_given_x_si #
    ###################
    params = {}
    shared_config = [(obs_dim + obs_dim), 500, 500]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_hi_given_x_si = InfNet(rng=rng, Xd=X_sym, \
            params=params, shared_param_dicts=None)
    q_hi_given_x_si.init_biases(0.2)


    ################################################################
    # Define parameters for the MultiStageModel, and initialize it #
    ################################################################
    print("Building the MultiStageModel...")
    msm_params = {}
    msm_params['x_type'] = x_type
    msm_params['obs_transform'] = 'sigmoid'
    MSM = MultiStageModel(rng=rng, x_in=X_sym, \
            p_s0_obs_given_z_obs=p_s0_obs_given_z_obs, \
            p_hi_given_si=p_hi_given_si, \
            p_sip1_given_si_hi=p_sip1_given_si_hi, \
            q_z_given_x=q_z_given_x, \
            q_hi_given_x_si=q_hi_given_x_si, \
            obs_dim=obs_dim, z_dim=z_dim, h_dim=h_dim, \
            model_init_obs=True, ir_steps=2, \
            params=msm_params)
    obs_mean = (0.9 * np.mean(Xtr, axis=0)) + 0.05
    obs_mean_logit = np.log(obs_mean / (1.0 - obs_mean))
    MSM.set_input_bias(-obs_mean)
    MSM.set_obs_bias(0.1*obs_mean_logit)

    ################################################################
    # Apply some updates, to check that they aren't totally broken #
    ################################################################
    costs = [0. for i in range(10)]
    learn_rate = 0.0003
    momentum = 0.8
    for i in range(300000):
        scale = min(1.0, ((i+1) / 10000.0))
        extra_kl = max(0.0, ((50000.0 - i) / 50000.0))
        if (((i + 1) % 10000) == 0):
            learn_rate = learn_rate * 0.95
        # randomly sample a minibatch
        tr_idx = npr.randint(low=0,high=tr_samples,size=(batch_size,))
        Xb = binarize_data(Xtr.take(tr_idx, axis=0))
        Xb = Xb.astype(theano.config.floatX)
        # set sgd and objective function hyperparams for this update
        MSM.set_sgd_params(lr_1=scale*learn_rate, lr_2=scale*learn_rate, \
                           mom_1=(scale*momentum), mom_2=0.98)
        MSM.set_train_switch(1.0)
        MSM.set_l1l2_weight(1.0)
        MSM.set_lam_nll(lam_nll=1.0)
        MSM.set_lam_kld(lam_kld_1=(1.0+extra_kl), lam_kld_2=(1.0+extra_kl))
        MSM.set_lam_l2w(1e-6)
        MSM.set_kzg_weight(0.01)
        # perform a minibatch update and record the cost for this batch
        result = MSM.train_joint(Xb, batch_reps)
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 500) == 0):
            costs = [(v / 500.0) for v in costs]
            print("-- batch {0:d} --".format(i))
            print("    joint_cost: {0:.4f}".format(costs[0]))
            print("    nll_cost  : {0:.4f}".format(costs[1]))
            print("    kld_cost  : {0:.4f}".format(costs[2]))
            print("    reg_cost  : {0:.4f}".format(costs[3]))
            costs = [0.0 for v in costs]
        if (((i % 2000) == 0) or ((i < 10000) and ((i % 1000) == 0))):
            Xva = row_shuffle(Xva)
            # draw some independent random samples from the model
            samp_count = 200
            model_samps = MSM.sample_from_prior(samp_count)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count): 
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MX_SAMPLES_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # visualize some important weights in the model
            file_name = "MX_INF_1_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.inf_1_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MX_INF_2_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.inf_2_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MX_GEN_1_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_1_weights.get_value(borrow=False)
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MX_GEN_2_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_2_weights.get_value(borrow=False)
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MX_GEN_INF_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_inf_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            # compute information about posterior KLds on validation set
            post_klds = MSM.compute_post_klds(Xva[0:5000])
            file_name = "MX_H0_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[0].shape[1]), \
                    np.mean(post_klds[0], axis=0), file_name)
            file_name = "MX_HI_COND_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[1].shape[1]), \
                    np.mean(post_klds[1], axis=0), file_name)
            file_name = "MX_HI_GLOB_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[2].shape[1]), \
                    np.mean(post_klds[2], axis=0), file_name)
            # compute information about free-energy on validation set
            file_name = "MX_FREE_ENERGY_b{0:d}.png".format(i)
            fe_terms = MSM.compute_fe_terms(binarize_data(Xva[0:5000]), 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            print("    nll_bound : {0:.4f}".format(fe_mean))
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
    return
Ejemplo n.º 6
0
def pretrain_osm(lam_kld=0.0):
    # Initialize a source of randomness
    rng = np.random.RandomState(1234)

    # Load some data to train/validate/test with
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, zero_mean=False)
    Xtr = datasets[0][0]
    Xtr = Xtr.get_value(borrow=False)
    Xva = datasets[2][0]
    Xva = Xva.get_value(borrow=False)
    print("Xtr.shape: {0:s}, Xva.shape: {1:s}".format(str(Xtr.shape),
                                                      str(Xva.shape)))

    # get and set some basic dataset information
    Xtr_mean = np.mean(Xtr, axis=0)
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 100
    batch_reps = 5

    # setup some symbolic variables and stuff
    Xd = T.matrix('Xd_base')
    Xc = T.matrix('Xc_base')
    Xm = T.matrix('Xm_base')
    data_dim = Xtr.shape[1]
    prior_sigma = 1.0

    ##########################
    # NETWORK CONFIGURATIONS #
    ##########################
    gn_params = {}
    shared_config = [PRIOR_DIM, 1000, 1000]
    top_config = [shared_config[-1], data_dim]
    gn_params['shared_config'] = shared_config
    gn_params['mu_config'] = top_config
    gn_params['sigma_config'] = top_config
    gn_params['activation'] = relu_actfun
    gn_params['init_scale'] = 1.4
    gn_params['lam_l2a'] = 0.0
    gn_params['vis_drop'] = 0.0
    gn_params['hid_drop'] = 0.0
    gn_params['bias_noise'] = 0.0
    gn_params['input_noise'] = 0.0
    # choose some parameters for the continuous inferencer
    in_params = {}
    shared_config = [data_dim, 1000, 1000]
    top_config = [shared_config[-1], PRIOR_DIM]
    in_params['shared_config'] = shared_config
    in_params['mu_config'] = top_config
    in_params['sigma_config'] = top_config
    in_params['activation'] = relu_actfun
    in_params['init_scale'] = 1.4
    in_params['lam_l2a'] = 0.0
    in_params['vis_drop'] = 0.0
    in_params['hid_drop'] = 0.0
    in_params['bias_noise'] = 0.0
    in_params['input_noise'] = 0.0
    # Initialize the base networks for this OneStageModel
    IN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=in_params, shared_param_dicts=None)
    GN = InfNet(rng=rng, Xd=Xd, prior_sigma=prior_sigma, \
            params=gn_params, shared_param_dicts=None)
    # Initialize biases in IN and GN
    IN.init_biases(0.2)
    GN.init_biases(0.2)

    #########################
    # INITIALIZE THE GIPAIR #
    #########################
    osm_params = {}
    osm_params['x_type'] = 'bernoulli'
    osm_params['xt_transform'] = 'sigmoid'
    osm_params['logvar_bound'] = LOGVAR_BOUND
    OSM = OneStageModel(rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, \
            p_x_given_z=GN, q_z_given_x=IN, \
            x_dim=data_dim, z_dim=PRIOR_DIM, params=osm_params)
    OSM.set_lam_l2w(1e-5)
    safe_mean = (0.9 * Xtr_mean) + 0.05
    safe_mean_logit = np.log(safe_mean / (1.0 - safe_mean))
    OSM.set_output_bias(safe_mean_logit)
    OSM.set_input_bias(-Xtr_mean)

    ######################
    # BASIC VAE TRAINING #
    ######################
    out_file = open(RESULT_PATH + "pt_osm_results.txt", 'wb')
    # Set initial learning rate and basic SGD hyper parameters
    obs_costs = np.zeros((batch_size, ))
    costs = [0. for i in range(10)]
    learn_rate = 0.0005
    for i in range(150000):
        scale = min(1.0, float(i) / 10000.0)
        if ((i > 1) and ((i % 20000) == 0)):
            learn_rate = learn_rate * 0.9
        # do a minibatch update of the model, and compute some costs
        tr_idx = npr.randint(low=0, high=tr_samples, size=(batch_size, ))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xc_batch = 0.0 * Xd_batch
        Xm_batch = 0.0 * Xd_batch
        # do a minibatch update of the model, and compute some costs
        OSM.set_sgd_params(lr_1=(scale * learn_rate), mom_1=0.5, mom_2=0.98)
        OSM.set_lam_nll(1.0)
        OSM.set_lam_kld(lam_kld_1=(1.0 + (scale * (lam_kld - 1.0))),
                        lam_kld_2=0.0)
        result = OSM.train_joint(Xd_batch, Xc_batch, Xm_batch, batch_reps)
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 1000) == 0):
            # record and then reset the cost trackers
            costs = [(v / 1000.0) for v in costs]
            str_1 = "-- batch {0:d} --".format(i)
            str_2 = "    joint_cost: {0:.4f}".format(costs[0])
            str_3 = "    nll_cost  : {0:.4f}".format(costs[1])
            str_4 = "    kld_cost  : {0:.4f}".format(costs[2])
            str_5 = "    reg_cost  : {0:.4f}".format(costs[3])
            costs = [0.0 for v in costs]
            # print out some diagnostic information
            joint_str = "\n".join([str_1, str_2, str_3, str_4, str_5])
            print(joint_str)
            out_file.write(joint_str + "\n")
            out_file.flush()
        if ((i % 2000) == 0):
            Xva = row_shuffle(Xva)
            model_samps = OSM.sample_from_prior(500)
            file_name = RESULT_PATH + "pt_osm_samples_b{0:d}_XG.png".format(i)
            utils.visualize_samples(model_samps, file_name, num_rows=20)
            # compute information about free-energy on validation set
            file_name = RESULT_PATH + "pt_osm_free_energy_b{0:d}.png".format(i)
            fe_terms = OSM.compute_fe_terms(Xva[0:2500], 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            fe_str = "    nll_bound : {0:.4f}".format(fe_mean)
            print(fe_str)
            out_file.write(fe_str + "\n")
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
            # compute information about posterior KLds on validation set
            file_name = RESULT_PATH + "pt_osm_post_klds_b{0:d}.png".format(i)
            post_klds = OSM.compute_post_klds(Xva[0:2500])
            post_dim_klds = np.mean(post_klds, axis=0)
            utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, \
                    file_name)
        if ((i % 5000) == 0):
            IN.save_to_file(f_name=RESULT_PATH +
                            "pt_osm_params_b{0:d}_IN.pkl".format(i))
            GN.save_to_file(f_name=RESULT_PATH +
                            "pt_osm_params_b{0:d}_GN.pkl".format(i))
    IN.save_to_file(f_name=RESULT_PATH + "pt_osm_params_IN.pkl")
    GN.save_to_file(f_name=RESULT_PATH + "pt_osm_params_GN.pkl")
    return
Ejemplo n.º 7
0
def test_gip_sigma_scale_mnist():
    from LogPDFs import cross_validate_sigma

    # Simple test code, to check that everything is basically functional.
    print("TESTING...")

    # Initialize a source of randomness
    rng = np.random.RandomState(12345)

    # Load some data to train/validate/test with
    dataset = "data/mnist.pkl.gz"
    datasets = load_udm(dataset, zero_mean=False)
    Xtr = datasets[0][0]
    Xtr = Xtr.get_value(borrow=False)
    Xva = datasets[2][0]
    Xva = Xva.get_value(borrow=False)
    print("Xtr.shape: {0:s}, Xva.shape: {1:s}".format(str(Xtr.shape), str(Xva.shape)))

    # get and set some basic dataset information
    tr_samples = Xtr.shape[0]
    batch_size = 100
    Xtr_mean = np.mean(Xtr, axis=0, keepdims=True)
    Xtr_mean = (0.0 * Xtr_mean) + np.mean(Xtr)
    Xc_mean = np.repeat(Xtr_mean, batch_size, axis=0).astype(theano.config.floatX)

    # Symbolic inputs
    Xd = T.matrix(name="Xd")
    Xc = T.matrix(name="Xc")
    Xm = T.matrix(name="Xm")
    Xt = T.matrix(name="Xt")

    # Load inferencer and generator from saved parameters
    gn_fname = "MNIST_WALKOUT_TEST_MED_KLD/pt_osm_params_b80000_GN.pkl"
    in_fname = "MNIST_WALKOUT_TEST_MED_KLD/pt_osm_params_b80000_IN.pkl"
    IN = load_infnet_from_file(f_name=in_fname, rng=rng, Xd=Xd)
    GN = load_infnet_from_file(f_name=gn_fname, rng=rng, Xd=Xd)
    x_dim = IN.shared_layers[0].in_dim
    z_dim = IN.mu_layers[-1].out_dim
    # construct a GIPair with the loaded InfNet and GenNet
    osm_params = {}
    osm_params["x_type"] = "gaussian"
    osm_params["xt_transform"] = "sigmoid"
    osm_params["logvar_bound"] = LOGVAR_BOUND
    OSM = OneStageModel(
        rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, p_x_given_z=GN, q_z_given_x=IN, x_dim=x_dim, z_dim=z_dim, params=osm_params
    )
    # compute variational likelihood bound and its sub-components
    Xva = row_shuffle(Xva)
    Xb = Xva[0:5000]
    file_name = "AX_MNIST_MAX_KLD_POST_KLDS.png"
    post_klds = OSM.compute_post_klds(Xb)
    post_dim_klds = np.mean(post_klds, axis=0)
    utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, file_name)
    # compute information about free-energy on validation set
    file_name = "AX_MNIST_MAX_KLD_FREE_ENERGY.png"
    fe_terms = OSM.compute_fe_terms(Xb, 20)
    utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, x_label="Posterior KLd", y_label="Negative Log-likelihood")

    # bound_results = OSM.compute_ll_bound(Xva)
    # ll_bounds = bound_results[0]
    # post_klds = bound_results[1]
    # log_likelihoods = bound_results[2]
    # max_lls = bound_results[3]
    # print("mean ll bound: {0:.4f}".format(np.mean(ll_bounds)))
    # print("mean posterior KLd: {0:.4f}".format(np.mean(post_klds)))
    # print("mean log-likelihood: {0:.4f}".format(np.mean(log_likelihoods)))
    # print("mean max log-likelihood: {0:.4f}".format(np.mean(max_lls)))
    # print("min ll bound: {0:.4f}".format(np.min(ll_bounds)))
    # print("max posterior KLd: {0:.4f}".format(np.max(post_klds)))
    # print("min log-likelihood: {0:.4f}".format(np.min(log_likelihoods)))
    # print("min max log-likelihood: {0:.4f}".format(np.min(max_lls)))
    # # compute some information about the approximate posteriors
    # post_stats = OSM.compute_post_stats(Xva, 0.0*Xva, 0.0*Xva)
    # all_post_klds = np.sort(post_stats[0].ravel()) # post KLds for each obs and dim
    # obs_post_klds = np.sort(post_stats[1]) # summed post KLds for each obs
    # post_dim_klds = post_stats[2] # average post KLds for each post dim
    # post_dim_vars = post_stats[3] # average squared mean for each post dim
    # utils.plot_line(np.arange(all_post_klds.shape[0]), all_post_klds, "AAA_ALL_POST_KLDS.png")
    # utils.plot_line(np.arange(obs_post_klds.shape[0]), obs_post_klds, "AAA_OBS_POST_KLDS.png")
    # utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, "AAA_POST_DIM_KLDS.png")
    # utils.plot_stem(np.arange(post_dim_vars.shape[0]), post_dim_vars, "AAA_POST_DIM_VARS.png")

    # draw many samples from the GIP
    for i in range(5):
        tr_idx = npr.randint(low=0, high=tr_samples, size=(100,))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xs = []
        for row in range(3):
            Xs.append([])
            for col in range(3):
                sample_lists = OSM.sample_from_chain(Xd_batch[0:10, :], loop_iters=100, sigma_scale=1.0)
                Xs[row].append(group_chains(sample_lists["data samples"]))
        Xs, block_im_dim = block_video(Xs, (28, 28), (3, 3))
        to_video(Xs, block_im_dim, "AX_MNIST_MAX_KLD_CHAIN_VIDEO_{0:d}.avi".format(i), frame_rate=10)
    file_name = "AX_MNIST_MAX_KLD_PRIOR_SAMPLE.png"
    Xs = OSM.sample_from_prior(20 * 20)
    utils.visualize_samples(Xs, file_name, num_rows=20)
    # # test Parzen density estimator built from prior samples
    # Xs = OSM.sample_from_prior(10000)
    # [best_sigma, best_ll, best_lls] = \
    #         cross_validate_sigma(Xs, Xva, [0.12, 0.14, 0.15, 0.16, 0.18], 20)
    # sort_idx = np.argsort(best_lls)
    # sort_idx = sort_idx[0:400]
    # utils.plot_line(np.arange(sort_idx.shape[0]), best_lls[sort_idx], "A_MNIST_MAX_KLD_BEST_LLS_1.png")
    # utils.visualize_samples(Xva[sort_idx], "A_MNIST_MAX_KLD_BAD_DIGITS_1.png", num_rows=20)
    return
Ejemplo n.º 8
0
def test_with_model_init():
    ##########################
    # Get some training data #
    ##########################
    rng = np.random.RandomState(1234)
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, as_shared=False, zero_mean=False)
    Xtr = to_fX(datasets[0][0])
    Xva = to_fX(datasets[1][0])
    Ytr = datasets[0][1]
    Yva = datasets[1][1]
    Xtr_class_groups = make_class_groups(Xtr, Ytr)

    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]
    batch_size = 300

    BD = lambda ary: binarize_data(ary)

    ############################################################
    # Setup some parameters for the Iterative Refinement Model #
    ############################################################
    obs_dim = Xtr.shape[1]
    z_dim = 32
    h_dim = 100
    ir_steps = 2
    init_scale = 1.0
    
    x_type = 'bernoulli'

    # some InfNet instances to build the TwoStageModel from
    x_in = T.matrix('x_in')
    x_pos = T.matrix('x_pos')
    y_in = T.lvector('y_in')

    #################
    # p_hi_given_si #
    #################
    params = {}
    shared_config = [obs_dim, 500, 500]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_hi_given_si = InfNet(rng=rng, Xd=x_in, \
            params=params, shared_param_dicts=None)
    p_hi_given_si.init_biases(0.2)
    ######################
    # p_sip1_given_si_hi #
    ######################
    params = {}
    shared_config = [(h_dim + obs_dim), 500, 500]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_sip1_given_si_hi = InfNet(rng=rng, Xd=x_in, \
            params=params, shared_param_dicts=None)
    p_sip1_given_si_hi.init_biases(0.2)
    ################
    # p_s0_given_z #
    ################
    params = {}
    shared_config = [z_dim, 500, 500]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_s0_given_z = InfNet(rng=rng, Xd=x_in, \
            params=params, shared_param_dicts=None)
    p_s0_given_z.init_biases(0.2)
    ###############
    # q_z_given_x #
    ###############
    params = {}
    shared_config = [obs_dim, (500, 4), (500, 4)]
    top_config = [shared_config[-1], z_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.2
    params['hid_drop'] = 0.5
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_z_given_x = InfNet(rng=rng, Xd=x_in, \
            params=params, shared_param_dicts=None)
    q_z_given_x.init_biases(0.0)
    ###################
    # q_hi_given_x_si #
    ###################
    params = {}
    shared_config = [(obs_dim + obs_dim), 800, 800]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = relu_actfun
    params['init_scale'] = init_scale
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_hi_given_x_si = InfNet(rng=rng, Xd=x_in, \
            params=params, shared_param_dicts=None)
    q_hi_given_x_si.init_biases(0.2)


    ################################################################
    # Define parameters for the MultiStageModel, and initialize it #
    ################################################################
    print("Building the MultiStageModel...")
    msm_params = {}
    msm_params['x_type'] = x_type
    msm_params['obs_transform'] = 'sigmoid'
    MSM = MultiStageModelSS(rng=rng, \
            x_in=x_in, x_pos=x_pos, y_in=y_in, \
            p_s0_given_z=p_s0_given_z, \
            p_hi_given_si=p_hi_given_si, \
            p_sip1_given_si_hi=p_sip1_given_si_hi, \
            q_z_given_x=q_z_given_x, \
            q_hi_given_x_si=q_hi_given_x_si, \
            class_count=10, \
            obs_dim=obs_dim, z_dim=z_dim, h_dim=h_dim, \
            ir_steps=ir_steps, params=msm_params)
    MSM.set_lam_class(lam_class=20.0)
    MSM.set_lam_nll(lam_nll=1.0)
    MSM.set_lam_kld(lam_kld_z=1.0, lam_kld_q2p=0.9, \
                    lam_kld_p2q=0.1)
    MSM.set_lam_l2w(1e-4)
    MSM.set_drop_rate(0.0)
    MSM.q_hi_given_x_si.set_bias_noise(0.0)
    MSM.p_hi_given_si.set_bias_noise(0.0)
    MSM.p_sip1_given_si_hi.set_bias_noise(0.0)

    ################################################################
    # Apply some updates, to check that they aren't totally broken #
    ################################################################
    out_file = open("MSS_A_RESULTS.txt", 'wb')
    costs = [0. for i in range(10)]
    learn_rate = 0.0002
    momentum = 0.5
    batch_idx = np.arange(batch_size) + tr_samples
    for i in range(250000):
        scale = min(1.0, ((i+1) / 2000.0))
        if (((i + 1) % 10000) == 0):
            learn_rate = learn_rate * 0.95
        if (i > 20000):
            momentum = 0.90
        else:
            momentum = 0.50
        # get the indices of training samples for this batch update
        batch_idx += batch_size
        if (np.max(batch_idx) >= tr_samples):
            # we finished an "epoch", so we rejumble the training set
            Xtr, Ytr = row_shuffle(Xtr, Ytr)
            batch_idx = np.arange(batch_size)
        # set sgd and objective function hyperparams for this update
        MSM.set_sgd_params(lr_1=scale*learn_rate, lr_2=scale*learn_rate, \
                           mom_1=scale*momentum, mom_2=0.99)
        MSM.set_train_switch(1.0)
        # perform a minibatch update and record the cost for this batch
        Xi_tr = Xtr.take(batch_idx, axis=0)
        Yi_tr = Ytr.take(batch_idx, axis=0)
        Xp_tr, Xn_tr = sample_class_groups(Yi_tr, Xtr_class_groups)
        result = MSM.train_joint(BD(Xi_tr), BD(Xp_tr), Yi_tr)
        costs = [(costs[j] + result[j]) for j in range(len(result)-1)]
        # output useful information about training progress
        if ((i % 500) == 0):
            costs = [(v / 500.0) for v in costs]
            str1 = "-- batch {0:d} --".format(i)
            str2 = "    joint_cost  : {0:.4f}".format(costs[0])
            str3 = "    class_cost  : {0:.4f}".format(costs[1])
            str4 = "    nll_cost    : {0:.4f}".format(costs[2])
            str5 = "    kld_cost    : {0:.4f}".format(costs[3])
            str6 = "    reg_cost    : {0:.4f}".format(costs[4])
            joint_str = "\n".join([str1, str2, str3, str4, str5, str6])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            costs = [0.0 for v in costs]
        if (((i % 2000) == 0) or ((i < 10000) and ((i % 1000) == 0))):
            # Get some validation samples for computing diagnostics
            Xva, Yva = row_shuffle(Xva, Yva)
            Xb_va = Xva[0:2500]
            Yb_va = Yva[0:2500]
            # draw some independent random samples from the model
            samp_count = 200
            model_samps = MSM.sample_from_prior(samp_count)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count):
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MSS_A_SAMPLES_IND_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # draw some conditional random samples from the model
            Xs = Xb_va[0:50] # only use validation set samples
            Xs = np.repeat(Xs, 4, axis=0)
            samp_count = Xs.shape[0]
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # draw some conditional random samples from the model
            model_samps = MSM.sample_from_input(BD(Xs), guided_decoding=False)
            model_samps.append(Xs)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count): 
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MSS_A_SAMPLES_CND_UD_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # compute information about posterior KLds on validation set
            raw_costs = MSM.compute_raw_costs(BD(Xb_va), BD(Xb_va))
            init_nll, init_kld, q2p_kld, p2q_kld, step_nll, step_kld = raw_costs
            file_name = "MSS_A_H0_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(init_kld.shape[1]), \
                    np.mean(init_kld, axis=0), file_name)
            file_name = "MSS_A_HI_Q2P_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(q2p_kld.shape[1]), \
                    np.mean(q2p_kld, axis=0), file_name)
            file_name = "MSS_A_HI_P2Q_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(p2q_kld.shape[1]), \
                    np.mean(p2q_kld, axis=0), file_name)
            # draw weights for the initial encoder/classifier
            file_name = "MSS_A_QZX_WEIGHTS_b{0:d}.png".format(i)
            W = q_z_given_x.shared_layers[0].W.get_value(borrow=False).T
            utils.visualize_samples(W, file_name, num_rows=20)
            # compute free-energy terms on training samples
            fe_terms = MSM.compute_fe_terms(BD(Xtr[0:2500]), BD(Xtr[0:2500]), 30)
            fe_nll = np.mean(fe_terms[0])
            fe_kld = np.mean(fe_terms[1])
            fe_joint = fe_nll + fe_kld
            joint_str = "    vfe-tr: {0:.4f}, nll: ({1:.4f}, {2:.4f}, {3:.4f}), kld: ({4:.4f}, {5:.4f}, {6:.4f})".format( \
                    fe_joint, fe_nll, np.min(fe_terms[0]), np.max(fe_terms[0]), fe_kld, np.min(fe_terms[1]), np.max(fe_terms[1]))
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            # compute free-energy terms on validation samples
            fe_terms = MSM.compute_fe_terms(BD(Xb_va), BD(Xb_va), 30)
            fe_nll = np.mean(fe_terms[0])
            fe_kld = np.mean(fe_terms[1])
            fe_joint = fe_nll + fe_kld
            joint_str = "    vfe-va: {0:.4f}, nll: ({1:.4f}, {2:.4f}, {3:.4f}), kld: ({4:.4f}, {5:.4f}, {6:.4f})".format( \
                    fe_joint, fe_nll, np.min(fe_terms[0]), np.max(fe_terms[0]), fe_kld, np.min(fe_terms[1]), np.max(fe_terms[1]))
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            # compute multi-sample estimate of classification error
            err_rate, err_idx, y_preds = MSM.class_error(Xb_va, Yb_va, \
                    samples=30, prep_func=BD)
            joint_str = "    va-class-error: {0:.4f}".format(err_rate)
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            # draw some conditional random samples from the model
            Xs = Xb_va[err_idx] # use validation samples with class errors
            if (Xs.shape[0] > 50):
                Xs = Xs[:50]
            Xs = np.repeat(Xs, 4, axis=0)
            if ((Xs.shape[0] % 20) != 0):
                # round-off the number of error examples, for nice display
                remainder = Xs.shape[0] % 20
                Xs = Xs[:-remainder]
            samp_count = Xs.shape[0]
            # draw some conditional random samples from the model
            model_samps = MSM.sample_from_input(BD(Xs), guided_decoding=False)
            model_samps.append(Xs)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count): 
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MSS_A_SAMPLES_CND_ERR_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
Ejemplo n.º 9
0
def test_with_model_init():
    ##########################
    # Get some training data #
    ##########################
    rng = np.random.RandomState(1234)
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, zero_mean=False)
    Xtr_shared = datasets[0][0]
    Xva_shared = datasets[1][0]
    Xtr = Xtr_shared.get_value(borrow=False).astype(theano.config.floatX)
    Xva = Xva_shared.get_value(borrow=False).astype(theano.config.floatX)
    tr_samples = Xtr.shape[0]
    batch_size = 500
    batch_reps = 1

    ############################################################
    # Setup some parameters for the Iterative Refinement Model #
    ############################################################
    obs_dim = Xtr.shape[1]
    z_rnn_dim = 25
    z_obs_dim = 5
    jnt_dim = obs_dim + z_rnn_dim
    h_dim = 100
    x_type = 'bernoulli'
    prior_sigma = 1.0

    # some InfNet instances to build the TwoStageModel from
    X_sym = T.matrix('X_sym')

    ########################
    # p_s0_obs_given_z_obs #
    ########################
    params = {}
    shared_config = [z_obs_dim, 250, 250]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = softplus_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 1e-3
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_s0_obs_given_z_obs = InfNet(rng=rng, Xd=X_sym, prior_sigma=prior_sigma, \
            params=params, shared_param_dicts=None)
    p_s0_obs_given_z_obs.init_biases(0.2)
    #################
    # p_hi_given_si #
    #################
    params = {}
    shared_config = [jnt_dim, 500, 500]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = softplus_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_hi_given_si = InfNet(rng=rng, Xd=X_sym, prior_sigma=prior_sigma, \
            params=params, shared_param_dicts=None)
    p_hi_given_si.init_biases(0.2)
    ######################
    # p_sip1_given_si_hi #
    ######################
    params = {}
    shared_config = [(h_dim + z_rnn_dim), 500, 500]
    top_config = [shared_config[-1], obs_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = softplus_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_sip1_given_si_hi = InfNet(rng=rng, Xd=X_sym, prior_sigma=prior_sigma, \
            params=params, shared_param_dicts=None)
    p_sip1_given_si_hi.init_biases(0.2)
    ###############
    # q_z_given_x #
    ###############
    params = {}
    shared_config = [obs_dim, 250, 250]
    top_config = [shared_config[-1], (z_rnn_dim + z_obs_dim)]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = softplus_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_z_given_x = InfNet(rng=rng, Xd=X_sym, prior_sigma=prior_sigma, \
            params=params, shared_param_dicts=None)
    q_z_given_x.init_biases(0.2)
    ###################
    # q_hi_given_x_si #
    ###################
    params = {}
    shared_config = [(obs_dim + jnt_dim), 500, 500]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = softplus_actfun
    params['init_scale'] = 1.2
    params['lam_l2a'] = 0.0
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_hi_given_x_si = InfNet(rng=rng, Xd=X_sym, prior_sigma=prior_sigma, \
            params=params, shared_param_dicts=None)
    q_hi_given_x_si.init_biases(0.2)


    ################################################################
    # Define parameters for the MultiStageModel, and initialize it #
    ################################################################
    print("Building the MultiStageModel...")
    msm_params = {}
    msm_params['x_type'] = x_type
    msm_params['obs_transform'] = 'sigmoid'
    MSM = MultiStageModel(rng=rng, x_in=X_sym, \
            p_s0_obs_given_z_obs=p_s0_obs_given_z_obs, \
            p_hi_given_si=p_hi_given_si, \
            p_sip1_given_si_hi=p_sip1_given_si_hi, \
            q_z_given_x=q_z_given_x, \
            q_hi_given_x_si=q_hi_given_x_si, \
            obs_dim=obs_dim, z_rnn_dim=z_rnn_dim, z_obs_dim=z_obs_dim, \
            h_dim=h_dim, model_init_obs=False, model_init_rnn=True, \
            ir_steps=3, params=msm_params)
    obs_mean = (0.9 * np.mean(Xtr, axis=0)) + 0.05
    obs_mean_logit = np.log(obs_mean / (1.0 - obs_mean))
    MSM.set_input_bias(-obs_mean)
    MSM.set_obs_bias(0.1*obs_mean_logit)

    ################################################################
    # Apply some updates, to check that they aren't totally broken #
    ################################################################
    costs = [0. for i in range(10)]
    learn_rate = 0.003
    momentum = 0.5
    for i in range(300000):
        scale = min(1.0, ((i+1) / 5000.0))
        l1l2_weight = 1.0 #min(1.0, ((i+1) / 2500.0))
        if (((i + 1) % 10000) == 0):
            learn_rate = learn_rate * 0.92
        if (i > 100000):
            momentum = 0.80
        if (i > 50000):
            momentum = 0.65
        else:
            momentum = 0.50
        # randomly sample a minibatch
        tr_idx = npr.randint(low=0,high=tr_samples,size=(batch_size,))
        Xb = binarize_data(Xtr.take(tr_idx, axis=0))
        Xb = Xb.astype(theano.config.floatX)
        # set sgd and objective function hyperparams for this update
        MSM.set_sgd_params(lr_1=scale*learn_rate, lr_2=scale*learn_rate, \
                mom_1=(scale*momentum), mom_2=0.99)
        MSM.set_train_switch(1.0)
        MSM.set_l1l2_weight(l1l2_weight)
        MSM.set_lam_nll(lam_nll=1.0)
        MSM.set_lam_kld(lam_kld_1=1.0, lam_kld_2=1.0)
        MSM.set_lam_l2w(1e-5)
        MSM.set_kzg_weight(0.01)
        # perform a minibatch update and record the cost for this batch
        result = MSM.train_joint(Xb, batch_reps)
        costs = [(costs[j] + result[j]) for j in range(len(result))]
        if ((i % 500) == 0):
            costs = [(v / 500.0) for v in costs]
            print("-- batch {0:d} --".format(i))
            print("    joint_cost: {0:.4f}".format(costs[0]))
            print("    nll_cost  : {0:.4f}".format(costs[1]))
            print("    kld_cost  : {0:.4f}".format(costs[2]))
            print("    reg_cost  : {0:.4f}".format(costs[3]))
            costs = [0.0 for v in costs]
        if (((i % 2000) == 0) or ((i < 10000) and ((i % 1000) == 0))):
            Xva = row_shuffle(Xva)
            # draw some independent random samples from the model
            samp_count = 200
            model_samps = MSM.sample_from_prior(samp_count)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count): 
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "MZ_SAMPLES_b{0:d}.png".format(i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # visualize some important weights in the model
            file_name = "MZ_INF_1_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.inf_1_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MZ_INF_2_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.inf_2_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MZ_GEN_1_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_1_weights.get_value(borrow=False)
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MZ_GEN_2_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_2_weights.get_value(borrow=False)
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            file_name = "MZ_GEN_INF_WEIGHTS_b{0:d}.png".format(i)
            W = MSM.gen_inf_weights.get_value(borrow=False).T
            utils.visualize_samples(W[:,:obs_dim], file_name, num_rows=20)
            # compute information about posterior KLds on validation set
            post_klds = MSM.compute_post_klds(Xva[0:5000])
            file_name = "MZ_H0_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[0].shape[1]), \
                    np.mean(post_klds[0], axis=0), file_name)
            file_name = "MZ_HI_COND_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[1].shape[1]), \
                    np.mean(post_klds[1], axis=0), file_name)
            file_name = "MZ_HI_GLOB_KLDS_b{0:d}.png".format(i)
            utils.plot_stem(np.arange(post_klds[2].shape[1]), \
                    np.mean(post_klds[2], axis=0), file_name)
            # compute information about free-energy on validation set
            file_name = "MZ_FREE_ENERGY_b{0:d}.png".format(i)
            fe_terms = MSM.compute_fe_terms(binarize_data(Xva[0:5000]), 20)
            fe_mean = np.mean(fe_terms[0]) + np.mean(fe_terms[1])
            print("    nll_bound : {0:.4f}".format(fe_mean))
            utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
                    x_label='Posterior KLd', y_label='Negative Log-likelihood')
    return
Ejemplo n.º 10
0
def test_gip_sigma_scale_mnist():
    from LogPDFs import cross_validate_sigma
    # Simple test code, to check that everything is basically functional.
    print("TESTING...")

    # Initialize a source of randomness
    rng = np.random.RandomState(12345)

    # Load some data to train/validate/test with
    dataset = 'data/mnist.pkl.gz'
    datasets = load_udm(dataset, zero_mean=False)
    Xtr = datasets[0][0]
    Xtr = Xtr.get_value(borrow=False)
    Xva = datasets[2][0]
    Xva = Xva.get_value(borrow=False)
    print("Xtr.shape: {0:s}, Xva.shape: {1:s}".format(str(Xtr.shape),
                                                      str(Xva.shape)))

    # get and set some basic dataset information
    tr_samples = Xtr.shape[0]
    batch_size = 100
    Xtr_mean = np.mean(Xtr, axis=0, keepdims=True)
    Xtr_mean = (0.0 * Xtr_mean) + np.mean(Xtr)
    Xc_mean = np.repeat(Xtr_mean, batch_size,
                        axis=0).astype(theano.config.floatX)

    # Symbolic inputs
    Xd = T.matrix(name='Xd')
    Xc = T.matrix(name='Xc')
    Xm = T.matrix(name='Xm')
    Xt = T.matrix(name='Xt')

    # Load inferencer and generator from saved parameters
    gn_fname = "MNIST_WALKOUT_TEST_MED_KLD/pt_osm_params_b80000_GN.pkl"
    in_fname = "MNIST_WALKOUT_TEST_MED_KLD/pt_osm_params_b80000_IN.pkl"
    IN = load_infnet_from_file(f_name=in_fname, rng=rng, Xd=Xd)
    GN = load_infnet_from_file(f_name=gn_fname, rng=rng, Xd=Xd)
    x_dim = IN.shared_layers[0].in_dim
    z_dim = IN.mu_layers[-1].out_dim
    # construct a GIPair with the loaded InfNet and GenNet
    osm_params = {}
    osm_params['x_type'] = 'gaussian'
    osm_params['xt_transform'] = 'sigmoid'
    osm_params['logvar_bound'] = LOGVAR_BOUND
    OSM = OneStageModel(rng=rng, Xd=Xd, Xc=Xc, Xm=Xm, \
            p_x_given_z=GN, q_z_given_x=IN, \
            x_dim=x_dim, z_dim=z_dim, params=osm_params)
    # compute variational likelihood bound and its sub-components
    Xva = row_shuffle(Xva)
    Xb = Xva[0:5000]
    file_name = "AX_MNIST_MAX_KLD_POST_KLDS.png"
    post_klds = OSM.compute_post_klds(Xb)
    post_dim_klds = np.mean(post_klds, axis=0)
    utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, \
            file_name)
    # compute information about free-energy on validation set
    file_name = "AX_MNIST_MAX_KLD_FREE_ENERGY.png"
    fe_terms = OSM.compute_fe_terms(Xb, 20)
    utils.plot_scatter(fe_terms[1], fe_terms[0], file_name, \
            x_label='Posterior KLd', y_label='Negative Log-likelihood')

    # bound_results = OSM.compute_ll_bound(Xva)
    # ll_bounds = bound_results[0]
    # post_klds = bound_results[1]
    # log_likelihoods = bound_results[2]
    # max_lls = bound_results[3]
    # print("mean ll bound: {0:.4f}".format(np.mean(ll_bounds)))
    # print("mean posterior KLd: {0:.4f}".format(np.mean(post_klds)))
    # print("mean log-likelihood: {0:.4f}".format(np.mean(log_likelihoods)))
    # print("mean max log-likelihood: {0:.4f}".format(np.mean(max_lls)))
    # print("min ll bound: {0:.4f}".format(np.min(ll_bounds)))
    # print("max posterior KLd: {0:.4f}".format(np.max(post_klds)))
    # print("min log-likelihood: {0:.4f}".format(np.min(log_likelihoods)))
    # print("min max log-likelihood: {0:.4f}".format(np.min(max_lls)))
    # # compute some information about the approximate posteriors
    # post_stats = OSM.compute_post_stats(Xva, 0.0*Xva, 0.0*Xva)
    # all_post_klds = np.sort(post_stats[0].ravel()) # post KLds for each obs and dim
    # obs_post_klds = np.sort(post_stats[1]) # summed post KLds for each obs
    # post_dim_klds = post_stats[2] # average post KLds for each post dim
    # post_dim_vars = post_stats[3] # average squared mean for each post dim
    # utils.plot_line(np.arange(all_post_klds.shape[0]), all_post_klds, "AAA_ALL_POST_KLDS.png")
    # utils.plot_line(np.arange(obs_post_klds.shape[0]), obs_post_klds, "AAA_OBS_POST_KLDS.png")
    # utils.plot_stem(np.arange(post_dim_klds.shape[0]), post_dim_klds, "AAA_POST_DIM_KLDS.png")
    # utils.plot_stem(np.arange(post_dim_vars.shape[0]), post_dim_vars, "AAA_POST_DIM_VARS.png")

    # draw many samples from the GIP
    for i in range(5):
        tr_idx = npr.randint(low=0, high=tr_samples, size=(100, ))
        Xd_batch = Xtr.take(tr_idx, axis=0)
        Xs = []
        for row in range(3):
            Xs.append([])
            for col in range(3):
                sample_lists = OSM.sample_from_chain(Xd_batch[0:10,:], loop_iters=100, \
                        sigma_scale=1.0)
                Xs[row].append(group_chains(sample_lists['data samples']))
        Xs, block_im_dim = block_video(Xs, (28, 28), (3, 3))
        to_video(Xs,
                 block_im_dim,
                 "AX_MNIST_MAX_KLD_CHAIN_VIDEO_{0:d}.avi".format(i),
                 frame_rate=10)
    file_name = "AX_MNIST_MAX_KLD_PRIOR_SAMPLE.png"
    Xs = OSM.sample_from_prior(20 * 20)
    utils.visualize_samples(Xs, file_name, num_rows=20)
    # # test Parzen density estimator built from prior samples
    # Xs = OSM.sample_from_prior(10000)
    # [best_sigma, best_ll, best_lls] = \
    #         cross_validate_sigma(Xs, Xva, [0.12, 0.14, 0.15, 0.16, 0.18], 20)
    # sort_idx = np.argsort(best_lls)
    # sort_idx = sort_idx[0:400]
    # utils.plot_line(np.arange(sort_idx.shape[0]), best_lls[sort_idx], "A_MNIST_MAX_KLD_BEST_LLS_1.png")
    # utils.visualize_samples(Xva[sort_idx], "A_MNIST_MAX_KLD_BAD_DIGITS_1.png", num_rows=20)
    return
def test_mnist(step_type='add',
               imp_steps=6,
               occ_dim=15,
               drop_prob=0.0):
    #########################################
    # Format the result tag more thoroughly #
    #########################################
    dp_int = int(100.0 * drop_prob)
    result_tag = "{}GPSI_OD{}_DP{}_IS{}_{}_NA".format(RESULT_PATH, occ_dim, dp_int, imp_steps, step_type)

    ##########################
    # Get some training data #
    ##########################
    rng = np.random.RandomState(1234)
    Xtr, Xva, Xte = load_binarized_mnist(data_path='./data/')
    Xtr = np.vstack((Xtr, Xva))
    Xva = Xte
    #del Xte
    tr_samples = Xtr.shape[0]
    va_samples = Xva.shape[0]

    ##########################
    # Get some training data #
    ##########################
    # rng = np.random.RandomState(1234)
    # dataset = 'data/mnist.pkl.gz'
    # datasets = load_udm(dataset, as_shared=False, zero_mean=False)
    # Xtr = datasets[0][0]
    # Xva = datasets[1][0]
    # Xte = datasets[2][0]
    # # Merge validation set and training set, and test on test set.
    # #Xtr = np.concatenate((Xtr, Xva), axis=0)
    # #Xva = Xte
    # Xtr = to_fX(shift_and_scale_into_01(Xtr))
    # Xva = to_fX(shift_and_scale_into_01(Xva))
    # tr_samples = Xtr.shape[0]
    # va_samples = Xva.shape[0]
    batch_size = 200
    batch_reps = 1
    all_pix_mean = np.mean(np.mean(Xtr, axis=1))
    data_mean = to_fX( all_pix_mean * np.ones((Xtr.shape[1],)) )

    ############################################################
    # Setup some parameters for the Iterative Refinement Model #
    ############################################################
    x_dim = Xtr.shape[1]
    s_dim = x_dim
    h_dim = 50
    z_dim = 100
    init_scale = 0.6

    x_in_sym = T.matrix('x_in_sym')
    x_out_sym = T.matrix('x_out_sym')
    x_mask_sym = T.matrix('x_mask_sym')

    ###############
    # p_h_given_x #
    ###############
    params = {}
    shared_config = [x_dim, 250]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = 'xg' #init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_h_given_x = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_h_given_x.init_biases(0.0)
    ################
    # p_s0_given_h #
    ################
    params = {}
    shared_config = [h_dim, 250]
    output_config = [s_dim, s_dim, s_dim]
    params['shared_config'] = shared_config
    params['output_config'] = output_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = 'xg' #init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_s0_given_h = HydraNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_s0_given_h.init_biases(0.0)
    #################
    # p_zi_given_xi #
    #################
    params = {}
    shared_config = [(x_dim + x_dim), 500, 500]
    top_config = [shared_config[-1], z_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_zi_given_xi = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_zi_given_xi.init_biases(0.0)
    ###################
    # p_sip1_given_zi #
    ###################
    params = {}
    shared_config = [z_dim, 500, 500]
    output_config = [s_dim, s_dim, s_dim]
    params['shared_config'] = shared_config
    params['output_config'] = output_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_sip1_given_zi = HydraNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_sip1_given_zi.init_biases(0.0)
    ################
    # p_x_given_si #
    ################
    params = {}
    shared_config = [s_dim]
    output_config = [x_dim, x_dim]
    params['shared_config'] = shared_config
    params['output_config'] = output_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    p_x_given_si = HydraNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    p_x_given_si.init_biases(0.0)
    ###############
    # q_h_given_x #
    ###############
    params = {}
    shared_config = [x_dim, 250]
    top_config = [shared_config[-1], h_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = 'xg' #init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_h_given_x = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    q_h_given_x.init_biases(0.0)
    #################
    # q_zi_given_xi #
    #################
    params = {}
    shared_config = [(x_dim + x_dim), 500, 500]
    top_config = [shared_config[-1], z_dim]
    params['shared_config'] = shared_config
    params['mu_config'] = top_config
    params['sigma_config'] = top_config
    params['activation'] = tanh_actfun #relu_actfun
    params['init_scale'] = init_scale
    params['vis_drop'] = 0.0
    params['hid_drop'] = 0.0
    params['bias_noise'] = 0.0
    params['input_noise'] = 0.0
    params['build_theano_funcs'] = False
    q_zi_given_xi = InfNet(rng=rng, Xd=x_in_sym, \
            params=params, shared_param_dicts=None)
    q_zi_given_xi.init_biases(0.0)

    ###########################################################
    # Define parameters for the GPSImputer, and initialize it #
    ###########################################################
    print("Building the GPSImputer...")
    gpsi_params = {}
    gpsi_params['x_dim'] = x_dim
    gpsi_params['h_dim'] = h_dim
    gpsi_params['z_dim'] = z_dim
    gpsi_params['s_dim'] = s_dim
    # switch between direct construction and construction via p_x_given_si
    gpsi_params['use_p_x_given_si'] = False
    gpsi_params['imp_steps'] = imp_steps
    gpsi_params['step_type'] = step_type
    gpsi_params['x_type'] = 'bernoulli'
    gpsi_params['obs_transform'] = 'sigmoid'
    GPSI = GPSImputerWI(rng=rng,
            x_in=x_in_sym, x_out=x_out_sym, x_mask=x_mask_sym, \
            p_h_given_x=p_h_given_x, \
            p_s0_given_h=p_s0_given_h, \
            p_zi_given_xi=p_zi_given_xi, \
            p_sip1_given_zi=p_sip1_given_zi, \
            p_x_given_si=p_x_given_si, \
            q_h_given_x=q_h_given_x, \
            q_zi_given_xi=q_zi_given_xi, \
            params=gpsi_params, \
            shared_param_dicts=None)

    ################################################################
    # Apply some updates, to check that they aren't totally broken #
    ################################################################
    log_name = "{}_RESULTS.txt".format(result_tag)
    out_file = open(log_name, 'wb')
    costs = [0. for i in range(10)]
    learn_rate = 0.0002
    momentum = 0.5
    batch_idx = np.arange(batch_size) + tr_samples
    for i in range(250000):
        scale = min(1.0, ((i+1) / 5000.0))
        lam_scale = 1.0 - min(1.0, ((i+1) / 100000.0)) # decays from 1.0->0.0
        if (((i + 1) % 15000) == 0):
            learn_rate = learn_rate * 0.93
        if (i > 10000):
            momentum = 0.90
        else:
            momentum = 0.75
        # get the indices of training samples for this batch update
        batch_idx += batch_size
        if (np.max(batch_idx) >= tr_samples):
            # we finished an "epoch", so we rejumble the training set
            Xtr = row_shuffle(Xtr)
            batch_idx = np.arange(batch_size)
        # set sgd and objective function hyperparams for this update
        GPSI.set_sgd_params(lr=scale*learn_rate, \
                            mom_1=scale*momentum, mom_2=0.98)
        GPSI.set_train_switch(1.0)
        GPSI.set_lam_nll(lam_nll=1.0)
        GPSI.set_lam_kld(lam_kld_p=0.05, lam_kld_q=0.95, \
                         lam_kld_g=(0.1 * lam_scale), lam_kld_s=(0.1 * lam_scale))
        GPSI.set_lam_l2w(1e-5)
        # perform a minibatch update and record the cost for this batch
        xb = to_fX( Xtr.take(batch_idx, axis=0) )
        xi, xo, xm = construct_masked_data(xb, drop_prob=drop_prob, \
                                        occ_dim=occ_dim, data_mean=data_mean)
        result = GPSI.train_joint(xi, xo, xm, batch_reps)
        # do diagnostics and general training tracking
        costs = [(costs[j] + result[j]) for j in range(len(result)-1)]
        if ((i % 250) == 0):
            costs = [(v / 250.0) for v in costs]
            str1 = "-- batch {0:d} --".format(i)
            str2 = "    joint_cost: {0:.4f}".format(costs[0])
            str3 = "    nll_bound : {0:.4f}".format(costs[1])
            str4 = "    nll_cost  : {0:.4f}".format(costs[2])
            str5 = "    kld_cost  : {0:.4f}".format(costs[3])
            str6 = "    reg_cost  : {0:.4f}".format(costs[4])
            joint_str = "\n".join([str1, str2, str3, str4, str5, str6])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
            costs = [0.0 for v in costs]
        if ((i % 1000) == 0):
            Xva = row_shuffle(Xva)
            # record an estimate of performance on the test set
            xi, xo, xm = construct_masked_data(Xva[0:5000], drop_prob=drop_prob, \
                                               occ_dim=occ_dim, data_mean=data_mean)
            nll, kld = GPSI.compute_fe_terms(xi, xo, xm, sample_count=10)
            vfe = np.mean(nll) + np.mean(kld)
            str1 = "    va_nll_bound : {}".format(vfe)
            str2 = "    va_nll_term  : {}".format(np.mean(nll))
            str3 = "    va_kld_q2p   : {}".format(np.mean(kld))
            joint_str = "\n".join([str1, str2, str3])
            print(joint_str)
            out_file.write(joint_str+"\n")
            out_file.flush()
        if ((i % 2000) == 0):
            GPSI.save_to_file("{}_PARAMS.pkl".format(result_tag))
            # Get some validation samples for evaluating model performance
            xb = to_fX( Xva[0:100] )
            xi, xo, xm = construct_masked_data(xb, drop_prob=drop_prob, \
                                    occ_dim=occ_dim, data_mean=data_mean)
            xi = np.repeat(xi, 2, axis=0)
            xo = np.repeat(xo, 2, axis=0)
            xm = np.repeat(xm, 2, axis=0)
            # draw some sample imputations from the model
            samp_count = xi.shape[0]
            _, model_samps = GPSI.sample_imputer(xi, xo, xm, use_guide_policy=False)
            seq_len = len(model_samps)
            seq_samps = np.zeros((seq_len*samp_count, model_samps[0].shape[1]))
            idx = 0
            for s1 in range(samp_count):
                for s2 in range(seq_len):
                    seq_samps[idx] = model_samps[s2][s1]
                    idx += 1
            file_name = "{0:s}_samples_ng_b{1:d}.png".format(result_tag, i)
            utils.visualize_samples(seq_samps, file_name, num_rows=20)
            # show KLds and NLLs on a step-by-step basis
            xb = to_fX( Xva[0:1000] )
            xi, xo, xm = construct_masked_data(xb, drop_prob=drop_prob, \
                                    occ_dim=occ_dim, data_mean=data_mean)
            step_costs = GPSI.compute_per_step_cost(xi, xo, xm)
            step_nlls = step_costs[0]
            step_klds = step_costs[1]
            step_nums = np.arange(step_nlls.shape[0])
            file_name = "{0:s}_NLL_b{1:d}.png".format(result_tag, i)
            utils.plot_stem(step_nums, step_nlls, file_name)
            file_name = "{0:s}_KLD_b{1:d}.png".format(result_tag, i)
            utils.plot_stem(step_nums, step_klds, file_name)