Python train.Training类代码示例

nodes = [h1, h2, h3, h4, cost]
rnn = Net(inputs=inputs, inputs_dim=inputs_dim, nodes=nodes)
cost = unpack(rnn.build_recurrent_graph(output_args=[cost]))
cost = cost.mean()
cost.name = 'cost'
model.graphs = [rnn]

optimizer = Adam(
    lr=0.001
)

extension = [
    GradientClipping(batch_size=batch_size),
    EpochCount(100),
    Monitoring(freq=100,
               ddout=[cost]),
    Picklize(freq=200, path=save_path)
]

mainloop = Training(
    name='toy_bb_gflstm',
    data=Iterator(trdata, batch_size),
    model=model,
    optimizer=optimizer,
    cost=cost,
    outputs=[cost],
    extension=extension
)
mainloop.run()

#.........这里部分代码省略.........

    params = OrderedDict()

    for node in nodes:
        if node.initialize() is not None:
            params.update(node.initialize())

    params = init_tparams(params)

    s_0 = rnn.get_init_state(batch_size)

    x_1_temp = x_1.fprop([x], params)


    def inner_fn(x_t, s_tm1):

        s_t = rnn.fprop([[x_t], [s_tm1]], params)

        return s_t

    ((s_temp), updates) = theano.scan(fn=inner_fn,
                                      sequences=[x_1_temp],
                                      outputs_info=[s_0])

    for k, v in updates.iteritems():
        k.default_update = v

    s_temp = concatenate([s_0[None, :, :], s_temp[:-1]], axis=0)
    theta_1_temp = theta_1.fprop([s_temp], params)
    theta_mu_temp = theta_mu.fprop([theta_1_temp], params)
    theta_sig_temp = theta_sig.fprop([theta_1_temp], params)
    corr_temp = corr.fprop([theta_1_temp], params)
    binary_temp = binary.fprop([theta_1_temp], params)

    x_shape = x.shape
    x_in = x.reshape((x_shape[0]*x_shape[1], -1))
    theta_mu_in = theta_mu_temp.reshape((x_shape[0]*x_shape[1], -1))
    theta_sig_in = theta_sig_temp.reshape((x_shape[0]*x_shape[1], -1))
    corr_in = corr_temp.reshape((x_shape[0]*x_shape[1], -1))
    binary_in = binary_temp.reshape((x_shape[0]*x_shape[1], -1))

    recon = BiGauss(x_in, theta_mu_in, theta_sig_in, corr_in, binary_in)
    recon = recon.reshape((x_shape[0], x_shape[1]))
    recon = recon * mask
    recon_term = recon.sum(axis=0).mean()
    recon_term.name = 'nll'

    max_x = x.max()
    mean_x = x.mean()
    min_x = x.min()
    max_x.name = 'max_x'
    mean_x.name = 'mean_x'
    min_x.name = 'min_x'

    max_theta_mu = theta_mu_in.max()
    mean_theta_mu = theta_mu_in.mean()
    min_theta_mu = theta_mu_in.min()
    max_theta_mu.name = 'max_theta_mu'
    mean_theta_mu.name = 'mean_theta_mu'
    min_theta_mu.name = 'min_theta_mu'

    max_theta_sig = theta_sig_in.max()
    mean_theta_sig = theta_sig_in.mean()
    min_theta_sig = theta_sig_in.min()
    max_theta_sig.name = 'max_theta_sig'
    mean_theta_sig.name = 'mean_theta_sig'
    min_theta_sig.name = 'min_theta_sig'

    model.inputs = [x, mask]
    model._params = params
    model.nodes = nodes

    optimizer = Adam(
        lr=lr
    )

    extension = [
        GradientClipping(batch_size=batch_size),
        EpochCount(epoch),
        Monitoring(freq=monitoring_freq,
                   ddout=[recon_term,
                          max_theta_sig, mean_theta_sig, min_theta_sig,
                          max_x, mean_x, min_x,
                          max_theta_mu, mean_theta_mu, min_theta_mu],
                   data=[Iterator(valid_data, batch_size)]),
        Picklize(freq=monitoring_freq, path=save_path),
        EarlyStopping(freq=monitoring_freq, path=save_path),
        WeightNorm()
    ]

    mainloop = Training(
        name=pkl_name,
        data=Iterator(train_data, batch_size),
        model=model,
        optimizer=optimizer,
        cost=recon_term,
        outputs=[recon_term],
        extension=extension
    )
    mainloop.run()

#.........这里部分代码省略.........
    theta_sig_in = theta_sig_temp.reshape((x_shape[0] * x_shape[1], -1))
    corr_in = corr_temp.reshape((x_shape[0] * x_shape[1], -1))
    binary_in = binary_temp.reshape((x_shape[0] * x_shape[1], -1))

    recon = BiGauss(x_in, theta_mu_in, theta_sig_in, corr_in, binary_in)
    recon = recon.reshape((x_shape[0], x_shape[1]))
    recon = recon * mask
    recon_term = recon.sum(axis=0).mean()
    recon_term.name = "recon_term"

    kl_temp = kl_temp * mask
    kl_term = kl_temp.sum(axis=0).mean()
    kl_term.name = "kl_term"

    nll_upper_bound = recon_term + kl_term
    nll_upper_bound.name = "nll_upper_bound"

    max_x = x.max()
    mean_x = x.mean()
    min_x = x.min()
    max_x.name = "max_x"
    mean_x.name = "mean_x"
    min_x.name = "min_x"

    max_theta_mu = theta_mu_in.max()
    mean_theta_mu = theta_mu_in.mean()
    min_theta_mu = theta_mu_in.min()
    max_theta_mu.name = "max_theta_mu"
    mean_theta_mu.name = "mean_theta_mu"
    min_theta_mu.name = "min_theta_mu"

    max_theta_sig = theta_sig_in.max()
    mean_theta_sig = theta_sig_in.mean()
    min_theta_sig = theta_sig_in.min()
    max_theta_sig.name = "max_theta_sig"
    mean_theta_sig.name = "mean_theta_sig"
    min_theta_sig.name = "min_theta_sig"

    max_phi_sig = phi_sig_temp.max()
    mean_phi_sig = phi_sig_temp.mean()
    min_phi_sig = phi_sig_temp.min()
    max_phi_sig.name = "max_phi_sig"
    mean_phi_sig.name = "mean_phi_sig"
    min_phi_sig.name = "min_phi_sig"

    max_prior_sig = prior_sig_temp.max()
    mean_prior_sig = prior_sig_temp.mean()
    min_prior_sig = prior_sig_temp.min()
    max_prior_sig.name = "max_prior_sig"
    mean_prior_sig.name = "mean_prior_sig"
    min_prior_sig.name = "min_prior_sig"

    model.inputs = [x, mask]
    model.params = params
    model.nodes = nodes

    optimizer = Adam(lr=lr)

    extension = [
        GradientClipping(batch_size=batch_size),
        EpochCount(epoch),
        Monitoring(
            freq=monitoring_freq,
            ddout=[
                nll_upper_bound,
                recon_term,
                kl_term,
                max_phi_sig,
                mean_phi_sig,
                min_phi_sig,
                max_prior_sig,
                mean_prior_sig,
                min_prior_sig,
                max_theta_sig,
                mean_theta_sig,
                min_theta_sig,
                max_x,
                mean_x,
                min_x,
                max_theta_mu,
                mean_theta_mu,
                min_theta_mu,
            ],
            data=[Iterator(valid_data, batch_size)],
        ),
        Picklize(freq=monitoring_freq, path=save_path),
        EarlyStopping(freq=monitoring_freq, path=save_path, channel=channel_name),
        WeightNorm(),
    ]

    mainloop = Training(
        name=pkl_name,
        data=Iterator(train_data, batch_size),
        model=model,
        optimizer=optimizer,
        cost=nll_upper_bound,
        outputs=[nll_upper_bound],
        extension=extension,
    )
    mainloop.run()

#.........这里部分代码省略.........
    m_theta_sig_temp = theta_sig.fprop([m_theta_4_temp], params)
    m_coeff_temp = coeff.fprop([m_theta_4_temp], params)

    m_kl_temp = KLGaussianGaussian(m_phi_mu_temp, m_phi_sig_temp, m_prior_mu_temp, m_prior_sig_temp)

    m_x_shape = m_x.shape
    m_x_in = m_x.reshape((m_x_shape[0]*m_x_shape[1], -1))
    m_theta_mu_in = m_theta_mu_temp.reshape((m_x_shape[0]*m_x_shape[1], -1))
    m_theta_sig_in = m_theta_sig_temp.reshape((m_x_shape[0]*m_x_shape[1], -1))
    m_coeff_in = m_coeff_temp.reshape((m_x_shape[0]*m_x_shape[1], -1))

    m_recon = GMM(m_x_in, m_theta_mu_in, m_theta_sig_in, m_coeff_in)
    m_recon_term = m_recon.mean()
    m_kl_term = m_kl_temp.mean()
    m_nll_upper_bound = m_recon_term + m_kl_term
    m_nll_upper_bound.name = 'nll_upper_bound'
    m_recon_term.name = 'recon_term'
    m_kl_term.name = 'kl_term'

    max_x = m_x.max()
    mean_x = m_x.mean()
    min_x = m_x.min()
    max_x.name = 'max_x'
    mean_x.name = 'mean_x'
    min_x.name = 'min_x'

    max_theta_mu = m_theta_mu_in.max()
    mean_theta_mu = m_theta_mu_in.mean()
    min_theta_mu = m_theta_mu_in.min()
    max_theta_mu.name = 'max_theta_mu'
    mean_theta_mu.name = 'mean_theta_mu'
    min_theta_mu.name = 'min_theta_mu'

    max_theta_sig = m_theta_sig_in.max()
    mean_theta_sig = m_theta_sig_in.mean()
    min_theta_sig = m_theta_sig_in.min()
    max_theta_sig.name = 'max_theta_sig'
    mean_theta_sig.name = 'mean_theta_sig'
    min_theta_sig.name = 'min_theta_sig'

    max_phi_sig = m_phi_sig_temp.max()
    mean_phi_sig = m_phi_sig_temp.mean()
    min_phi_sig = m_phi_sig_temp.min()
    max_phi_sig.name = 'max_phi_sig'
    mean_phi_sig.name = 'mean_phi_sig'
    min_phi_sig.name = 'min_phi_sig'

    max_prior_sig = m_prior_sig_temp.max()
    mean_prior_sig = m_prior_sig_temp.mean()
    min_prior_sig = m_prior_sig_temp.min()
    max_prior_sig.name = 'max_prior_sig'
    mean_prior_sig.name = 'mean_prior_sig'
    min_prior_sig.name = 'min_prior_sig'

    model.inputs = [x]
    model.params = params
    model.nodes = nodes
    model.set_updates(shared_updates)

    optimizer = Adam(
        lr=lr
    )

    monitor_fn = theano.function(inputs=[m_x],
                                 outputs=[m_nll_upper_bound, m_recon_term, m_kl_term,
                                          max_phi_sig, mean_phi_sig, min_phi_sig,
                                          max_prior_sig, mean_prior_sig, min_prior_sig,
                                          max_theta_sig, mean_theta_sig, min_theta_sig,
                                          max_x, mean_x, min_x,
                                          max_theta_mu, mean_theta_mu, min_theta_mu],
                                 on_unused_input='ignore')

    extension = [
        GradientClipping(batch_size=batch_size, check_nan=1),
        EpochCount(epoch),
        Monitoring(freq=monitoring_freq,
                   monitor_fn=monitor_fn,
                   ddout=[m_nll_upper_bound, m_recon_term, m_kl_term,
                          max_phi_sig, mean_phi_sig, min_phi_sig,
                          max_prior_sig, mean_prior_sig, min_prior_sig,
                          max_theta_sig, mean_theta_sig, min_theta_sig,
                          max_x, mean_x, min_x,
                          max_theta_mu, mean_theta_mu, min_theta_mu],
                   data=[Iterator(train_data, m_batch_size, start=0, end=112640),
                         Iterator(valid_data, m_batch_size, start=2040064, end=2152704)]),
        Picklize(freq=monitoring_freq, force_save_freq=force_saving_freq, path=save_path),
        EarlyStopping(freq=monitoring_freq, force_save_freq=force_saving_freq, path=save_path, channel=channel_name),
        WeightNorm()
    ]

    mainloop = Training(
        name=pkl_name,
        data=Iterator(train_data, batch_size, start=0, end=2040064),
        model=model,
        optimizer=optimizer,
        cost=nll_upper_bound,
        outputs=[nll_upper_bound],
        extension=extension
    )
    mainloop.run()

#.........这里部分代码省略.........

    max_theta_sig = m_theta_sig_temp.max()
    mean_theta_sig = m_theta_sig_temp.mean()
    min_theta_sig = m_theta_sig_temp.min()
    max_theta_sig.name = "max_theta_sig"
    mean_theta_sig.name = "mean_theta_sig"
    min_theta_sig.name = "min_theta_sig"

    max_phi_sig = m_phi_sig_temp.max()
    mean_phi_sig = m_phi_sig_temp.mean()
    min_phi_sig = m_phi_sig_temp.min()
    max_phi_sig.name = "max_phi_sig"
    mean_phi_sig.name = "mean_phi_sig"
    min_phi_sig.name = "min_phi_sig"

    max_prior_sig = m_prior_sig_temp.max()
    mean_prior_sig = m_prior_sig_temp.mean()
    min_prior_sig = m_prior_sig_temp.min()
    max_prior_sig.name = "max_prior_sig"
    mean_prior_sig.name = "mean_prior_sig"
    min_prior_sig.name = "min_prior_sig"

    model.inputs = [x]
    model.params = params
    model.nodes = nodes
    model.set_updates(shared_updates)

    optimizer = Adam(lr=lr)

    monitor_fn = theano.function(
        inputs=[m_x],
        outputs=[
            m_nll_upper_bound,
            m_recon_term,
            m_kl_term,
            max_phi_sig,
            mean_phi_sig,
            min_phi_sig,
            max_prior_sig,
            mean_prior_sig,
            min_prior_sig,
            max_theta_sig,
            mean_theta_sig,
            min_theta_sig,
            max_x,
            mean_x,
            min_x,
            max_theta_mu,
            mean_theta_mu,
            min_theta_mu,
        ],
        on_unused_input="ignore",
    )

    extension = [
        GradientClipping(batch_size=batch_size, check_nan=1),
        EpochCount(epoch),
        Monitoring(
            freq=monitoring_freq,
            monitor_fn=monitor_fn,
            ddout=[
                m_nll_upper_bound,
                m_recon_term,
                m_kl_term,
                max_phi_sig,
                mean_phi_sig,
                min_phi_sig,
                max_prior_sig,
                mean_prior_sig,
                min_prior_sig,
                max_theta_sig,
                mean_theta_sig,
                min_theta_sig,
                max_x,
                mean_x,
                min_x,
                max_theta_mu,
                mean_theta_mu,
                min_theta_mu,
            ],
            data=[
                Iterator(train_data, m_batch_size, start=0, end=112640),
                Iterator(valid_data, m_batch_size, start=2040064, end=2152704),
            ],
        ),
        Picklize(freq=monitoring_freq, force_save_freq=force_saving_freq, path=save_path),
        EarlyStopping(freq=monitoring_freq, force_save_freq=force_saving_freq, path=save_path, channel=channel_name),
        WeightNorm(),
    ]

    mainloop = Training(
        name=pkl_name,
        data=KIter(train_data, batch_size, start=0, end=2040064),
        model=model,
        optimizer=optimizer,
        cost=nll_upper_bound,
        outputs=[nll_upper_bound],
        extension=extension,
    )
    mainloop.run()