コード例 #1
0
ファイル: train_cae.py プロジェクト: StevenLOL/video_predict
def main(specstr=default_specstr, z_dim=256, num_epochs=10, ch=3, init_from='',
        img_size=64, pxsh=0.5, data_file='', batch_size=8, save_to='params'):

    # build expressions for the output, loss, gradient
    input_var = T.tensor4('inputs')
    print('building specstr {} - zdim {}'.format(specstr, z_dim))
    cae = m.build_cae_nopoolinv(input_var, shape=(img_size,img_size), channels=ch, 
            specstr=specstr.format(z_dim))
    l_list = nn.layers.get_all_layers(cae)
    pred = nn.layers.get_output(cae)
    loss = nn.objectives.squared_error(pred, input_var.flatten(2)).mean()
    params = nn.layers.get_all_params(cae, trainable=True)
    grads = nn.updates.total_norm_constraint(T.grad(loss, params), 10)
    updates = nn.updates.adam(grads, params, learning_rate=1e-3)
    te_pred = nn.layers.get_output(cae, deterministic=True)
    te_loss = nn.objectives.squared_error(te_pred, input_var.flatten(2)).mean()

    # training functions
    print('compiling functions')
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], te_loss)

    # compile functions for encode/decode to test later
    enc_layer = l_list[next(i for i in xrange(len(l_list)) if l_list[i].name=='encode')]
    enc_fn = theano.function([input_var], nn.layers.get_output(enc_layer, deterministic=True))
    dec_fn = lambda z: nn.layers.get_output(cae, deterministic=True,
        inputs={l_list[0]:np.zeros((z.shape[0],ch,img_size,img_size),dtype=theano.config.floatX),
            enc_layer:z}).eval().reshape(-1,ch,img_size,img_size)

    # load params if requested, run training
    if len(init_from) > 0:
        print('loading params from {}'.format(init_from))
        load_params(cae, init_from)
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    print('training for {} epochs'.format(num_epochs))
    hist = u.train_with_hdf5(data, num_epochs=num_epochs,
        train_fn = train_fn,
        test_fn = val_fn,
        tr_transform=lambda x: u.raw_to_floatX(x[0], pixel_shift=pxsh, center=False),
        te_transform=lambda x: u.raw_to_floatX(x[0], pixel_shift=pxsh, center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    tg = u.raw_to_floatX(imb, pixel_shift=pxsh, square=True, center=True)
    pr = dec_fn(enc_fn(tg))
    for i in range(pr.shape[0]):
        u.get_image_pair(tg, pr,index=i,shift=pxsh).save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('cae_train_hist.csv', np.asarray(hist), delimiter=',', fmt='%.5f')
    u.save_params(cae, os.path.join(save_to, 'cae_{}.npz'.format(hist[-1,-1])))
コード例 #2
0
ファイル: train_vcae.py プロジェクト: StevenLOL/video_predict
def main(L=2, img_size=64, pxsh=0., z_dim=32, n_hid=1024, num_epochs=12, binary='True',
        init_from='', data_file='', batch_size=128, save_to='params', max_per_epoch=-1):
    binary = binary.lower()=='true'

    # Create VAE model
    input_var = T.tensor4('inputs')
    print("Building model and compiling functions...")
    print("L = {}, z_dim = {}, n_hid = {}, binary={}".format(L, z_dim, n_hid, binary))
    l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
           m.build_vcae(input_var, L=L, binary=binary, z_dim=z_dim, n_hid=n_hid)
        
    if len(init_from) > 0:
        print('loading from {}'.format(init_from))
        load_params(l_x, init_from)

    # compile functions
    loss, _ = u.build_vae_loss(input_var, *l_tup, deterministic=False, binary=binary, L=L)
    test_loss, test_prediction = u.build_vae_loss(input_var, *l_tup, deterministic=True,
            binary=binary, L=L)
    params = nn.layers.get_all_params(l_x, trainable=True)
    updates = nn.updates.adam(loss, params, learning_rate=3e-5)
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], test_loss)
    ae_fn = theano.function([input_var], test_prediction)

    # run training loop
    print('training for {} epochs'.format(num_epochs))
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    hist = u.train_with_hdf5(data, num_epochs=num_epochs, train_fn=train_fn, test_fn=val_fn,
            max_per_epoch=max_per_epoch,
            tr_transform=lambda x: u.raw_to_floatX(x[0], pixel_shift=pxsh, center=False),
            te_transform=lambda x: u.raw_to_floatX(x[0], pixel_shift=pxsh, center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    orig_images = u.raw_to_floatX(imb, pixel_shift=pxsh)
    autoencoded_images = ae_fn(orig_images)
    for i in range(autoencoded_images.shape[0]):
        u.get_image_pair(orig_images, autoencoded_images, index=i, shift=pxsh) \
                .save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('vcae_train_hist.csv', np.asarray(hist), delimiter=',', fmt='%.5f')
    u.save_params(l_x, os.path.join(save_to, 'vcae_{}.npz'.format(hist[-1,-1])))
コード例 #3
0
ファイル: train_cae.py プロジェクト: zcrwind/video_predict
def main(specstr=default_specstr,
         z_dim=256,
         num_epochs=10,
         ch=3,
         init_from='',
         img_size=64,
         pxsh=0.5,
         data_file='',
         batch_size=8,
         save_to='params'):

    # build expressions for the output, loss, gradient
    input_var = T.tensor4('inputs')
    print('building specstr {} - zdim {}'.format(specstr, z_dim))
    cae = m.build_cae_nopoolinv(input_var,
                                shape=(img_size, img_size),
                                channels=ch,
                                specstr=specstr.format(z_dim))
    l_list = nn.layers.get_all_layers(cae)
    pred = nn.layers.get_output(cae)
    loss = nn.objectives.squared_error(pred, input_var.flatten(2)).mean()
    params = nn.layers.get_all_params(cae, trainable=True)
    grads = nn.updates.total_norm_constraint(T.grad(loss, params), 10)
    updates = nn.updates.adam(grads, params, learning_rate=1e-3)
    te_pred = nn.layers.get_output(cae, deterministic=True)
    te_loss = nn.objectives.squared_error(te_pred, input_var.flatten(2)).mean()

    # training functions
    print('compiling functions')
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], te_loss)

    # compile functions for encode/decode to test later
    enc_layer = l_list[next(i for i in xrange(len(l_list))
                            if l_list[i].name == 'encode')]
    enc_fn = theano.function([input_var],
                             nn.layers.get_output(enc_layer,
                                                  deterministic=True))
    dec_fn = lambda z: nn.layers.get_output(
        cae,
        deterministic=True,
        inputs={
            l_list[0]:
            np.zeros((z.shape[0], ch, img_size, img_size),
                     dtype=theano.config.floatX),
            enc_layer:
            z
        }).eval().reshape(-1, ch, img_size, img_size)

    # load params if requested, run training
    if len(init_from) > 0:
        print('loading params from {}'.format(init_from))
        load_params(cae, init_from)
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    print('training for {} epochs'.format(num_epochs))
    hist = u.train_with_hdf5(data,
                             num_epochs=num_epochs,
                             train_fn=train_fn,
                             test_fn=val_fn,
                             tr_transform=lambda x: u.raw_to_floatX(
                                 x[0], pixel_shift=pxsh, center=False),
                             te_transform=lambda x: u.raw_to_floatX(
                                 x[0], pixel_shift=pxsh, center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    tg = u.raw_to_floatX(imb, pixel_shift=pxsh, square=True, center=True)
    pr = dec_fn(enc_fn(tg))
    for i in range(pr.shape[0]):
        u.get_image_pair(tg, pr, index=i,
                         shift=pxsh).save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('cae_train_hist.csv',
               np.asarray(hist),
               delimiter=',',
               fmt='%.5f')
    u.save_params(cae, os.path.join(save_to, 'cae_{}.npz'.format(hist[-1,
                                                                      -1])))
コード例 #4
0
ファイル: train_vcae.py プロジェクト: zcrwind/video_predict
def main(L=2,
         img_size=64,
         pxsh=0.,
         z_dim=32,
         n_hid=1024,
         num_epochs=12,
         binary='True',
         init_from='',
         data_file='',
         batch_size=128,
         save_to='params',
         max_per_epoch=-1):
    binary = binary.lower() == 'true'

    # Create VAE model
    input_var = T.tensor4('inputs')
    print("Building model and compiling functions...")
    print("L = {}, z_dim = {}, n_hid = {}, binary={}".format(
        L, z_dim, n_hid, binary))
    l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
           m.build_vcae(input_var, L=L, binary=binary, z_dim=z_dim, n_hid=n_hid)

    if len(init_from) > 0:
        print('loading from {}'.format(init_from))
        load_params(l_x, init_from)

    # compile functions
    loss, _ = u.build_vae_loss(input_var,
                               *l_tup,
                               deterministic=False,
                               binary=binary,
                               L=L)
    test_loss, test_prediction = u.build_vae_loss(input_var,
                                                  *l_tup,
                                                  deterministic=True,
                                                  binary=binary,
                                                  L=L)
    params = nn.layers.get_all_params(l_x, trainable=True)
    updates = nn.updates.adam(loss, params, learning_rate=3e-5)
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], test_loss)
    ae_fn = theano.function([input_var], test_prediction)

    # run training loop
    print('training for {} epochs'.format(num_epochs))
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    hist = u.train_with_hdf5(data,
                             num_epochs=num_epochs,
                             train_fn=train_fn,
                             test_fn=val_fn,
                             max_per_epoch=max_per_epoch,
                             tr_transform=lambda x: u.raw_to_floatX(
                                 x[0], pixel_shift=pxsh, center=False),
                             te_transform=lambda x: u.raw_to_floatX(
                                 x[0], pixel_shift=pxsh, center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    orig_images = u.raw_to_floatX(imb, pixel_shift=pxsh)
    autoencoded_images = ae_fn(orig_images)
    for i in range(autoencoded_images.shape[0]):
        u.get_image_pair(orig_images, autoencoded_images, index=i, shift=pxsh) \
                .save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('vcae_train_hist.csv',
               np.asarray(hist),
               delimiter=',',
               fmt='%.5f')
    u.save_params(l_x, os.path.join(save_to, 'vcae_{}.npz'.format(hist[-1,
                                                                       -1])))
コード例 #5
0
def main(data_file = '', num_epochs=10, batch_size = 128, L=2, z_dim=256,
        n_hid=1500, binary='false', img_size = 64, init_from = '', save_to='params',
        split_layer='conv7', pxsh = 0.5, specstr = c.pf_cae_specstr,
        cae_weights=c.pf_cae_params, deconv_weights = c.pf_deconv_params):
    binary = binary.lower() == 'true'

    # pre-trained function for extracting convolutional features from images
    cae = m.build_cae(input_var=None, specstr=specstr, shape=(img_size,img_size))
    laydict = dict((l.name, l) for l in nn.layers.get_all_layers(cae))
    convshape = nn.layers.get_output_shape(laydict[split_layer])
    convs_from_img, _ = m.encoder_decoder(cae_weights, specstr=specstr, layersplit=split_layer,
            shape=(img_size, img_size))
    # pre-trained function for returning to images from convolutional features
    img_from_convs = m.deconvoluter(deconv_weights, specstr=specstr, shape=convshape)

    # Create VAE model
    print("Building model and compiling functions...")
    print("L = {}, z_dim = {}, n_hid = {}, binary={}".format(L, z_dim, n_hid, binary))
    input_var = T.tensor4('inputs')
    c,w,h = convshape[1], convshape[2], convshape[3]
    l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
            m.build_vae(input_var, L=L, binary=binary, z_dim=z_dim, n_hid=n_hid,
                   shape=(w,h), channels=c)

    if len(init_from) > 0:
        print("loading from {}".format(init_from))
        u.load_params(l_x, init_from)
    
    # build loss, updates, training, prediction functions
    loss,_ = u.build_vae_loss(input_var, *l_tup, deterministic=False, binary=binary, L=L)
    test_loss, test_prediction = u.build_vae_loss(input_var, *l_tup, deterministic=True,
            binary=binary, L=L)

    lr = theano.shared(nn.utils.floatX(1e-5))
    params = nn.layers.get_all_params(l_x, trainable=True)
    updates = nn.updates.adam(loss, params, learning_rate=lr)
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], test_loss)
    ae_fn = theano.function([input_var], test_prediction)

    # run training loop
    def data_transform(x, do_center):
        floatx_ims = u.raw_to_floatX(x, pixel_shift=pxsh, square=True, center=do_center)
        return convs_from_img(floatx_ims)

    print("training for {} epochs".format(num_epochs))
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    hist = u.train_with_hdf5(data, num_epochs=num_epochs, train_fn=train_fn, test_fn=val_fn,
                             tr_transform=lambda x: data_transform(x[0], do_center=False),
                             te_transform=lambda x: data_transform(x[0], do_center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    orig_feats = data_transform(imb, do_center=True)
    reconstructed_feats = ae_fn(orig_feats).reshape(orig_feats.shape)
    orig_feats_deconv = img_from_convs(orig_feats)
    reconstructed_feats_deconv = img_from_convs(reconstructed_feats)
    for i in range(reconstructed_feats_deconv.shape[0]):
        u.get_image_pair(orig_feats_deconv, reconstructed_feats_deconv, index=i, shift=pxsh)\
                .save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('vae_convs_train_hist.csv', np.asarray(hist), delimiter=',', fmt='%.5f')
    u.save_params(l_x, os.path.join(save_to, 'vae_convs_{}.npz'.format(hist[-1,-1])))
コード例 #6
0
def main(data_file='',
         num_epochs=10,
         batch_size=128,
         L=2,
         z_dim=256,
         n_hid=1500,
         binary='false',
         img_size=64,
         init_from='',
         save_to='params',
         split_layer='conv7',
         pxsh=0.5,
         specstr=c.pf_cae_specstr,
         cae_weights=c.pf_cae_params,
         deconv_weights=c.pf_deconv_params):
    binary = binary.lower() == 'true'

    # pre-trained function for extracting convolutional features from images
    cae = m.build_cae(input_var=None,
                      specstr=specstr,
                      shape=(img_size, img_size))
    laydict = dict((l.name, l) for l in nn.layers.get_all_layers(cae))
    convshape = nn.layers.get_output_shape(laydict[split_layer])
    convs_from_img, _ = m.encoder_decoder(cae_weights,
                                          specstr=specstr,
                                          layersplit=split_layer,
                                          shape=(img_size, img_size))
    # pre-trained function for returning to images from convolutional features
    img_from_convs = m.deconvoluter(deconv_weights,
                                    specstr=specstr,
                                    shape=convshape)

    # Create VAE model
    print("Building model and compiling functions...")
    print("L = {}, z_dim = {}, n_hid = {}, binary={}".format(
        L, z_dim, n_hid, binary))
    input_var = T.tensor4('inputs')
    c, w, h = convshape[1], convshape[2], convshape[3]
    l_tup = l_z_mu, l_z_ls, l_x_mu_list, l_x_ls_list, l_x_list, l_x = \
            m.build_vae(input_var, L=L, binary=binary, z_dim=z_dim, n_hid=n_hid,
                   shape=(w,h), channels=c)

    if len(init_from) > 0:
        print("loading from {}".format(init_from))
        u.load_params(l_x, init_from)

    # build loss, updates, training, prediction functions
    loss, _ = u.build_vae_loss(input_var,
                               *l_tup,
                               deterministic=False,
                               binary=binary,
                               L=L)
    test_loss, test_prediction = u.build_vae_loss(input_var,
                                                  *l_tup,
                                                  deterministic=True,
                                                  binary=binary,
                                                  L=L)

    lr = theano.shared(nn.utils.floatX(1e-5))
    params = nn.layers.get_all_params(l_x, trainable=True)
    updates = nn.updates.adam(loss, params, learning_rate=lr)
    train_fn = theano.function([input_var], loss, updates=updates)
    val_fn = theano.function([input_var], test_loss)
    ae_fn = theano.function([input_var], test_prediction)

    # run training loop
    def data_transform(x, do_center):
        floatx_ims = u.raw_to_floatX(x,
                                     pixel_shift=pxsh,
                                     square=True,
                                     center=do_center)
        return convs_from_img(floatx_ims)

    print("training for {} epochs".format(num_epochs))
    data = u.DataH5PyStreamer(data_file, batch_size=batch_size)
    hist = u.train_with_hdf5(
        data,
        num_epochs=num_epochs,
        train_fn=train_fn,
        test_fn=val_fn,
        tr_transform=lambda x: data_transform(x[0], do_center=False),
        te_transform=lambda x: data_transform(x[0], do_center=True))

    # generate examples, save training history
    te_stream = data.streamer(shuffled=True)
    imb, = next(te_stream.get_epoch_iterator())
    orig_feats = data_transform(imb, do_center=True)
    reconstructed_feats = ae_fn(orig_feats).reshape(orig_feats.shape)
    orig_feats_deconv = img_from_convs(orig_feats)
    reconstructed_feats_deconv = img_from_convs(reconstructed_feats)
    for i in range(reconstructed_feats_deconv.shape[0]):
        u.get_image_pair(orig_feats_deconv, reconstructed_feats_deconv, index=i, shift=pxsh)\
                .save('output_{}.jpg'.format(i))
    hist = np.asarray(hist)
    np.savetxt('vae_convs_train_hist.csv',
               np.asarray(hist),
               delimiter=',',
               fmt='%.5f')
    u.save_params(
        l_x, os.path.join(save_to, 'vae_convs_{}.npz'.format(hist[-1, -1])))