예제 #1
0
    # Compute prior probabilities per class
    train_y = mnist.binarize_labels(train_y)
    prior_y = train_y.mean(axis=1).reshape((10,1))

    # Create model
    n_x = 28*28
    n_y = 10
    n_z = 50
    n_hidden = 500,500
    updates = None
    model = GPUVAE_YZ_X(updates, n_x, n_y, n_hidden, n_z, n_hidden, 'softplus', 'softplus', type_px='bernoulli', type_qz='gaussianmarg', type_pz='gaussianmarg', prior_sd=1, uniform_y=True)

    # Load parameters
    dir = 'models/mnist_yz_x_50-500-500/'
    ndict.set_value(model.v, ndict.loadz(dir+'v_best.ndict.tar.gz'))
    ndict.set_value(model.w, ndict.loadz(dir+'w_best.ndict.tar.gz'))

else:
    raise Exception("Unknown dataset")

# Make predictions on test set
def get_lowerbound():
    lb = np.zeros((n_y,test_x.shape[1]))
    for _class in range(n_y):
        y = np.zeros((n_y,test_x.shape[1]))
        y[_class,:] = 1
        _lb = model.eval({'x': test_x.astype(np.float32), 'y':y.astype(np.float32)}, {})
        lb[_class,:] = _lb
    return lb
예제 #2
0
    
    n_y = 10
    n_batch_w = 10
    
    colorImg = True
    binarize = False
    
    if True:
        n_hidden = (500,500)
        n_z = 300
        dir = 'models/svhn_yz_x_300-500-500/'
        from anglepy.models import GPUVAE_YZ_X
        model = GPUVAE_YZ_X(None, n_x, n_y, n_hidden, n_z, n_hidden[::-1], nonlinear, nonlinear, type_px, type_qz=type_qz, type_pz=type_pz, prior_sd=100, init_sd=1e-2)
        w = ndict.loadz(dir+'w_best.ndict.tar.gz')
        v = ndict.loadz(dir+'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)

        # PCA
        f_enc, f_dec = pp.PCA_fromfile(dir+'pca_params.ndict.tar.gz')
        
if dataset == 'mnist':
    # MNIST
    import anglepy.data.mnist as mnist
    train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(size=28, binarize_y=True)
    f_enc, f_dec = lambda x:x, lambda x:x
    
    n_x = 28*28
    dim_input = (28,28)
    type_qz = 'gaussianmarg'
    type_pz = 'gaussianmarg'
예제 #3
0
def main(n_z, n_hidden, dataset, seed, gfx=True, _size=None):
    '''Learn a variational auto-encoder with generative model p(x,y,z)=p(y)p(z)p(x|y,z).
    x and y are (always) observed.
    I.e. this cannot be used for semi-supervised learning
    '''
    assert (type(n_hidden) == tuple or type(n_hidden) == list)
    assert type(n_z) == int
    assert isinstance(dataset, basestring)
    
    print 'gpulearn_yz_x', n_z, n_hidden, dataset, seed
    
    comment = ''
    if os.environ.has_key('prior') and bool(int(os.environ['prior'])) == True:
        comment += 'prior-'
    if os.environ.has_key('default') and bool(int(os.environ['default'])) == True:
        comment += 'default-'
    else:
        comment += 'not_default-'
    
    import time
    logdir = 'results/gpulearn_yz_x_'+dataset+'_'+str(n_z)+'-'+str(n_hidden)+comment+'-'+str(int(time.time()))+'/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print 'logdir:', logdir
    
    np.random.seed(seed)
    
    # Init data
    if dataset == 'mnist':
        '''
        What works well:
        100-2-100 (Generated digits stay bit shady)
        1000-2-1000 (Needs pretty long training)
        '''
        import anglepy.data.mnist as mnist
        
        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(size, binarize_y=True)
        f_enc, f_dec = lambda x:x, lambda x:x
        
        if os.environ.has_key('prior') and bool(int(os.environ['prior'])) == True:
            color.printBlue('Loading prior')
            mnist_prior = sio.loadmat('data/mnist_prior/mnist_prior.mat')
            train_mean_prior = mnist_prior['z_train']
            valid_mean_prior = mnist_prior['z_valid']
        else:    
            train_mean_prior = np.zeros((n_z,train_x.shape[1]))
            valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
            
        x = {'x': train_x[:,:].astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': train_y[:,:].astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32),'y': valid_y.astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 2
        type_px = 'bernoulli'
        #print 'Network Structure:', n_z, 

    elif dataset == 'mnist_basic': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_'
        tmp = sio.loadmat(data_dir+'train.mat')
        #color.printRed(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
    elif dataset == 'rectangle': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'rectangles_'
        tmp = sio.loadmat(data_dir+'train.mat')
        color.printRed(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,1000:]
        valid_y = train_y[1000:]
        train_x = train_x[:,:1000]
        train_y = train_y[:1000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 2
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 1000
        n_valid = 200
        n_test = 50000
        n_batch = 500
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
    
    elif dataset == 'convex': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'convex_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,6000:]
        valid_y = train_y[6000:]
        train_x = train_x[:,:6000]
        train_y = train_y[:6000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 2
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 6000
        n_valid = 2000
        n_test = 50000
        n_batch = 120
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'rectangle_image': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'rectangles_im_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 2
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'mnist_rot': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_all_rotation_normalized_float_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'mnist_back_rand': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_background_random_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'mnist_back_image': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_background_images_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'mnist_back_image_rot': 
        # MNIST
        size = 28
        data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_all_background_images_rotation_normalized_'
        tmp = sio.loadmat(data_dir+'train.mat')
        train_x = tmp['x_train'].T
        train_y = tmp['t_train'].T.astype(np.int32)
        # validation 2000
        valid_x = train_x[:,10000:]
        valid_y = train_y[10000:]
        train_x = train_x[:,:10000]
        train_y = train_y[:10000]
        tmp = sio.loadmat(data_dir+'test.mat')
        test_x = tmp['x_test'].T
        test_y = tmp['t_test'].T.astype(np.int32)
        
        print train_x.shape
        print train_y.shape
        print test_x.shape
        print test_y.shape
        
        f_enc, f_dec = pp.Identity()
        train_mean_prior = np.zeros((n_z,train_x.shape[1]))
        test_mean_prior = np.zeros((n_z,test_x.shape[1]))
        valid_mean_prior = np.zeros((n_z,valid_x.shape[1]))
        '''
        x = {'x': train_x.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        '''
        x = {'x': train_x.astype(np.float32), 'mean_prior': train_mean_prior.astype(np.float32), 'y': labelToMat(train_y).astype(np.float32)}
        x_train = x
        x_valid = {'x': valid_x.astype(np.float32), 'mean_prior': valid_mean_prior.astype(np.float32), 'y': labelToMat(valid_y).astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32), 'mean_prior': test_mean_prior.astype(np.float32), 'y': labelToMat(test_y).astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 10000
        n_valid = 2000
        n_test = 50000
        n_batch = 200
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
    else:
        raise Exception('Unknown dataset.')
  



    # Init model
    n_hidden_q = n_hidden
    n_hidden_p = n_hidden
    from anglepy.models import GPUVAE_YZ_X
    if os.environ.has_key('default') and bool(int(os.environ['default'])) == True:
        updates = get_adam_optimizer(alpha=3e-4, beta1=0.9, beta2=0.999, weight_decay=0)
    else:
        updates = get_adam_optimizer(alpha=3e-4, beta1=0.1, beta2=0.001, weight_decay=1000.0/50000.0)
    model = GPUVAE_YZ_X(updates, n_x, n_y, n_hidden_q, n_z, n_hidden_p[::-1], 'softplus', 'softplus', type_px=type_px, type_qz='gaussianmarg', type_pz='gaussianmarg', prior_sd=1, uniform_y=True)
    
    if False:
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414094291/'
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414163488/'
        w = ndict.loadz(dir+'w_best.ndict.tar.gz')
        v = ndict.loadz(dir+'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)
    
    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]

    # Fixed sample for visualisation
    z_sample = {'z': np.repeat(np.random.standard_normal(size=(n_z, 12)), 12, axis=1).astype(np.float32)}
    y_sample = {'y': np.tile(np.random.multinomial(1, [1./n_y]*n_y, size=12).T, (1, 12))}
    
    # Progress hook
    def hook(epoch, t, ll):
        
        if epoch%10 != 0:
            return
        
        ll_valid, _ = model.est_loglik(x_valid, n_samples=L_valid, n_batch=n_batch, byteToFloat=byteToFloat)
            
        if math.isnan(ll_valid):
            print "NaN detected. Reverting to saved best parameters"
            ndict.set_value(model.v, ndict.loadz(logdir+'v.ndict.tar.gz'))
            ndict.set_value(model.w, ndict.loadz(logdir+'w.ndict.tar.gz'))
            return
            
        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir+'v_best')
            ndict.savez(ndict.get_value(model.w), logdir+'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if False and ll_valid_stats[1] > 1000:
                print "Finished"
                with open(logdir+'hook.txt', 'a') as f:
                    print >>f, "Finished"
                exit()

        # Log
        ndict.savez(ndict.get_value(model.v), logdir+'v')
        ndict.savez(ndict.get_value(model.w), logdir+'w')
        print epoch, t, ll, ll_valid
        with open(logdir+'hook.txt', 'a') as f:
            print >>f, t, ll, ll_valid
        
        if gfx:   
            # Graphics
            
            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}
                
            tail = '-'+str(epoch)+'.png'
            
            image = paramgraphics.mat_to_img(f_dec(v['w0x'][:].T), dim_input, True, colorImg=colorImg)
            image.save(logdir+'q_w0x'+tail, 'PNG')
            
            image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]), dim_input, True, colorImg=colorImg)
            image.save(logdir+'out_w'+tail, 'PNG')
            
            _x = {'y': np.random.multinomial(1, [1./n_y]*n_y, size=144).T}
            _, _, _z_confab = model.gen_xz(_x, {}, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples'+tail, 'PNG')
            
            _, _, _z_confab = model.gen_xz(y_sample, z_sample, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples_fixed'+tail, 'PNG')
            
            if n_z == 2:
                
                import ImageFont
                import ImageDraw
                
                n_width = 10
                submosaic_offset = 15
                submosaic_width = (dim_input[1]*n_width)
                submosaic_height = (dim_input[0]*n_width)
                mosaic = Image.new("RGB", (submosaic_width*mosaic_w, submosaic_offset+submosaic_height*mosaic_h))
                
                for digit in range(0,n_y):
                    if digit >= mosaic_h*mosaic_w: continue
                    
                    _x = {}
                    n_batch_plot = n_width*n_width
                    _x['y'] = np.zeros((n_y,n_batch_plot))
                    _x['y'][digit,:] = 1
                    _z = {'z':np.zeros((2,n_width**2))}
                    for i in range(0,n_width):
                        for j in range(0,n_width):
                            _z['z'][0,n_width*i+j] = scipy.stats.norm.ppf(float(i)/n_width+0.5/n_width)
                            _z['z'][1,n_width*i+j] = scipy.stats.norm.ppf(float(j)/n_width+0.5/n_width)
                    
                    _x, _, _z_confab = model.gen_xz(_x, _z, n_batch=n_batch_plot)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg, tile_spacing=(0,0))
                    
                    #image.save(logdir+'samples_digit_'+str(digit)+'_'+tail, 'PNG')
                    mosaic_x = (digit%mosaic_w)*submosaic_width
                    mosaic_y = submosaic_offset+int(digit/mosaic_w)*submosaic_height
                    mosaic.paste(image, (mosaic_x, mosaic_y))
                
                draw = ImageDraw.Draw(mosaic)
                draw.text((1,1),"Epoch #"+str(epoch)+" Loss="+str(int(ll)))
                    
                #plt.savefig(logdir+'mosaic'+tail, format='PNG')
                mosaic.save(logdir+'mosaic'+tail, 'PNG')
                
                #x_samples = _x['x']
                #image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg)
                #image.save(logdir+'samples2'+tail, 'PNG')
        
    # Optimize
    dostep = epoch_vae_adam(model, x, n_batch=n_batch, bernoulli_x=bernoulli_x, byteToFloat=byteToFloat)
    loop_va(dostep, hook)
    
    pass
예제 #4
0
    def hook(epoch, t, ll):
        
        if epoch%10 != 0:
            return
        
        ll_valid, _ = model.est_loglik(x_valid, n_samples=L_valid, n_batch=n_batch, byteToFloat=byteToFloat)
            
        if math.isnan(ll_valid):
            print "NaN detected. Reverting to saved best parameters"
            ndict.set_value(model.v, ndict.loadz(logdir+'v.ndict.tar.gz'))
            ndict.set_value(model.w, ndict.loadz(logdir+'w.ndict.tar.gz'))
            return
            
        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir+'v_best')
            ndict.savez(ndict.get_value(model.w), logdir+'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if False and ll_valid_stats[1] > 1000:
                print "Finished"
                with open(logdir+'hook.txt', 'a') as f:
                    print >>f, "Finished"
                exit()

        # Log
        ndict.savez(ndict.get_value(model.v), logdir+'v')
        ndict.savez(ndict.get_value(model.w), logdir+'w')
        print epoch, t, ll, ll_valid
        with open(logdir+'hook.txt', 'a') as f:
            print >>f, t, ll, ll_valid
        
        if gfx:   
            # Graphics
            
            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}
                
            tail = '-'+str(epoch)+'.png'
            
            image = paramgraphics.mat_to_img(f_dec(v['w0x'][:].T), dim_input, True, colorImg=colorImg)
            image.save(logdir+'q_w0x'+tail, 'PNG')
            
            image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]), dim_input, True, colorImg=colorImg)
            image.save(logdir+'out_w'+tail, 'PNG')
            
            _x = {'y': np.random.multinomial(1, [1./n_y]*n_y, size=144).T}
            _, _, _z_confab = model.gen_xz(_x, {}, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples'+tail, 'PNG')
            
            _, _, _z_confab = model.gen_xz(y_sample, z_sample, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples_fixed'+tail, 'PNG')
            
            if n_z == 2:
                
                import ImageFont
                import ImageDraw
                
                n_width = 10
                submosaic_offset = 15
                submosaic_width = (dim_input[1]*n_width)
                submosaic_height = (dim_input[0]*n_width)
                mosaic = Image.new("RGB", (submosaic_width*mosaic_w, submosaic_offset+submosaic_height*mosaic_h))
                
                for digit in range(0,n_y):
                    if digit >= mosaic_h*mosaic_w: continue
                    
                    _x = {}
                    n_batch_plot = n_width*n_width
                    _x['y'] = np.zeros((n_y,n_batch_plot))
                    _x['y'][digit,:] = 1
                    _z = {'z':np.zeros((2,n_width**2))}
                    for i in range(0,n_width):
                        for j in range(0,n_width):
                            _z['z'][0,n_width*i+j] = scipy.stats.norm.ppf(float(i)/n_width+0.5/n_width)
                            _z['z'][1,n_width*i+j] = scipy.stats.norm.ppf(float(j)/n_width+0.5/n_width)
                    
                    _x, _, _z_confab = model.gen_xz(_x, _z, n_batch=n_batch_plot)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg, tile_spacing=(0,0))
                    
                    #image.save(logdir+'samples_digit_'+str(digit)+'_'+tail, 'PNG')
                    mosaic_x = (digit%mosaic_w)*submosaic_width
                    mosaic_y = submosaic_offset+int(digit/mosaic_w)*submosaic_height
                    mosaic.paste(image, (mosaic_x, mosaic_y))
                
                draw = ImageDraw.Draw(mosaic)
                draw.text((1,1),"Epoch #"+str(epoch)+" Loss="+str(int(ll)))
                    
                #plt.savefig(logdir+'mosaic'+tail, format='PNG')
                mosaic.save(logdir+'mosaic'+tail, 'PNG')
예제 #5
0
def main(n_z, n_hidden, dataset, seed, gfx=True, _size=None):
    '''Learn a variational auto-encoder with generative model p(x,y,z)=p(y)p(z)p(x|y,z).
    x and y are (always) observed.
    I.e. this cannot be used for semi-supervised learning
    '''
    assert (type(n_hidden) == tuple or type(n_hidden) == list)
    assert type(n_z) == int
    assert isinstance(dataset, basestring)
    
    print 'gpulearn_yz_x', n_z, n_hidden, dataset, seed
    
    import time
    logdir = 'results/gpulearn_yz_x_'+dataset+'_'+str(n_z)+'-'+str(n_hidden)+'-'+str(int(time.time()))+'/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print 'logdir:', logdir
    
    np.random.seed(seed)
    
    # Init data
    if dataset == 'mnist':
        '''
        What works well:
        100-2-100 (Generated digits stay bit shady)
        1000-2-1000 (Needs pretty long training)
        '''
        import anglepy.data.mnist as mnist
        
        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(size, binarize_y=True)
        f_enc, f_dec = lambda x:x, lambda x:x
        x = {'x': train_x[:,:].astype(np.float32), 'y': train_y[:,:].astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32), 'y': valid_y.astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        n_y = 10
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 2
        type_px = 'bernoulli'

    elif dataset == 'norb':
        # resized NORB dataset, reshuffled
        import anglepy.data.norb as norb
        size = _size #48
        train_x, train_y, test_x, test_y = norb.load_resized(size, binarize_y=True)
        _x = {'x': train_x, 'y': train_y}
        ndict.shuffleCols(_x)
        train_x = _x['x']
        train_y = _x['y']
        
        # Do PCA
        f_enc, f_dec, pca_params = pp.PCA(_x['x'][:,:10000], cutoff=2000, toFloat=False)
        ndict.savez(pca_params, logdir+'pca_params')
        
        x = {'x': f_enc(train_x).astype(np.float32), 'y':train_y.astype(np.float32)}
        x_valid = {'x': f_enc(test_x).astype(np.float32), 'y':test_y.astype(np.float32)}
        
        L_valid = 1
        n_x = x['x'].shape[0]
        n_y = 5
        dim_input = (size,size)
        n_batch = 1000 #23400/900 = 27
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 1
        type_px = 'gaussian'

    elif dataset == 'norb_instances': 
        # resized NORB dataset with the instances as classes
        import anglepy.data.norb2 as norb2
        size = _size #48
        x, y = norb2.load_numpy_subclasses(size, binarize_y=True)
        _x = {'x': x, 'y': y}
        ndict.shuffleCols(_x)
        
        # Do pre=processing
        if True:
            # Works
            f_enc, f_dec, pca_params = pp.PCA(_x['x'][:,:10000], cutoff=600, global_sd=True, toFloat=True)
            ndict.savez(pca_params, logdir+'pca_params')
        elif False:
            # Doesn't work
            f_enc, f_dec, pp_params = pp.normalize_noise(_x['x'][:,:50000], noise_sd=0.01, global_sd=True, toFloat=True)
        else:
            # Doesn't work
            f_enc, f_dec, params = pp.normalize_random(x=x[:,:10000], global_sd=True, toFloat=True)
            ndict.savez(params, logdir+'normalize_random_params')
        
        n_valid = 5000
        x = {'x': f_enc(_x['x'][:,:-n_valid]).astype(np.float32), 'y':_x['y'][:,:-n_valid].astype(np.float32)}
        x_valid = {'x': f_enc(_x['x'][:,:n_valid]).astype(np.float32), 'y':_x['y'][:,:n_valid].astype(np.float32)}
        
        L_valid = 1
        n_x = x['x'].shape[0]
        n_y = 50
        dim_input = (size,size)
        n_batch = 5000 #23400/900 = 27
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 1
        type_px = 'gaussian'

    elif dataset == 'svhn':    
        # SVHN dataset
        import anglepy.data.svhn as svhn
        size = 32
        train_x, train_y, test_x, test_y = svhn.load_numpy(False, binarize_y=True) #norb.load_resized(size, binarize_y=True)
        extra_x, extra_y = svhn.load_numpy_extra(False, binarize_y=True)
        x = {'x': np.hstack((train_x, extra_x)), 'y':np.hstack((train_y, extra_y))}
        ndict.shuffleCols(x)
        
        #f_enc, f_dec, (x_sd, x_mean) = pp.preprocess_normalize01(train_x, True)
        f_enc, f_dec, pca_params = pp.PCA(x['x'][:,:10000], cutoff=1000, toFloat=True)
        ndict.savez(pca_params, logdir+'pca_params')
        
        n_y = 10
        x = {'x': f_enc(x['x']).astype(np.float32), 'y': x['y'].astype(np.float32)}
        x_valid = {'x': f_enc(test_x).astype(np.float32), 'y': test_y.astype(np.float32)}
        L_valid = 1
        n_x = x['x'].shape[0]
        dim_input = (size,size)
        n_batch = 5000
        colorImg = True
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 2
        type_px = 'gaussian'
        
    # Init model
    n_hidden_q = n_hidden
    n_hidden_p = n_hidden
    from anglepy.models import GPUVAE_YZ_X
    updates = get_adam_optimizer(alpha=3e-4, beta1=0.9, beta2=0.999, weight_decay=0)
    model = GPUVAE_YZ_X(updates, n_x, n_y, n_hidden_q, n_z, n_hidden_p[::-1], 'softplus', 'softplus', type_px=type_px, type_qz='gaussianmarg', type_pz='gaussianmarg', prior_sd=1, uniform_y=True)
    
    if False:
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414094291/'
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414163488/'
        w = ndict.loadz(dir+'w_best.ndict.tar.gz')
        v = ndict.loadz(dir+'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)
    
    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]

    # Fixed sample for visualisation
    z_sample = {'z': np.repeat(np.random.standard_normal(size=(n_z, 12)), 12, axis=1).astype(np.float32)}
    y_sample = {'y': np.tile(np.random.multinomial(1, [1./n_y]*n_y, size=12).T, (1, 12))}
    
    # Progress hook
    def hook(epoch, t, ll):
        
        if epoch%10 != 0:
            return
        
        ll_valid, _ = model.est_loglik(x_valid, n_samples=L_valid, n_batch=n_batch, byteToFloat=byteToFloat)
            
        if math.isnan(ll_valid):
            print "NaN detected. Reverting to saved best parameters"
            ndict.set_value(model.v, ndict.loadz(logdir+'v.ndict.tar.gz'))
            ndict.set_value(model.w, ndict.loadz(logdir+'w.ndict.tar.gz'))
            return
            
        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir+'v_best')
            ndict.savez(ndict.get_value(model.w), logdir+'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if False and ll_valid_stats[1] > 1000:
                print "Finished"
                with open(logdir+'hook.txt', 'a') as f:
                    print >>f, "Finished"
                exit()

        # Log
        ndict.savez(ndict.get_value(model.v), logdir+'v')
        ndict.savez(ndict.get_value(model.w), logdir+'w')
        print epoch, t, ll, ll_valid
        with open(logdir+'hook.txt', 'a') as f:
            print >>f, t, ll, ll_valid
        
        if gfx:   
            # Graphics
            
            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}
                
            tail = '-'+str(epoch)+'.png'
            
            image = paramgraphics.mat_to_img(f_dec(v['w0x'][:].T), dim_input, True, colorImg=colorImg)
            image.save(logdir+'q_w0x'+tail, 'PNG')
            
            image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]), dim_input, True, colorImg=colorImg)
            image.save(logdir+'out_w'+tail, 'PNG')
            
            _x = {'y': np.random.multinomial(1, [1./n_y]*n_y, size=144).T}
            _, _, _z_confab = model.gen_xz(_x, {}, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples'+tail, 'PNG')
            
            _, _, _z_confab = model.gen_xz(y_sample, z_sample, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']), dim_input, colorImg=colorImg)
            image.save(logdir+'samples_fixed'+tail, 'PNG')
            
            if n_z == 2:
                
                import ImageFont
                import ImageDraw
                
                n_width = 10
                submosaic_offset = 15
                submosaic_width = (dim_input[1]*n_width)
                submosaic_height = (dim_input[0]*n_width)
                mosaic = Image.new("RGB", (submosaic_width*mosaic_w, submosaic_offset+submosaic_height*mosaic_h))
                
                for digit in range(0,n_y):
                    if digit >= mosaic_h*mosaic_w: continue
                    
                    _x = {}
                    n_batch_plot = n_width*n_width
                    _x['y'] = np.zeros((n_y,n_batch_plot))
                    _x['y'][digit,:] = 1
                    _z = {'z':np.zeros((2,n_width**2))}
                    for i in range(0,n_width):
                        for j in range(0,n_width):
                            _z['z'][0,n_width*i+j] = scipy.stats.norm.ppf(float(i)/n_width+0.5/n_width)
                            _z['z'][1,n_width*i+j] = scipy.stats.norm.ppf(float(j)/n_width+0.5/n_width)
                    
                    _x, _, _z_confab = model.gen_xz(_x, _z, n_batch=n_batch_plot)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg, tile_spacing=(0,0))
                    
                    #image.save(logdir+'samples_digit_'+str(digit)+'_'+tail, 'PNG')
                    mosaic_x = (digit%mosaic_w)*submosaic_width
                    mosaic_y = submosaic_offset+int(digit/mosaic_w)*submosaic_height
                    mosaic.paste(image, (mosaic_x, mosaic_y))
                
                draw = ImageDraw.Draw(mosaic)
                draw.text((1,1),"Epoch #"+str(epoch)+" Loss="+str(int(ll)))
                    
                #plt.savefig(logdir+'mosaic'+tail, format='PNG')
                mosaic.save(logdir+'mosaic'+tail, 'PNG')
                
                #x_samples = _x['x']
                #image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg)
                #image.save(logdir+'samples2'+tail, 'PNG')
        
    # Optimize
    dostep = epoch_vae_adam(model, x, n_batch=n_batch, bernoulli_x=bernoulli_x, byteToFloat=byteToFloat)
    loop_va(dostep, hook)
    
    pass
예제 #6
0
파일: run_sl.py 프로젝트: codeaudit/mmdgm
        
        
    # Compute prior probabilities per class
    train_y = mnist.binarize_labels(train_y)
    prior_y = train_y.mean(axis=1).reshape((10,1))

    # Create model
    n_x = 28*28
    n_y = 10
    n_hidden = 500,500
    updates = None
    print 'n_z:', n_z
    model = GPUVAE_YZ_X(updates, n_x, n_y, n_hidden, n_z, n_hidden, 'softplus', 'softplus', type_px='bernoulli', type_qz='gaussianmarg', type_pz='gaussianmarg', prior_sd=1, uniform_y=True)

    # Load parameters
    ndict.set_value(model.v, ndict.loadz(dir+'v_best.ndict.tar.gz'))
    ndict.set_value(model.w, ndict.loadz(dir+'w_best.ndict.tar.gz'))

elif dataset == 'mnist_basic':
    data_dir = os.environ['ML_DATA_PATH']+'/mnist_variations/'+'mnist_'
    
    tmp = sio.loadmat(data_dir+'train.mat')
    #color.printRed(data_dir+'train.mat')
    train_x = tmp['x_train'].T
    train_y = tmp['t_train'].T.astype(np.int32)
    # validation 2000
    valid_x = train_x[:,10000:]
    valid_y = train_y[10000:]
    train_x = train_x[:,:10000]
    train_y = train_y[:10000]
    tmp = sio.loadmat(data_dir+'test.mat')
예제 #7
0
def main(n_z, n_hidden, dataset, seed, comment, gfx=True):

    # Initialize logdir
    import time
    logdir = 'results/gpulearn_z_x_' + dataset + '_' + str(n_z) + '-' + str(
        n_hidden) + '_' + comment + '_' + str(int(time.time())) + '/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print 'logdir:', logdir
    print 'gpulearn_z_x', n_z, n_hidden, dataset, seed
    with open(logdir + 'hook.txt', 'a') as f:
        print >> f, 'learn_z_x', n_z, n_hidden, dataset, seed

    np.random.seed(seed)

    gfx_freq = 1

    weight_decay = 0
    f_enc, f_dec = lambda x: x, lambda x: x

    # Init data
    if dataset == 'mnist':
        import anglepy.data.mnist as mnist

        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(
            size)
        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (size, size)
        n_x = size * size
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 50000
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    if dataset == 'mnist_binarized':
        import anglepy.data.mnist_binarized as mnist_binarized
        # MNIST
        train_x, valid_x, test_x = mnist_binarized.load_numpy(28)
        x = {'x': np.hstack((train_x, valid_x)).astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (28, 28)
        n_x = 28 * 28
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'mog'
        nonlinear = 'rectlin'
        type_px = 'bernoulli'
        n_train = 60000
        n_batch = 1000
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    elif dataset == 'freyface':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy()
        np.random.shuffle(train_x)
        x = {'x': train_x.T[:, 0:n_train]}
        x_valid = {'x': train_x.T[:, n_train:]}
        L_valid = 1
        dim_input = (28, 20)
        n_x = 20 * 28
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'bounded01'
        nonlinear = 'tanh'  #tanh works better with freyface #'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    elif dataset == 'freyface_pca':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)

        f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        train_x = f_enc(train_x)

        x = {'x': train_x[:, 0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:, n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28, 20)
        n_x = train_x.shape[0]
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'freyface_bernoulli':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)

        x = {'x': train_x[:, 0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:, n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28, 20)
        n_x = train_x.shape[0]
        type_pz = 'gaussianmarg'
        type_px = 'bernoulli'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'norb':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        byteToFloat = False
        bernoulli_x = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'norb_pca':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        f_enc, f_dec, _ = pp.PCA(train_x, 0.999)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'norb_normalized':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        #f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        f_enc, f_dec, _ = pp.normalize(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'svhn':
        # SVHN dataset
        import anglepy.data.svhn as svhn
        size = 32
        train_x, train_y, test_x, test_y = svhn.load_numpy(
            False, binarize_y=True)  #norb.load_resized(size, binarize_y=True)
        extra_x, extra_y = svhn.load_numpy_extra(False, binarize_y=True)
        x = {
            'x': np.hstack((train_x, extra_x)),
            'y': np.hstack((train_y, extra_y))
        }
        ndict.shuffleCols(x)

        print 'Performing PCA, can take a few minutes... ',
        f_enc, f_dec, pca_params = pp.PCA(x['x'][:, :10000],
                                          cutoff=600,
                                          toFloat=True)
        ndict.savez(pca_params, logdir + 'pca_params')
        print 'Done.'

        n_y = 10
        x = {'x': f_enc(x['x']).astype(np.float32)}
        x_valid = {'x': f_enc(test_x).astype(np.float32)}
        L_valid = 1
        n_x = x['x'].shape[0]
        dim_input = (size, size)
        n_batch = 5000
        colorImg = True
        bernoulli_x = False
        byteToFloat = False
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'

    # Construct model
    from anglepy.models import GPUVAE_Z_X
    updates = get_adam_optimizer(learning_rate=3e-4, weight_decay=weight_decay)
    model = GPUVAE_Z_X(updates,
                       n_x,
                       n_hidden,
                       n_z,
                       n_hidden[::-1],
                       nonlinear,
                       nonlinear,
                       type_px,
                       type_qz=type_qz,
                       type_pz=type_pz,
                       prior_sd=100,
                       init_sd=1e-3)

    if False:
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412689061/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412676966/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412695481/'
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412695455/'
        #dir = '/Users/dpkingma/results/gpulearn_z_x_svhn_pca_300-(500, 500)__1413904756/'
        dir = '/home/ubuntu/results/gpulearn_z_x_mnist_50-[500, 500]__1414259423/'
        w = ndict.loadz(dir + 'w_best.ndict.tar.gz')
        v = ndict.loadz(dir + 'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)

    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]

    # Progress hook
    def hook(epoch, t, ll):

        if epoch % 10 != 0: return

        ll_valid, _ = model.est_loglik(x_valid,
                                       n_samples=L_valid,
                                       n_batch=n_batch,
                                       byteToFloat=byteToFloat)

        # Log
        ndict.savez(ndict.get_value(model.v), logdir + 'v')
        ndict.savez(ndict.get_value(model.w), logdir + 'w')

        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir + 'v_best')
            ndict.savez(ndict.get_value(model.w), logdir + 'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if ll_valid_stats[1] > 1000:
                print "Finished"
                with open(logdir + 'hook.txt', 'a') as f:
                    print >> f, "Finished"
                exit()

        print epoch, t, ll, ll_valid, ll_valid_stats
        with open(logdir + 'hook.txt', 'a') as f:
            print >> f, epoch, t, ll, ll_valid, ll_valid_stats

        # Graphics
        if gfx and epoch % gfx_freq == 0:

            #tail = '.png'
            tail = '-' + str(epoch) + '.png'

            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}

            if 'pca' not in dataset and 'random' not in dataset and 'normalized' not in dataset:

                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'q_w0' + tail, 'PNG')

                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]),
                                                 dim_input,
                                                 True,
                                                 colorImg=colorImg)
                image.save(logdir + 'out_w' + tail, 'PNG')

                if 'out_unif' in w:
                    image = paramgraphics.mat_to_img(f_dec(
                        w['out_unif'].reshape((-1, 1))),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'out_unif' + tail, 'PNG')

                if n_z == 2:
                    n_width = 10
                    import scipy.stats
                    z = {'z': np.zeros((2, n_width**2))}
                    for i in range(0, n_width):
                        for j in range(0, n_width):
                            z['z'][0, n_width * i + j] = scipy.stats.norm.ppf(
                                float(i) / n_width + 0.5 / n_width)
                            z['z'][1, n_width * i + j] = scipy.stats.norm.ppf(
                                float(j) / n_width + 0.5 / n_width)

                    x, _, _z = model.gen_xz({}, z, n_width**2)
                    if dataset == 'mnist':
                        x = 1 - _z['x']
                    image = paramgraphics.mat_to_img(f_dec(_z['x']), dim_input)
                    image.save(logdir + '2dmanifold' + tail, 'PNG')
                else:
                    _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples),
                                                     dim_input,
                                                     colorImg=colorImg)
                    image.save(logdir + 'samples' + tail, 'PNG')

                    #x_samples = _x['x']
                    #image = paramgraphics.mat_to_img(x_samples, dim_input, colorImg=colorImg)
                    #image.save(logdir+'samples2'+tail, 'PNG')

            else:
                # Model with preprocessing

                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'q_w0' + tail, 'PNG')

                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]),
                                                 dim_input,
                                                 True,
                                                 colorImg=colorImg)
                image.save(logdir + 'out_w' + tail, 'PNG')

                _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                x_samples = f_dec(_z_confab['x'])
                x_samples = np.minimum(np.maximum(x_samples, 0), 1)
                image = paramgraphics.mat_to_img(x_samples,
                                                 dim_input,
                                                 colorImg=colorImg)
                image.save(logdir + 'samples' + tail, 'PNG')

    # Optimize
    #SFO
    dostep = epoch_vae_adam(model,
                            x,
                            n_batch=n_batch,
                            bernoulli_x=bernoulli_x,
                            byteToFloat=byteToFloat)
    loop_va(dostep, hook)

    pass
예제 #8
0
def main(n_z, n_hidden, dataset, seed, comment, gfx=True):
    
    # Initialize logdir
    import time
    logdir = 'results/gpulearn_z_x_'+dataset+'_'+str(n_z)+'-'+str(n_hidden)+'_'+comment+'_'+str(int(time.time()))+'/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print 'logdir:', logdir
    print 'gpulearn_z_x', n_z, n_hidden, dataset, seed
    with open(logdir+'hook.txt', 'a') as f:
        print >>f, 'learn_z_x', n_z, n_hidden, dataset, seed
    
    np.random.seed(seed)

    gfx_freq = 1
    
    weight_decay = 0
    f_enc, f_dec = lambda x:x, lambda x:x

    # Init data
    if dataset == 'mnist':
        import anglepy.data.mnist as mnist
        
        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(size)
        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (size,size)
        n_x = size*size
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 50000
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    if dataset == 'mnist_binarized':
        import anglepy.data.mnist_binarized as mnist_binarized
        # MNIST
        train_x, valid_x, test_x = mnist_binarized.load_numpy(28)
        x = {'x': np.hstack((train_x, valid_x)).astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (28,28)
        n_x = 28*28
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'mog'
        nonlinear = 'rectlin'
        type_px = 'bernoulli'
        n_train = 60000
        n_batch = 1000
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch)/n_train
        
    elif dataset == 'freyface':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy()
        np.random.shuffle(train_x)
        x = {'x': train_x.T[:,0:n_train]}
        x_valid = {'x': train_x.T[:,n_train:]}
        L_valid = 1
        dim_input = (28,20)
        n_x = 20*28
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'bounded01'
        nonlinear = 'tanh'  #tanh works better with freyface #'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch)/n_train

    elif dataset == 'freyface_pca':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)
        
        f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        train_x = f_enc(train_x)
        
        x = {'x': train_x[:,0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:,n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28,20)
        n_x = train_x.shape[0]
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'freyface_bernoulli':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)
        
        x = {'x': train_x[:,0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:,n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28,20)
        n_x = train_x.shape[0]
        type_pz = 'gaussianmarg'
        type_px = 'bernoulli'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'norb':    
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size, binarize_y=True)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size,size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900 #23400/900 = 27
        colorImg = False
        #binarize = False
        byteToFloat = False
        bernoulli_x = False
        weight_decay= float(n_batch)/train_x.shape[1]
    
    elif dataset == 'norb_pca':    
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size, binarize_y=True)

        f_enc, f_dec, _ = pp.PCA(train_x, 0.999)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)
        
        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size,size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900 #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay= float(n_batch)/train_x.shape[1]

    elif dataset == 'norb_normalized':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size, binarize_y=True)

        #f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        f_enc, f_dec, _ = pp.normalize(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)
        
        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size,size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900 #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay= float(n_batch)/train_x.shape[1]
        
    elif dataset == 'svhn':
        # SVHN dataset
        import anglepy.data.svhn as svhn
        size = 32
        train_x, train_y, test_x, test_y = svhn.load_numpy(False, binarize_y=True) #norb.load_resized(size, binarize_y=True)
        extra_x, extra_y = svhn.load_numpy_extra(False, binarize_y=True)
        x = {'x': np.hstack((train_x, extra_x)), 'y':np.hstack((train_y, extra_y))}
        ndict.shuffleCols(x)
        
        print 'Performing PCA, can take a few minutes... ',
        f_enc, f_dec, pca_params = pp.PCA(x['x'][:,:10000], cutoff=600, toFloat=True)
        ndict.savez(pca_params, logdir+'pca_params')
        print 'Done.'
        
        n_y = 10
        x = {'x': f_enc(x['x']).astype(np.float32)}
        x_valid = {'x': f_enc(test_x).astype(np.float32)}
        L_valid = 1
        n_x = x['x'].shape[0]
        dim_input = (size,size)
        n_batch = 5000
        colorImg = True
        bernoulli_x = False
        byteToFloat = False
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
    
        
    # Construct model
    from anglepy.models import GPUVAE_Z_X
    updates = get_adam_optimizer(learning_rate=3e-4, weight_decay=weight_decay)
    model = GPUVAE_Z_X(updates, n_x, n_hidden, n_z, n_hidden[::-1], nonlinear, nonlinear, type_px, type_qz=type_qz, type_pz=type_pz, prior_sd=100, init_sd=1e-3)
    
    if False:
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412689061/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412676966/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412695481/'
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412695455/'
        #dir = '/Users/dpkingma/results/gpulearn_z_x_svhn_pca_300-(500, 500)__1413904756/'
        dir = '/home/ubuntu/results/gpulearn_z_x_mnist_50-[500, 500]__1414259423/'
        w = ndict.loadz(dir+'w_best.ndict.tar.gz')
        v = ndict.loadz(dir+'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)
    
    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]
    
    # Progress hook
    def hook(epoch, t, ll):
        
        if epoch%10 != 0: return
        
        ll_valid, _ = model.est_loglik(x_valid, n_samples=L_valid, n_batch=n_batch, byteToFloat=byteToFloat)
        
        # Log
        ndict.savez(ndict.get_value(model.v), logdir+'v')
        ndict.savez(ndict.get_value(model.w), logdir+'w')
        
        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir+'v_best')
            ndict.savez(ndict.get_value(model.w), logdir+'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if ll_valid_stats[1] > 1000:
                print "Finished"
                with open(logdir+'hook.txt', 'a') as f:
                    print >>f, "Finished"
                exit()
        
        print epoch, t, ll, ll_valid, ll_valid_stats
        with open(logdir+'hook.txt', 'a') as f:
            print >>f, epoch, t, ll, ll_valid, ll_valid_stats

        # Graphics
        if gfx and epoch%gfx_freq == 0:
            
            #tail = '.png'
            tail = '-'+str(epoch)+'.png'
            
            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}
                
            if 'pca' not in dataset and 'random' not in dataset and 'normalized' not in dataset:
                
                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T), dim_input, True, colorImg=colorImg)
                    image.save(logdir+'q_w0'+tail, 'PNG')
                
                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]), dim_input, True, colorImg=colorImg)
                image.save(logdir+'out_w'+tail, 'PNG')
                
                if 'out_unif' in w:
                    image = paramgraphics.mat_to_img(f_dec(w['out_unif'].reshape((-1,1))), dim_input, True, colorImg=colorImg)
                    image.save(logdir+'out_unif'+tail, 'PNG')
                
                if n_z == 2:
                    n_width = 10
                    import scipy.stats
                    z = {'z':np.zeros((2,n_width**2))}
                    for i in range(0,n_width):
                        for j in range(0,n_width):
                            z['z'][0,n_width*i+j] = scipy.stats.norm.ppf(float(i)/n_width+0.5/n_width)
                            z['z'][1,n_width*i+j] = scipy.stats.norm.ppf(float(j)/n_width+0.5/n_width)
                    
                    x, _, _z = model.gen_xz({}, z, n_width**2)
                    if dataset == 'mnist':
                        x = 1 - _z['x']
                    image = paramgraphics.mat_to_img(f_dec(_z['x']), dim_input)
                    image.save(logdir+'2dmanifold'+tail, 'PNG')
                else:
                    _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg)
                    image.save(logdir+'samples'+tail, 'PNG')
                    
                    #x_samples = _x['x']
                    #image = paramgraphics.mat_to_img(x_samples, dim_input, colorImg=colorImg)
                    #image.save(logdir+'samples2'+tail, 'PNG')
                    
            else:
                # Model with preprocessing
                
                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T), dim_input, True, colorImg=colorImg)
                    image.save(logdir+'q_w0'+tail, 'PNG')
                    
                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]), dim_input, True, colorImg=colorImg)
                image.save(logdir+'out_w'+tail, 'PNG')

                _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                x_samples = f_dec(_z_confab['x'])
                x_samples = np.minimum(np.maximum(x_samples, 0), 1)
                image = paramgraphics.mat_to_img(x_samples, dim_input, colorImg=colorImg)
                image.save(logdir+'samples'+tail, 'PNG')
                
                
                
    # Optimize
    #SFO
    dostep = epoch_vae_adam(model, x, n_batch=n_batch, bernoulli_x=bernoulli_x, byteToFloat=byteToFloat)
    loop_va(dostep, hook)
    
    pass
예제 #9
0
    
    if True:
        if False:
            n_hidden = (500,500)
            n_z = 300
            dir = 'models/svhn_yz_x_300-500-500/'
        else:
            n_hidden = (1000,1000)
            n_z = 300
            dir = 'models/svhn_yz_x_300-1000-1000/'
        
        from anglepy.models import GPUVAE_YZ_X
        model = GPUVAE_YZ_X(None, n_x, n_y, n_hidden, n_z, n_hidden[::-1], nonlinear, nonlinear, type_px, type_qz=type_qz, type_pz=type_pz, prior_sd=100, init_sd=1e-2)
        w = ndict.loadz(dir+'w_best.ndict.tar.gz')
        v = ndict.loadz(dir+'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)
        # PCA
        pca = ndict.loadz(dir+'pca_params.ndict.tar.gz')
        def f_dec(x):
            result = pca['eigvec'].dot(x * np.sqrt(pca['eigval'])) * pca['x_sd'] + pca['x_center']
            result = np.maximum(0, np.minimum(1, result))
            return result

if dataset == 'mnist':
    n_x = 28*28
    dim_input = (28,28)
    type_qz = 'gaussianmarg'
    type_pz = 'gaussianmarg'
    type_px = 'bernoulli'
    nonlinear = 'softplus'
예제 #10
0
def main(n_z, n_hidden, dataset, seed, comment, gfx=True):

    # Initialize logdir
    #---------------------
    # Setasouto:
    # Create the directory to save the outputs files and log.
    #---------------------
    import time
    logdir = 'results/gpulearn_z_x_' + dataset + '_' + str(n_z) + '-' + str(
        n_hidden) + '_' + comment + '_' + str(int(time.time())) + '/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print('logdir:', logdir)
    print('gpulearn_z_x', n_z, n_hidden, dataset, seed)
    with open(logdir + 'hook.txt', 'a') as f:
        print(f, 'learn_z_x', n_z, n_hidden, dataset, seed)

    np.random.seed(seed)

    gfx_freq = 1

    weight_decay = 0
    f_enc, f_dec = lambda x: x, lambda x: x

    # Init data
    if dataset == 'mnist':
        import anglepy.data.mnist as mnist

        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(
            size)
        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (size, size)
        n_x = size * size
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = 50000
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    if dataset == 'mnist_binarized':
        import anglepy.data.mnist_binarized as mnist_binarized
        # MNIST
        train_x, valid_x, test_x = mnist_binarized.load_numpy(28)
        x = {'x': np.hstack((train_x, valid_x)).astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        dim_input = (28, 28)
        n_x = 28 * 28
        n_y = 10
        type_qz = 'gaussianmarg'
        type_pz = 'mog'
        nonlinear = 'rectlin'
        type_px = 'bernoulli'
        n_train = 60000
        n_batch = 1000
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    elif dataset == 'freyface':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy()
        np.random.shuffle(train_x)
        x = {'x': train_x.T[:, 0:n_train]}
        x_valid = {'x': train_x.T[:, n_train:]}
        L_valid = 1
        dim_input = (28, 20)
        n_x = 20 * 28
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'bounded01'
        nonlinear = 'tanh'  #tanh works better with freyface #'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / n_train

    elif dataset == 'freyface_pca':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)

        f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        train_x = f_enc(train_x)

        x = {'x': train_x[:, 0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:, n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28, 20)
        n_x = train_x.shape[0]
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'freyface_bernoulli':
        # Frey's face
        import anglepy.data.freyface as freyface
        n_train = 1600
        train_x = freyface.load_numpy().T
        np.random.shuffle(train_x.T)

        x = {'x': train_x[:, 0:n_train].astype(np.float32)}
        x_valid = {'x': train_x[:, n_train:].astype(np.float32)}
        L_valid = 1
        dim_input = (28, 20)
        n_x = train_x.shape[0]
        type_pz = 'gaussianmarg'
        type_px = 'bernoulli'
        nonlinear = 'softplus'
        n_batch = 100
        colorImg = False
        bernoulli_x = False
        byteToFloat = False

    elif dataset == 'norb':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        byteToFloat = False
        bernoulli_x = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'norb_pca':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        f_enc, f_dec, _ = pp.PCA(train_x, 0.999)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'norb_normalized':
        # small NORB dataset
        import anglepy.data.norb as norb
        size = 48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)

        #f_enc, f_dec, _ = pp.PCA(train_x, 0.99)
        #f_enc, f_dec, _ = pp.normalize_random(train_x)
        f_enc, f_dec, _ = pp.normalize(train_x)
        train_x = f_enc(train_x)
        test_x = f_enc(test_x)

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': test_x.astype(np.float32)}
        L_valid = 1
        n_x = train_x.shape[0]
        dim_input = (size, size)
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'
        n_batch = 900  #23400/900 = 27
        colorImg = False
        #binarize = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / train_x.shape[1]

    elif dataset == 'svhn':
        # SVHN dataset
        import anglepy.data.svhn as svhn
        size = 32
        train_x, train_y, test_x, test_y = svhn.load_numpy(
            False, binarize_y=True)  #norb.load_resized(size, binarize_y=True)
        extra_x, extra_y = svhn.load_numpy_extra(False, binarize_y=True)
        x = {
            'x': np.hstack((train_x, extra_x)),
            'y': np.hstack((train_y, extra_y))
        }
        ndict.shuffleCols(x)

        print('Performing PCA, can take a few minutes... ',
              f_enc,
              f_dec,
              pca_params=pp.PCA(x['x'][:, :10000], cutoff=600, toFloat=True))
        ndict.savez(pca_params, logdir + 'pca_params')
        print('Done.')

        n_y = 10
        x = {'x': f_enc(x['x']).astype(np.float32)}
        x_valid = {'x': f_enc(test_x).astype(np.float32)}
        L_valid = 1
        n_x = x['x'].shape[0]
        dim_input = (size, size)
        n_batch = 5000
        colorImg = True
        bernoulli_x = False
        byteToFloat = False
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        type_px = 'gaussian'
        nonlinear = 'softplus'

    elif dataset == 'hyper':
        # Hyperspectral images:

        # Import 1 file of the dataset
        # TODO: import more files: Edit hyperspectralData.py

        #I added the hyperspectralData file in the anglepy library
        from hyperspectralData import HyperspectralData

        train_x, train_y, valid_x, valid_y, test_x, test_y = HyperspectralData(
        ).load_numpy(100000)

        #Dim input: How it has to be written like an image. We said that is:
        dim_input = (67, 4)
        n_x = train_x.shape[0]  #Dimension of our data vector.

        x = {'x': train_x.astype(np.float32)}
        x_valid = {'x': valid_x.astype(np.float32)}
        x_test = {'x': test_x.astype(np.float32)}
        L_valid = 1
        type_qz = 'gaussianmarg'
        type_pz = 'gaussianmarg'
        nonlinear = 'softplus'
        type_px = 'bernoulli'
        n_train = train_x.shape[1]
        n_batch = 1000
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        weight_decay = float(n_batch) / n_train
        #Write the hyperparameters used:
        with open(logdir + 'AA_hyperparameters.txt', 'w') as file:
            file.write("L_valid: " + str(L_valid) + '\n')
            file.write("type_qz: " + type_qz + '\n')
            file.write("type_pz: " + type_pz + '\n')
            file.write("Nonlinear: " + nonlinear + '\n')
            file.write("type_px: " + type_px + '\n')
            file.write("n_train: " + str(n_train) + '\n')
            file.write("n_batch: " + str(n_batch) + '\n')
            file.write("colorImg: " + str(colorImg) + '\n')
            file.write("bernoulli_x: " + str(bernoulli_x) + '\n')
            file.write("byteToFloat: " + str(byteToFloat) + '\n')
            file.close()
        # Write the headers for the csv file output:
        with open(logdir + 'AA_results.txt', 'w') as file:
            # Like a csv file:
            file.write("Step" + ',' + "TimeElapsed" + ',' +
                       "LowerboundMinibatch" + ',' + "LowerboundValid" + ',' +
                       "NumStepNotImproving" + '\n')
            file.close()

    # Construct model
    from anglepy.models import GPUVAE_Z_X
    updates = get_adam_optimizer(learning_rate=3e-4, weight_decay=weight_decay)
    model = GPUVAE_Z_X(updates,
                       n_x,
                       n_hidden,
                       n_z,
                       n_hidden[::-1],
                       nonlinear,
                       nonlinear,
                       type_px,
                       type_qz=type_qz,
                       type_pz=type_pz,
                       prior_sd=100,
                       init_sd=1e-3)
    #---------------
    # SetaSouto:
    # The [::-1] is to reverse the list.
    #---------------

    if False:
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412689061/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412676966/'
        #dir = '/Users/dpkingma/results/learn_z_x_svhn_bernoulli_300-(1000, 1000)_l1l2_sharing_and_1000HU_1412695481/'
        #dir = '/Users/dpkingma/results/learn_z_x_mnist_binarized_50-(500, 500)_mog_1412695455/'
        #dir = '/Users/dpkingma/results/gpulearn_z_x_svhn_pca_300-(500, 500)__1413904756/'
        dir = '/home/ubuntu/results/gpulearn_z_x_mnist_50-[500, 500]__1414259423/'
        w = ndict.loadz(dir + 'w_best.ndict.tar.gz')
        v = ndict.loadz(dir + 'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)

    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]

    # Progress hook
    def hook(epoch, t, ll):
        '''
        Documented by SetaSouto, may contains errors.

        :epoch: Number of the current step.
        :t: Time elapsed from the beginning.
        :ll: Loglikelihood (?).
        '''

        if epoch % 10 != 0: return

        ll_valid, _ = model.est_loglik(x_valid,
                                       n_samples=L_valid,
                                       n_batch=n_batch,
                                       byteToFloat=byteToFloat)

        # Log
        ndict.savez(ndict.get_value(model.v), logdir + 'v')
        ndict.savez(ndict.get_value(model.w), logdir + 'w')

        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir + 'v_best')
            ndict.savez(ndict.get_value(model.w), logdir + 'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if ll_valid_stats[1] > 100:
                print("Finished")
                with open(logdir + 'hook.txt', 'a') as f:
                    print(f, "Finished")
                exit()

        # This will be showing the current results and write them in a file:
        with open(logdir + 'AA_results.txt', 'a') as file:
            # Like a csv file:
            file.write(
                str(epoch) + ',' + str(t) + ',' + str(ll) + ',' +
                str(ll_valid) + ',' + str(ll_valid_stats[1]) + '\n')
            file.close()
        print("-------------------------")
        print("Current results:")
        print(" ")
        print("Step:", epoch)
        print("Time elapsed:", t)
        print("Loglikelihood minibatch:", ll)
        print("Loglikelihood validSet:", ll_valid)
        print("N not improving:", ll_valid_stats[1])
        #print(epoch, t, ll, ll_valid, ll_valid_stats)

        #This print the file where are written the stats.
        #with open(logdir+'hook.txt', 'a') as f:
        #print(f, epoch, t, ll, ll_valid, ll_valid_stats)

        # Graphics
        if gfx and epoch % gfx_freq == 0:

            #tail = '.png'
            tail = '-' + str(epoch) + '.png'

            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}

            if 'pca' not in dataset and 'random' not in dataset and 'normalized' not in dataset:

                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'q_w0' + tail, 'PNG')

                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]),
                                                 dim_input,
                                                 True,
                                                 colorImg=colorImg)
                image.save(logdir + 'out_w' + tail, 'PNG')

                if 'out_unif' in w:
                    image = paramgraphics.mat_to_img(f_dec(
                        w['out_unif'].reshape((-1, 1))),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'out_unif' + tail, 'PNG')

                if n_z == 2:
                    n_width = 10
                    import scipy.stats
                    z = {'z': np.zeros((2, n_width**2))}
                    for i in range(0, n_width):
                        for j in range(0, n_width):
                            z['z'][0, n_width * i + j] = scipy.stats.norm.ppf(
                                float(i) / n_width + 0.5 / n_width)
                            z['z'][1, n_width * i + j] = scipy.stats.norm.ppf(
                                float(j) / n_width + 0.5 / n_width)

                    x, _, _z = model.gen_xz({}, z, n_width**2)
                    if dataset == 'mnist':
                        x = 1 - _z['x']
                    image = paramgraphics.mat_to_img(f_dec(_z['x']), dim_input)
                    image.save(logdir + '2dmanifold' + tail, 'PNG')
                else:
                    _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples),
                                                     dim_input,
                                                     colorImg=colorImg)
                    image.save(logdir + 'samples' + tail, 'PNG')

                    #x_samples = _x['x']
                    #image = paramgraphics.mat_to_img(x_samples, dim_input, colorImg=colorImg)
                    #image.save(logdir+'samples2'+tail, 'PNG')

            else:
                # Model with preprocessing

                if 'w0' in v:
                    image = paramgraphics.mat_to_img(f_dec(v['w0'][:].T),
                                                     dim_input,
                                                     True,
                                                     colorImg=colorImg)
                    image.save(logdir + 'q_w0' + tail, 'PNG')

                image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]),
                                                 dim_input,
                                                 True,
                                                 colorImg=colorImg)
                image.save(logdir + 'out_w' + tail, 'PNG')

                _x, _, _z_confab = model.gen_xz({}, {}, n_batch=144)
                x_samples = f_dec(_z_confab['x'])
                x_samples = np.minimum(np.maximum(x_samples, 0), 1)
                image = paramgraphics.mat_to_img(x_samples,
                                                 dim_input,
                                                 colorImg=colorImg)
                image.save(logdir + 'samples' + tail, 'PNG')

    # Optimize
    #SFO
    dostep = epoch_vae_adam(model,
                            x,
                            n_batch=n_batch,
                            bernoulli_x=bernoulli_x,
                            byteToFloat=byteToFloat)
    loop_va(dostep, hook)

    pass
예제 #11
0
def main(n_z, n_hidden, dataset, seed, gfx=True, _size=None):
    '''Learn a variational auto-encoder with generative model p(x,y,z)=p(y)p(z)p(x|y,z).
    x and y are (always) observed.
    I.e. this cannot be used for semi-supervised learning
    '''
    assert (type(n_hidden) == tuple or type(n_hidden) == list)
    assert type(n_z) == int
    assert isinstance(dataset, str)

    print('gpulearn_yz_x', n_z, n_hidden, dataset, seed)

    import time
    logdir = 'results/gpulearn_yz_x_' + dataset + '_' + str(n_z) + '-' + str(
        n_hidden) + '-' + str(int(time.time())) + '/'
    if not os.path.exists(logdir): os.makedirs(logdir)
    print('logdir:', logdir)

    np.random.seed(seed)

    # Init data
    if dataset == 'mnist':
        '''
        What works well:
        100-2-100 (Generated digits stay bit shady)
        1000-2-1000 (Needs pretty long training)
        '''
        import anglepy.data.mnist as mnist

        # MNIST
        size = 28
        train_x, train_y, valid_x, valid_y, test_x, test_y = mnist.load_numpy(
            size, binarize_y=True)
        f_enc, f_dec = lambda x: x, lambda x: x
        x = {
            'x': train_x[:, :].astype(np.float32),
            'y': train_y[:, :].astype(np.float32)
        }
        x_valid = {
            'x': valid_x.astype(np.float32),
            'y': valid_y.astype(np.float32)
        }
        L_valid = 1
        dim_input = (size, size)
        n_x = size * size
        n_y = 10
        n_batch = 1000
        colorImg = False
        bernoulli_x = True
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 2
        type_px = 'bernoulli'

    elif dataset == 'norb':
        # resized NORB dataset, reshuffled
        import anglepy.data.norb as norb
        size = _size  #48
        train_x, train_y, test_x, test_y = norb.load_resized(size,
                                                             binarize_y=True)
        _x = {'x': train_x, 'y': train_y}
        ndict.shuffleCols(_x)
        train_x = _x['x']
        train_y = _x['y']

        # Do PCA
        f_enc, f_dec, pca_params = pp.PCA(_x['x'][:, :10000],
                                          cutoff=2000,
                                          toFloat=False)
        ndict.savez(pca_params, logdir + 'pca_params')

        x = {
            'x': f_enc(train_x).astype(np.float32),
            'y': train_y.astype(np.float32)
        }
        x_valid = {
            'x': f_enc(test_x).astype(np.float32),
            'y': test_y.astype(np.float32)
        }

        L_valid = 1
        n_x = x['x'].shape[0]
        n_y = 5
        dim_input = (size, size)
        n_batch = 1000  #23400/900 = 27
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 1
        type_px = 'gaussian'

    elif dataset == 'norb_instances':
        # resized NORB dataset with the instances as classes
        import anglepy.data.norb2 as norb2
        size = _size  #48
        x, y = norb2.load_numpy_subclasses(size, binarize_y=True)
        _x = {'x': x, 'y': y}
        ndict.shuffleCols(_x)

        # Do pre=processing
        if True:
            # Works
            f_enc, f_dec, pca_params = pp.PCA(_x['x'][:, :10000],
                                              cutoff=600,
                                              global_sd=True,
                                              toFloat=True)
            ndict.savez(pca_params, logdir + 'pca_params')
        elif False:
            # Doesn't work
            f_enc, f_dec, pp_params = pp.normalize_noise(_x['x'][:, :50000],
                                                         noise_sd=0.01,
                                                         global_sd=True,
                                                         toFloat=True)
        else:
            # Doesn't work
            f_enc, f_dec, params = pp.normalize_random(x=x[:, :10000],
                                                       global_sd=True,
                                                       toFloat=True)
            ndict.savez(params, logdir + 'normalize_random_params')

        n_valid = 5000
        x = {
            'x': f_enc(_x['x'][:, :-n_valid]).astype(np.float32),
            'y': _x['y'][:, :-n_valid].astype(np.float32)
        }
        x_valid = {
            'x': f_enc(_x['x'][:, :n_valid]).astype(np.float32),
            'y': _x['y'][:, :n_valid].astype(np.float32)
        }

        L_valid = 1
        n_x = x['x'].shape[0]
        n_y = 50
        dim_input = (size, size)
        n_batch = 5000  #23400/900 = 27
        colorImg = False
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 1
        type_px = 'gaussian'

    elif dataset == 'svhn':
        # SVHN dataset
        import anglepy.data.svhn as svhn
        size = 32
        train_x, train_y, test_x, test_y = svhn.load_numpy(
            False, binarize_y=True)  #norb.load_resized(size, binarize_y=True)
        extra_x, extra_y = svhn.load_numpy_extra(False, binarize_y=True)
        x = {
            'x': np.hstack((train_x, extra_x)),
            'y': np.hstack((train_y, extra_y))
        }
        ndict.shuffleCols(x)

        #f_enc, f_dec, (x_sd, x_mean) = pp.preprocess_normalize01(train_x, True)
        f_enc, f_dec, pca_params = pp.PCA(x['x'][:, :10000],
                                          cutoff=1000,
                                          toFloat=True)
        ndict.savez(pca_params, logdir + 'pca_params')

        n_y = 10
        x = {
            'x': f_enc(x['x']).astype(np.float32),
            'y': x['y'].astype(np.float32)
        }
        x_valid = {
            'x': f_enc(test_x).astype(np.float32),
            'y': test_y.astype(np.float32)
        }
        L_valid = 1
        n_x = x['x'].shape[0]
        dim_input = (size, size)
        n_batch = 5000
        colorImg = True
        bernoulli_x = False
        byteToFloat = False
        mosaic_w = 5
        mosaic_h = 2
        type_px = 'gaussian'

    # Init model
    n_hidden_q = n_hidden
    n_hidden_p = n_hidden
    from anglepy.models import GPUVAE_YZ_X
    updates = get_adam_optimizer(alpha=3e-4,
                                 beta1=0.9,
                                 beta2=0.999,
                                 weight_decay=0)
    model = GPUVAE_YZ_X(updates,
                        n_x,
                        n_y,
                        n_hidden_q,
                        n_z,
                        n_hidden_p[::-1],
                        'softplus',
                        'softplus',
                        type_px=type_px,
                        type_qz='gaussianmarg',
                        type_pz='gaussianmarg',
                        prior_sd=1,
                        uniform_y=True)

    if False:
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414094291/'
        dir = '/home/ubuntu/results/gpulearn_yz_x_svhn_300-(500, 500)-1414163488/'
        w = ndict.loadz(dir + 'w_best.ndict.tar.gz')
        v = ndict.loadz(dir + 'v_best.ndict.tar.gz')
        ndict.set_value(model.w, w)
        ndict.set_value(model.v, v)

    # Some statistics for optimization
    ll_valid_stats = [-1e99, 0]

    # Fixed sample for visualisation
    z_sample = {
        'z':
        np.repeat(np.random.standard_normal(size=(n_z, 12)), 12,
                  axis=1).astype(np.float32)
    }
    y_sample = {
        'y':
        np.tile(
            np.random.multinomial(1, [1. / n_y] * n_y, size=12).T, (1, 12))
    }

    # Progress hook
    def hook(epoch, t, ll):

        if epoch % 10 != 0:
            return

        ll_valid, _ = model.est_loglik(x_valid,
                                       n_samples=L_valid,
                                       n_batch=n_batch,
                                       byteToFloat=byteToFloat)

        if math.isnan(ll_valid):
            print("NaN detected. Reverting to saved best parameters")
            ndict.set_value(model.v, ndict.loadz(logdir + 'v.ndict.tar.gz'))
            ndict.set_value(model.w, ndict.loadz(logdir + 'w.ndict.tar.gz'))
            return

        if ll_valid > ll_valid_stats[0]:
            ll_valid_stats[0] = ll_valid
            ll_valid_stats[1] = 0
            ndict.savez(ndict.get_value(model.v), logdir + 'v_best')
            ndict.savez(ndict.get_value(model.w), logdir + 'w_best')
        else:
            ll_valid_stats[1] += 1
            # Stop when not improving validation set performance in 100 iterations
            if False and ll_valid_stats[1] > 1000:
                print("Finished")
                with open(logdir + 'hook.txt', 'a') as f:
                    print("Finished", file=f)
                exit()

        # Log
        ndict.savez(ndict.get_value(model.v), logdir + 'v')
        ndict.savez(ndict.get_value(model.w), logdir + 'w')
        print(epoch, t, ll, ll_valid)
        with open(logdir + 'hook.txt', 'a') as f:
            print(t, ll, ll_valid, file=f)

        if gfx:
            # Graphics

            v = {i: model.v[i].get_value() for i in model.v}
            w = {i: model.w[i].get_value() for i in model.w}

            tail = '-' + str(epoch) + '.png'

            image = paramgraphics.mat_to_img(f_dec(v['w0x'][:].T),
                                             dim_input,
                                             True,
                                             colorImg=colorImg)
            image.save(logdir + 'q_w0x' + tail, 'PNG')

            image = paramgraphics.mat_to_img(f_dec(w['out_w'][:]),
                                             dim_input,
                                             True,
                                             colorImg=colorImg)
            image.save(logdir + 'out_w' + tail, 'PNG')

            _x = {'y': np.random.multinomial(1, [1. / n_y] * n_y, size=144).T}
            _, _, _z_confab = model.gen_xz(_x, {}, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']),
                                             dim_input,
                                             colorImg=colorImg)
            image.save(logdir + 'samples' + tail, 'PNG')

            _, _, _z_confab = model.gen_xz(y_sample, z_sample, n_batch=144)
            image = paramgraphics.mat_to_img(f_dec(_z_confab['x']),
                                             dim_input,
                                             colorImg=colorImg)
            image.save(logdir + 'samples_fixed' + tail, 'PNG')

            if n_z == 2:

                import Image
                import ImageFont
                import ImageDraw

                n_width = 10
                submosaic_offset = 15
                submosaic_width = (dim_input[1] * n_width)
                submosaic_height = (dim_input[0] * n_width)
                mosaic = Image.new(
                    "RGB", (submosaic_width * mosaic_w,
                            submosaic_offset + submosaic_height * mosaic_h))

                for digit in range(0, n_y):
                    if digit >= mosaic_h * mosaic_w: continue

                    _x = {}
                    n_batch_plot = n_width * n_width
                    _x['y'] = np.zeros((n_y, n_batch_plot))
                    _x['y'][digit, :] = 1
                    _z = {'z': np.zeros((2, n_width**2))}
                    for i in range(0, n_width):
                        for j in range(0, n_width):
                            _z['z'][0, n_width * i + j] = scipy.stats.norm.ppf(
                                float(i) / n_width + 0.5 / n_width)
                            _z['z'][1, n_width * i + j] = scipy.stats.norm.ppf(
                                float(j) / n_width + 0.5 / n_width)

                    _x, _, _z_confab = model.gen_xz(_x,
                                                    _z,
                                                    n_batch=n_batch_plot)
                    x_samples = _z_confab['x']
                    image = paramgraphics.mat_to_img(f_dec(x_samples),
                                                     dim_input,
                                                     colorImg=colorImg,
                                                     tile_spacing=(0, 0))

                    #image.save(logdir+'samples_digit_'+str(digit)+'_'+tail, 'PNG')
                    mosaic_x = (digit % mosaic_w) * submosaic_width
                    mosaic_y = submosaic_offset + int(
                        digit / mosaic_w) * submosaic_height
                    mosaic.paste(image, (mosaic_x, mosaic_y))

                draw = ImageDraw.Draw(mosaic)
                draw.text((1, 1),
                          "Epoch #" + str(epoch) + " Loss=" + str(int(ll)))

                #plt.savefig(logdir+'mosaic'+tail, format='PNG')
                mosaic.save(logdir + 'mosaic' + tail, 'PNG')

                #x_samples = _x['x']
                #image = paramgraphics.mat_to_img(f_dec(x_samples), dim_input, colorImg=colorImg)
                #image.save(logdir+'samples2'+tail, 'PNG')

    # Optimize
    dostep = epoch_vae_adam(model,
                            x,
                            n_batch=n_batch,
                            bernoulli_x=bernoulli_x,
                            byteToFloat=byteToFloat)
    loop_va(dostep, hook)

    pass