nb_threads = 2 # for input queue
stddev_init_nn = 0.01 # neural net initialization
log_dir = 'logs_svae'
verbose = False # log device placement
# set size_minibatch=64
schedule = create_schedule({
'dataset': 'auto',
'method': 'svae-cvi',
'lr': [0.0003], # adam stepsize
'lrcvi': [0.2], # cvi stepsize (convex combination)
'decay_rate': [0.95],
'K': 10, # nb components
'L': [6], # latent dimensionality
'U': 50, # hidden units
'seed': 0
})
# set size_minibatch=100 (as in Johnson+, 2016); initialize phi_gmm with gmm_prior (line 181)
schedule = create_schedule({
'dataset': 'pinwheel',
'method': 'svae-cvi',
'lr': [0.01], # adam stepsize
'lrcvi': [0.1], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': [2], # latent dimensionality
'U': 50, # hidden units
def boxplot_noisy_pinwheel():
# plots performance of latent SMM for increasing noise level
path = '../noisy_pinwheel_wu'
schedule_svae_smm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'svae-cvi-smm',
'lr': [0.01], # adam stepsize
'lrcvi': [0.005], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'DoF': 5,
'noise_level': 0.1,
'seed': range(1, 4)
})
_, mse1, mse1m, mse1s = load_m_std(path, schedule_svae_smm, 'mse_te')
schedule_svae_smm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'svae-cvi-smm',
'lr': [0.01], # adam stepsize
'lrcvi': [0.005], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'DoF': 5,
'noise_level': 0.3,
'seed': range(1, 4)
})
_, mse3, mse3m, mse3s = load_m_std(path, schedule_svae_smm, 'mse_te')
schedule_svae_smm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'svae-cvi-smm',
'lr': [0.01], # adam stepsize
'lrcvi': [0.0005], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'DoF': 5,
'noise_level': 0.5,
'seed': range(1, 4)
})
_, mse5, mse5m, mse5s = load_m_std(path, schedule_svae_smm, 'mse_te')
schedule_svae_smm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'svae-cvi-smm',
'lr': [0.01], # adam stepsize
'lrcvi': [0.0005], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'DoF': 5,
'noise_level': 0.7,
'seed': range(1, 4)
})
_, mse7, mse7m, mse7s = load_m_std(path, schedule_svae_smm, 'mse_te')
lsmm_mean = np.vstack([mse1m[-1], mse3m[-1], mse5m[-1], mse7m[-1]])
lsmm_std = np.vstack([mse1s[-1], mse3s[-1], mse5s[-1], mse7s[-1]])
# extract svae gmm
lgmm_mean = np.zeros(4)
lgmm_std = np.zeros(4)
for i, lvl in enumerate([0.1, 0.3, 0.5, 0.7]):
schedule_svae_gmm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'svae-cvi',
'lr': [0.01], # adam stepsize
'lrcvi': [0.005], # cvi stepsize (convex combination)
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'DoF': 5,
'noise_level': lvl,
'seed': range(1, 4)
})
_, mse, mean, std = load_m_std(path, schedule_svae_gmm, 'mse_te')
lgmm_mean[i] = mean[-1]
lgmm_std[i] = std[-1]
# extract simple gmm
gmm_mean = np.zeros(4)
gmm_std = np.zeros(4)
for i, lvl in enumerate([0.1, 0.3, 0.5, 0.7]):
schedule_svae_gmm = create_schedule({
'dataset': 'noisy-pinwheel',
'method': 'gmm',
'K': 10, # nb components
'noise_level': lvl,
'seed': range(1, 4)
})
_, mse, mean, std = load_m_std(path, schedule_svae_gmm, 'mse_te')
gmm_mean[i] = mean[-1]
gmm_std[i] = std[-1]
fig, ax = newfig(0.5, ratio_hw=customized_figsize)
# ax.set_xscale('log')
ax.set_yscale('log')
#(4, 3)
ax.errorbar(np.transpose([1, 2, 3, 4]),
gmm_mean,
yerr=gmm_std,
color="C0",
label="GMM")
ax.errorbar(np.transpose([1, 2, 3, 4]),
lgmm_mean,
yerr=lgmm_std,
color="C3",
label="SAN-GMM")
ax.errorbar(np.transpose([1, 2, 3, 4]),
lsmm_mean,
yerr=lsmm_std,
color="C8",
label="SAN-TMM")
ax.set_xticklabels(["10%", "30%", "50%", "70%"])
ax.set_xticks([1, 2, 3, 4])
ax.set_xlim([0.7, 4.3])
ax.set_ylim([0.5, 200])
ax.set_xlabel('Ratio Outliers')
ax.set_ylabel('MSE')
ax.grid(which='both', linewidth=width_gridlines)
ax.legend()
save_plot('noisy-pinwheel_errorbars', path='./')
def compare_auto():
mpl.rcParams.update({"legend.fontsize": 4})
# path_svae = '../raiden_auto/auto_svae'
path_svae_smm = '../auto_svae_smm'
path_svae = '../logs_svae'
path_vae = '../logs_vae'
path_gmm = '../auto_gmm'
# load SVAE-GMM
schedule_svae = create_schedule({
'dataset': 'auto',
'method': 'svae-cvi',
'lr': [0.0003], # adam stepsize
'lrcvi': [0.2], # cvi stepsize (convex combination)
'decay_rate': [0.95],
'K': 10, # nb components
'L': [6], # latent dimensionality
'U': 50, # hidden units
'seed': range(0, 10) # todo
})
# load SVAE-SMM
schedule_svae_smm = create_schedule({
'dataset': 'auto',
'method': 'svae-cvi-smm',
'lr': [0.0003], # adam stepsize
'lrcvi': [0.2], # cvi stepsize (convex combination)
'decay_rate': [0.95],
'K': 10, # nb components
'L': [6], # latent dimensionality
'U': 50, # hidden units
'DoF': 5,
'seed': range(10) # todo
})
# load VAE
schedule_vae = create_schedule({
'dataset': 'auto',
'method': 'vae',
'lr': [0.0003], # adam stepsize
'L': [6], # latent dimensionality
'U': 50, # hidden units
'seed': range(10)
})
# load VAE
schedule_gmm = create_schedule({
'method': 'gmm',
'dataset': 'auto',
'K': 5,
'seed': range(10)
})
label_svae_smm = 'SAN-SMM'
label_svae = 'SAN-GMM'
label_vae = 'VAE'
label_gmm = 'GMM'
fig, ax = newfig(0.5, ratio_hw=customized_figsize)
plot_summary(path_gmm, schedule_gmm, 'mse_te', ax, label_gmm, marker='o')
plot_summary(path_vae, schedule_vae, 'mse_te', ax, label_vae, marker='+')
plot_summary(path_svae,
schedule_svae,
'mse_te',
ax,
label_svae,
marker='D')
plot_summary(path_svae_smm,
schedule_svae_smm,
'mse_te',
ax,
label_svae_smm,
marker='s')
ax.set_xscale('log')
ax.set_yscale('log')
ax.set_xlim([400, 100000])
ax.set_ylim([1, 200])
ax.grid(which='both', linewidth=width_gridlines)
ax.set_xlabel('Iterations')
ax.set_ylabel('MSE')
# import matplotlib.lines as mlines
# handles, labels = ax.get_legend_handles_labels()
# handles.insert(2, mlines.Line2D([], [], color='C1', marker='x', markersize=7))
# labels.insert(2, 'SVAE (Johnson et. al.)')
# ax.legend(handles, labels, loc='lower left')
ax.legend()
save_plot('auto_mse_te_smm', path='./')
def compare_pinwheel():
# path = '../pinwheel'
path_san = '../pinwheel_new'
path_mjj = '../../svae/experiments/pinwheel_new'
# load SVAE
schedule_svae = create_schedule({
'dataset': 'pinwheel',
'method': 'svae-cvi',
'lr': [0.01], # adam stepsize
'lrcvi': [0.1], # cvi stepsize (convex combination)
'decay_rate': 1,
'delay': 0,
'K': 10, # nb components
'L': 2, # latent dimensionality
'U': 40, # hidden units
'seed': range(10)
})
# load VAE
schedule_vae = create_schedule({
'dataset': 'pinwheel',
'method': 'vae',
'lr': [0.005], # adam stepsize
'L': 2, # latent dimensionality
'U': 40, # hidden units
'seed': range(10)
})
# load GMM
schedule_gmm = create_schedule({
'method': 'gmm',
'dataset': 'pinwheel',
'K': 10,
'seed': range(10)
})
# load SVAE-Johnson
schedule_mjj = create_schedule({
'method': 'svae_mjj',
'dataset': 'pinwheel',
'inner_loop': [100],
'sgd_step_size': [10],
'adam_step_size': [0.005],
'K': 10,
'seed': range(10)
})
label_svae = 'SAN-GMM'
label_mjj = 'SVAE (Johnson et. al.)'
label_vae = 'VAE'
label_gmm = 'GMM'
# ax = ax3
fig, ax = newfig(0.5, ratio_hw=customized_figsize)
ax.set_xscale('log')
ax.set_yscale('log')
ax.grid(which='both', linewidth=width_gridlines)
ax.set_xlim([100, 10000])
ax.set_ylim([0.5, 100])
plot_summary(path_san,
schedule_gmm,
'mse_te',
ax,
label_gmm,
marker='o',
start=1)
plot_summary(path_san,
schedule_vae,
'mse_te',
ax,
label_vae,
marker='+',
start=1)
plot_summary(path_mjj,
schedule_mjj,
'test_rmse_100',
ax,
label_mjj,
marker='x',
start=1)
plot_summary(path_san,
schedule_svae,
'mse_te',
ax,
label_svae,
marker='D',
start=1)
ax.set_xlabel('Iterations')
ax.set_ylabel('MSE')
ax.grid(which='both', linewidth=width_gridlines)
# ax.legend()
# plt.show()
save_plot('pinwheel_mse_te', path='./')
measurement_freq = 2500
imputation_freq = 50000
checkpoint_freq = 25000
seed_data = 0
log_dir = '../logs_vae'
nb_threads = 5 # for input queue
stddev_init = 0.01
schedule = create_schedule({
'dataset': 'auto',
'method': 'vae',
'lr': [0.0003], # adam stepsize
'L': [6], # latent dimensionality
'U': 50, # hidden units
'seed': range(0, 10)
})
# schedule = create_schedule({
# 'dataset': 'mnist-small',
# 'method': 'vae',
# 'lr': [0.001], # adam stepsize
# 'L': [10], # latent dimensionality
# 'U': 50, # hidden units
# 'seed': 0
# })
####################################################################################################################
path_dataset = '../datasets'
ratio_tr = 0.7
ratio_val = None
missing_data_ratio = 0.1
nb_iters = 200
seed = 0
seed_data = 0
log_dir = '../debug'
schedule = create_schedule({
'method': 'smm',
'dataset': 'aggregation', # 't4.8k',
'K': 10,
'kappa': 9999,
'seed': 0
})
####################################################################################################################
for config_id, config in enumerate(schedule):
K = config['K']
kappa = config['kappa']
seed = config['seed']
dataset = config['dataset']
print("Experiment %d with config\n%s\n" % (config_id, str(config)))
ratio_val = None
missing_data_ratio = 0.1
nb_iters = 500
measurement_freq = 10
K = 10 # nb components
seed = 0
seed_data = 0
log_dir = '../pinwheel_new'
schedule = create_schedule({
'method': 'gmm',
'dataset': 'pinwheel',
'K': 8,
'seed': range(10)
})
#
# schedule = create_schedule({
# 'method': 'gmm',
# 'dataset': 'auto',
# 'K': 10,
# 'seed': 10
# })
schedule = create_schedule({
'method': 'gmm',
'dataset': 'aggregation',
'K': 10,
'seed': 0
ratio_val = None
missing_data_ratio = 0.1
nb_iters = 500
measurement_freq = 10
K = 10 # nb components
seed = 0
seed_data = 0
log_dir = '../pinwheel_new'
schedule = create_schedule({
'method': 'gmm',
'dataset': 'pinwheel',
'K': 8,
'seed': range(10)
})
#
# schedule = create_schedule({
# 'method': 'gmm',
# 'dataset': 'auto',
# 'K': 10,
# 'seed': 10
# })
####################################################################################################################
for config_id, config in enumerate(schedule):
K = config['K']
seed = config['seed']
