def plot_mvmm(mvmm, inches=8, save_dir=None):
"""
Plots loss history and estimated Pi matrix.
"""
if save_dir is not None:
os.makedirs(save_dir, exist_ok=True)
# TODO: maybe add Pi start
loss_vals = mvmm.opt_data_['history']['loss_val']
################
# Loss history #
################
plot_loss_history(loss_vals,
loss_name='Observed data negative log-likelihood')
if save_dir is not None:
fpath = join(save_dir, 'loss_history.png')
save_fig(fpath)
###############
# Pi estimate
################
plt.figure(figsize=(inches, inches))
plot_Pi(mvmm.weights_mat_)
plt.title("Estimated Pi")
if save_dir is not None:
fpath = join(save_dir, 'Pi_est.png')
save_fig(fpath)
def plot_loss_history_from_idx(idx):
estimator = sp_mvmm.estimators_[idx]
if idx == 0:
loss_vals = estimator.opt_data_['history']['loss_val']
# title = 'MVMM'
else:
loss_vals = estimator.opt_data_['adpt_opt_data']['history'][
'loss_val']
# title = 'Penalty = {:1.2f}'.format(estimator.eval_pen_base)
plot_loss_history(loss_vals)
def plot_sp_opt_history(sp_mvmm_gs, save_dir):
for idx, est in enumerate(sp_mvmm_gs.estimators_):
n_blocks = sp_mvmm_gs.est_n_blocks_[idx]
if type(est) == MVMM:
loss_vals = est.opt_data_['history']['obs_nll']
loss_name = 'Obs NLL'
else:
loss_vals = est.opt_data_['adpt_opt_data']['history']['loss_val']
loss_name = "eval pen loss"
plot_loss_history(loss_vals, loss_name=loss_name)
plt.title('{} steps'.format(len(loss_vals)))
os.makedirs(save_dir, exist_ok=True)
save_fig(
os.path.join(save_dir,
'opt_history__n_blocks_{}.png'.format(n_blocks)))
def plot_log_pen_mvmm(mvmm, inches=8, save_dir=None):
if save_dir is not None:
os.makedirs(save_dir, exist_ok=True)
# info = get_log_pen_mvmm_info(mvmm)
# if save_dir is not None:
# # TODO: save this
# save_dir
# else:
# print(info)
Pi = mvmm.weights_mat_
zero_thresh = 1e-10 # not sure if we need this
summary, Pi_comm = community_summary(Pi, zero_thresh=zero_thresh)
Pi_symlap_spec = eigh_Lsym_bp(Pi)[0]
if 'init_params' in mvmm.opt_data_['history']:
Pi_init = mvmm.opt_data_['history']['weights'].reshape(
Pi.shape) # TODO: check this
Pi_init_symlap_spec = eigh_Lsym_bp(Pi_init)[0]
else:
Pi_init = None
obs_nll = mvmm.opt_data_['history']['obs_nll']
loss_vals = mvmm.opt_data_['history']['loss_val']
####################
# Initial weights #
###################
if Pi_init is not None:
plt.figure(figsize=(inches, inches))
plot_Pi(Pi_init)
plt.title('weights initial value')
if save_dir is not None:
fpath = join(save_dir, 'weights_init.png')
save_fig(fpath)
######################
# Estimated weights #
######################
plt.figure(figsize=(2 * inches, inches))
plt.subplot(1, 2, 1)
plot_Pi(Pi)
plt.title('weights estimate, n_blocks={}'.format(summary['n_communities']))
plt.subplot(1, 2, 2)
plot_Pi(Pi, mask=Pi_comm < zero_thresh)
plt.title('weights estimate, block diagonal perm')
if save_dir is not None:
fpath = join(save_dir, 'weights_est.png')
save_fig(fpath)
##########################
# Spectrum of BD weights #
##########################
plt.figure(figsize=(inches, inches))
idxs = np.arange(1, len(Pi_symlap_spec) + 1)
plt.plot(idxs, Pi_symlap_spec, marker='.', label='Estimate')
if Pi_init is not None:
plt.plot(idxs, Pi_init_symlap_spec, marker='.', label="Initial")
plt.title('weights estimate spectrum')
plt.ylim(0)
plt.legend()
if save_dir is not None:
fpath = join(save_dir, 'weights_spectrum.png')
save_fig(fpath)
###########################
# Obs NLL for entire path #
###########################
plt.figure(figsize=[inches, inches])
plot_loss_history(obs_nll, loss_name="Obs NLL")
if save_dir is not None:
fpath = join(save_dir, 'obs_nll.png')
save_fig(fpath)
plt.figure(figsize=[inches, inches])
plot_loss_history(loss_vals, loss_name="log penalized obs nll")
if save_dir is not None:
fpath = join(save_dir, 'loss_vals.png')
save_fig(fpath)
def plot_bd_mvmm(mvmm, inches=8, save_dir=None):
"""
Initial BD weights, Estimated BD weights, spectrums of both
Number of steps in each adaptive stage
Evals of entire path
Loss history for each segment
"""
if save_dir is not None:
os.makedirs(save_dir, exist_ok=True)
info = get_bd_mvmm_info(mvmm)
if save_dir is not None:
# TODO: save this
save_dir
else:
print(info)
# BD weight estimate
bd_weights = mvmm.bd_weights_
zero_thresh = mvmm.zero_thresh
summary, Pi_comm = community_summary(bd_weights, zero_thresh=zero_thresh)
bd_weights_symlap_spec = eigh_Lsym_bp(bd_weights)[0]
# initial BD weights
bd_weights_init = mvmm.opt_data_['adpt_opt_data']['adapt_pen_history'][
'opt_data'][0]['history']['init_params']['bd_weights']
bd_weights_init_symlap_spec = eigh_Lsym_bp(bd_weights_init)[0]
# optimization history
adpt_history = mvmm.opt_data_['adpt_opt_data']['adapt_pen_history'][
'opt_data']
if 'raw_eval_sum' in adpt_history[0]['history']:
n_steps = [
len(adpt_history[i]['history']['raw_eval_sum'])
for i in range(len(adpt_history))
]
n_steps_cumsum = np.cumsum(n_steps)
raw_eval_sum = \
np.concatenate([adpt_history[i]['history']['raw_eval_sum']
for i in range(len(adpt_history))])
else:
raw_eval_sum = None
n_steps = None
n_steps_cumsum = None
obs_nll = np.concatenate([
adpt_history[i]['history']['obs_nll'] for i in range(len(adpt_history))
])
if mvmm.opt_data_['ft_opt_data'] is not None:
fine_tune_obs_nll = mvmm.opt_data_['ft_opt_data']['history']['obs_nll']
else:
fine_tune_obs_nll = None
######################
# Initial BD weights #
######################
plt.figure(figsize=(inches, inches))
plot_Pi(bd_weights_init)
plt.title('BD weights initial value')
if save_dir is not None:
fpath = join(save_dir, 'BD_weights_init.png')
save_fig(fpath)
########################
# Estimated BD weights #
########################
plt.figure(figsize=(2 * inches, inches))
plt.subplot(1, 2, 1)
plot_Pi(bd_weights)
plt.title('BD weights estimate, n_blocks={}'.format(
summary['n_communities']))
plt.subplot(1, 2, 2)
plot_Pi(bd_weights, mask=Pi_comm < zero_thresh)
plt.title('BD weights estimate, block diagonal perm')
if save_dir is not None:
fpath = join(save_dir, 'BD_weights_est.png')
save_fig(fpath)
##########################
# Spectrum of BD weights #
##########################
plt.figure(figsize=(inches, inches))
idxs = np.arange(1, len(bd_weights_symlap_spec) + 1)
plt.plot(idxs, bd_weights_symlap_spec, marker='.', label='Estimate')
plt.plot(idxs, bd_weights_init_symlap_spec, marker='.', label="Initial")
plt.title('BD weights estimate spectrum')
plt.ylim(0)
plt.legend()
if save_dir is not None:
fpath = join(save_dir, 'BD_weights_spectrum.png')
save_fig(fpath)
##################################
# Number of steps for each stage #
##################################
if n_steps is not None:
plt.figure(figsize=(inches, inches))
idxs = np.arange(1, len(n_steps) + 1)
plt.plot(idxs, n_steps, marker='.')
plt.ylim(0)
plt.ylabel("Number of steps")
plt.xlabel("Adaptive stage")
if save_dir is not None:
fpath = join(save_dir, 'n_steps.png')
save_fig(fpath)
###########################
# Obs NLL for entire path #
###########################
plt.figure(figsize=[inches, inches])
plot_loss_history(obs_nll, loss_name="Obs NLL (entire path)")
if save_dir is not None:
fpath = join(save_dir, 'path_obs_nll.png')
save_fig(fpath)
#########################
# Evals for entire path #
#########################
if raw_eval_sum is not None:
plt.figure(figsize=[inches, inches])
plt.plot(np.log10(raw_eval_sum), marker='.')
plt.ylabel('log10(sum smallest evals)')
plt.xlabel('step')
plt.title('Eigenvalue history (entire path)')
for s in n_steps_cumsum:
plt.axvline(s - 1, color='grey')
if save_dir is not None:
fpath = join(save_dir, 'path_evals.png')
save_fig(fpath)
###########################
# Losses for each segment #
###########################
if save_dir is not None:
segment_dir = join(save_dir, 'segments')
os.makedirs(segment_dir, exist_ok=True)
for i in range(len(adpt_history)):
loss_vals = adpt_history[i]['history']['loss_val']
plot_loss_history(loss_vals,
'loss val, adapt segment {}'.format(i + 1))
if save_dir is not None:
fpath = join(segment_dir, 'loss_history_{}.png'.format(i + 1))
save_fig(fpath)
##########################
# fine tune loss history #
##########################
if fine_tune_obs_nll is not None:
plot_loss_history(fine_tune_obs_nll, 'fine tune obs NLL')
if save_dir is not None:
fpath = join(segment_dir, 'fine_tune_loss_history.png')
save_fig(fpath)
save_fig(
os.path.join(save_dir, 'log_pen_mvmm',
'{}_vs_{}.png'.format(param_name, measure)))
################
# loss history #
################
log_pen_mvmm_gs = models['log_pen_mvmm']
# mvmm = log_pen_mvmm_gs.estimators_[idx]
mvmm = log_pen_mvmm_gs.best_estimator_
start_loss_val = mvmm.start_.opt_data_['history']['obs_nll']
final_loss_val = mvmm.final_.opt_data_['history']['loss_val']
plot_loss_history(start_loss_val)
save_fig(os.path.join(save_dir, 'log_pen_mvmm', 'loss_history_start.png'))
plot_loss_history(final_loss_val)
save_fig(os.path.join(save_dir, 'log_pen_mvmm', 'loss_history_final.png'))
###########
# SP MVMM #
###########
if 'sp_mvmm' not in to_exclude:
os.makedirs(os.path.join(save_dir, 'sp_mvmm'), exist_ok=True)
models['sp_mvmm'].init_fit_data_.\
to_csv(os.path.join(save_dir, 'sp_mvmm',
'init_fit_data_.csv'))
# model selection
est_n_comp = estimator.n_components
res_writer.write("Cat data GMM estimated number of components: {}".
format(est_n_comp))
# _model_sel_dir = make_and_get_dir(model_sel_dir, 'cat_gmm')
# plt.figure(figsize=(inches, inches))
plot_model_selection(cat_model, save_dir=model_sel_dir,
name_stub='cat_gmm',
title='GMM on concatenated data',
inches=inches)
# optimization history
loss_history = estimator.opt_data_['history']['obs_nll']
plot_loss_history(loss_history,
loss_name='Observed data negative log-likelihood')
save_fig(os.path.join(opt_diag_dir, 'cat_best_model_opt_history.png'))
else:
print("cat isnt fit...")
########################
# view marginal models #
########################
n_views = len(view_models)
for v in range(n_views):
estimator = view_models[v].best_estimator_
sel_models['view_' + str(v)] = estimator