def plot_objectives(nnet, xp_path, title_suffix, xlabel, file_prefix):
"""
plot train, validation, and test objective in a combined plot
"""
objectives = OrderedDict([("train", nnet.diag['train']['objective'])])
if nnet.data.n_val > 0:
objectives["val"] = nnet.diag['val']['objective']
objectives["test"] = nnet.diag['test']['objective']
plot_line(objectives, title="Objective " + title_suffix, xlabel=xlabel, ylabel="Objective", log_scale=True,
export_pdf=(xp_path + "/" + file_prefix + "obj"))
def plot_accuracy(nnet, xp_path, title_suffix, xlabel, file_prefix):
"""
plot accuracy of train, val, and test set per epoch.
"""
acc = OrderedDict([("train", nnet.diag['train']['acc'])])
if nnet.data.n_val > 0:
acc["val"] = nnet.diag['val']['acc']
acc["test"] = nnet.diag['test']['acc']
plot_line(acc, title="Accuracy " + title_suffix, xlabel=xlabel, ylabel="Accuracy (%) ", y_min=-5, y_max=105,
export_pdf=(xp_path + "/" + file_prefix + "accuracy"))
def plot_center_c_diagnostics(nnet, xp_path, title_suffix, xlabel, file_prefix):
"""
plot the norm of center c and the means of the outputs on the train, val, and test set per epoch
"""
norms = OrderedDict([("c", nnet.diag['network']['c_norm']),
("mean_train", nnet.diag['train']['output_mean_norm'])])
if nnet.data.n_val > 0:
norms["mean_val"] = nnet.diag['val']['output_mean_norm']
norms["mean_test"] = nnet.diag['test']['output_mean_norm']
plot_line(norms, title="Norms in output space " + title_suffix, xlabel=xlabel, ylabel="norm values ",
log_scale=True, export_pdf=(xp_path + "/" + file_prefix + "c_mean_norms"))
diffs = OrderedDict([("train", nnet.diag['train']['c_mean_diff'])])
if nnet.data.n_val > 0:
diffs["val"] = nnet.diag['val']['c_mean_diff']
diffs["test"] = nnet.diag['test']['c_mean_diff']
plot_line(diffs, title="Distance from c to means " + title_suffix, xlabel=xlabel, ylabel="distance ",
log_scale=True, export_pdf=(xp_path + "/" + file_prefix + "c_mean_diffs"))
def plot_objective_with_parts(nnet, xp_path, title_suffix, xlabel, file_prefix, pretrain=False):
"""
plot train, validation, and test objective (and their parts)
"""
for which_set in ['train', 'val', 'test']:
if (which_set == 'val') & (nnet.data.n_val == 0):
continue
# Plot objective (and its parts)
objective = OrderedDict([("objective", nnet.diag[which_set]['objective']),
("emp. loss", nnet.diag[which_set]['emp_loss']),
("l2 penalty", nnet.diag['network']['l2_penalty'])])
if Cfg.reconstruction_penalty and not pretrain:
objective["reconstruction penalty"] = nnet.diag[which_set]['reconstruction_penalty']
if Cfg.svdd_loss and not pretrain:
objective["R"] = nnet.diag['network']['R']
title = which_set.title() + " objective " + title_suffix
plot_line(objective, title=title, xlabel=xlabel, ylabel="Objective", log_scale=True,
export_pdf=(xp_path + "/" + file_prefix + "obj_" + which_set))
def plot_parameter_norms(nnet, xp_path, title_suffix, xlabel, file_prefix):
"""
plot norms of network parameters (and parameter updates)
"""
# plot norms of parameters for each unit of dense layers
params = OrderedDict()
n_layer = 0
for layer in nnet.trainable_layers:
if layer.isdense:
for unit in range(layer.num_units):
name = "W" + str(n_layer + 1) + str(unit + 1)
params[name] = nnet.diag['network']['W_norms'][n_layer][unit, :]
if layer.b is not None:
name = "b" + str(n_layer + 1) + str(unit + 1)
params[name] = nnet.diag['network']['b_norms'][n_layer][unit, :]
n_layer += 1
plot_line(params, title="Norms of network parameters " + title_suffix, xlabel=xlabel, ylabel="Norm", log_scale=True,
export_pdf=(xp_path + "/" + file_prefix + "param_norms"))
# plot norms of parameter differences between updates for each layer
params = OrderedDict()
n_layer = 0
for layer in nnet.trainable_layers:
if layer.isdense | layer.isconv:
name = "dW" + str(n_layer + 1)
params[name] = nnet.diag['network']['dW_norms'][n_layer]
if layer.b is not None:
name = "db" + str(n_layer + 1)
params[name] = nnet.diag['network']['db_norms'][n_layer]
n_layer += 1
plot_line(params, title="Absolute differences of parameter updates " + title_suffix, xlabel=xlabel, ylabel="Norm",
log_scale=True, export_pdf=(xp_path + "/" + file_prefix + "param_diff_norms"))
def plot_auc(nnet, xp_path, title_suffix, xlabel, file_prefix):
"""
plot auc time series of train, val, and test set.
"""
auc = OrderedDict()
for which_set in ['train', 'val', 'test']:
if (which_set == 'val') & (nnet.data.n_val == 0):
continue
if which_set == 'train':
y = nnet.data._y_train
if which_set == 'val':
y = nnet.data._y_val
if which_set == 'test':
y = nnet.data._y_test
if sum(y) > 0:
auc[which_set] = nnet.diag[which_set]['auc']
plot_line(auc, title="AUC " + title_suffix, xlabel=xlabel, ylabel="AUC", y_min=-0.05, y_max=1.05,
export_pdf=(xp_path + "/" + file_prefix + "auc"))