def generate_rdm_all(nnet,
name,
rdm_type=analysis.SPEARMAN,
save_files=True,
title="RDM training combined",
from_file=False,
delete_blank_states=True,
collapse_rdm=True):
if not from_file:
if rdm_type != analysis.SPEARMAN:
raise Exception("not implemented")
hidden_both, accuracy_totals_both, accuracy_fullseqs_both = test_network_all(
nnet)
hidden_ari, accuracy_totals_ari, accuracy_fullseqs_ari = test_network_ari(
nnet, blanks=True)
hidden_bev, accuracy_totals_bev, accuracy_fullseqs_bev = test_network_bev(
nnet, blanks=True)
print("Both: {0}, {1}".format(accuracy_totals_both,
accuracy_fullseqs_both))
print("Ari: {0}, {1}".format(accuracy_totals_ari,
accuracy_fullseqs_ari))
print("Bev: {0}, {1}".format(accuracy_totals_bev,
accuracy_fullseqs_bev))
hidden = utils.flatten_onelevel(hidden_bev) +\
utils.flatten_onelevel(hidden_ari) +\
utils.flatten_onelevel(hidden_both)
rdmatrix = analysis.rdm_euclidian(hidden)
utils.save_object(name + "rdmat", rdmatrix)
else:
rdmatrix = utils.load_object(name + "rdmat")
return model2.process_matrix(rdmatrix, delete_blank_states)
def generate_rdm_all_gradient(nnet,
name,
blanks,
rdm_type=analysis.SPEARMAN,
save_files=True,
title="RDM training combined",
from_file=False,
delete_blank_states=True):
if not from_file:
if rdm_type != analysis.SPEARMAN:
raise Exception("not implemented")
hidden_both, accuracy_totals_both, accuracy_fullseqs_both = test_network_all(
nnet)
hidden_ari, accuracy_totals_ari, accuracy_fullseqs_ari = test_network_ari(
nnet, blanks)
hidden_bev, accuracy_totals_bev, accuracy_fullseqs_bev = test_network_bev(
nnet, blanks)
print("Both: {0}, {1}".format(accuracy_totals_both,
accuracy_fullseqs_both))
print("Ari: {0}, {1}".format(accuracy_totals_ari,
accuracy_fullseqs_ari))
print("Bev: {0}, {1}".format(accuracy_totals_bev,
accuracy_fullseqs_bev))
hidden = utils.flatten_onelevel(hidden_bev) +\
utils.flatten_onelevel(hidden_ari) +\
utils.flatten_onelevel(hidden_both)
hidden_left = []
hidden_right = []
for vector in hidden:
hidden_left.append(vector[:len(vector) // 2])
hidden_right.append(vector[len(vector) // 2:])
# Now cut the hidden layer in two.
rdmatrix_left = analysis.rdm_spearman(hidden_left)
rdmatrix_right = analysis.rdm_spearman(hidden_right)
# save the massive rdm for debug purposes (so that I don't have to generate it all over again everytime).
utils.save_object(name + "rdmatright", rdmatrix_right)
utils.save_object(name + "rdmatleft", rdmatrix_left)
else:
rdmatrix_left = utils.load_object(name + "rdmatleft")
rdmatrix_right = utils.load_object(name + "rdmatright")
rdmatrix_left, labels = model2.process_matrix(rdmatrix_left,
delete_blank_states)
rdmatrix_right, _ = model2.process_matrix(rdmatrix_right,
delete_blank_states)
return rdmatrix_left, rdmatrix_right, labels
def run_models_with_noise(models, noise):
goal_list = [
"g_1_make_coffee", "g_1_make_tea", "g_2_add_grounds", "g_2_add_cream",
"g_2_add_sugar", "g_2_drink", "g_2_dip_teabag"
]
all_action_targets_str = utils.flatten_onelevel(
[tce.action_list[goal] for goal in goal_list])
all_action_targets = [
utils.str_to_onehot(action, tce.TeaCoffeeData.actions_list)
for action in all_action_targets_str
]
# Gather actual outputs
all_outputs = []
for model in models:
all_outputs.append(run_model_with_noise(model, noise))
# Now check which actions and which goals they got correctly
num_bad_actions = np.zeros(len(all_action_targets))
for output in all_outputs:
for i, action in enumerate(output):
if (all_action_targets[i] != action).any():
num_bad_actions[i] += 1
print(all_action_targets_str[i],
utils.onehot_to_str(action, tce.TeaCoffeeData.actions_list))
return num_bad_actions
def generate_rdm_bev(nnet,
name,
rdm_type=analysis.SPEARMAN,
save_files=True,
title="RDM training beverage only"):
if rdm_type != analysis.SPEARMAN:
raise Exception("not implemented")
hidden, accuracy_totals, accuracy_fullseqs = test_network_bev(nnet)
print(accuracy_totals)
print(accuracy_fullseqs)
hidden = utils.flatten_onelevel(hidden)
rdmatrix = analysis.rdm_spearman(hidden)
labels = utils.flatten_onelevel(task.label_seqs_bev)
name = name.replace('.', '_') + '_' + rdm_type
if save_files:
np.savetxt(name + "_rdm_mat.txt", rdmatrix, delimiter="\t", fmt='%.2e')
analysis.plot_rdm(rdmatrix,
labels,
title + " spearman rho matrix",
figsize=30,
fontsize=0.6)
if save_files:
plt.savefig(name + '_rdm.png', dpi=300, bbox_inches='tight')
plt.clf()