def main():
"""
dialogs: dict, dialogs of the dataset.
emo_dict: dict, emotions of all utterances.
out_dict: dict, output logits of emotion classifier.
"""
dialogs = joblib.load('./data/dialog_iemocap.pkl')
emo_dict = joblib.load('./data/emo_all_iemocap.pkl')
out_dict = joblib.load('./data/outputs.pkl')
# set log
logging.basicConfig(stream=sys.stdout,
format='%(asctime)s %(levelname)s:%(message)s',
level=logging.INFO,
datefmt='%I:%M:%S')
# Split dialogs by speakers
spk_dialogs = utils.split_dialog(dialogs)
logging.info("Average number of speaker's utterances per dialog: %.3f" % np.mean(
[len(i) for i in spk_dialogs.values()]))
# arguments
args = parse_args()
print('=' * 60 + '\n')
logging.info('Parameters are:\n%s\n', json.dumps(vars(args), sort_keys=False, indent=4))
print('=' * 60 + '\n')
if args.transition_bias > 0:
# Use given p_0
p_0 = args.transition_bias
else:
# Estimate p_0 of ALL dialogs.
p_0, total_transit = utils.transition_bias(spk_dialogs, emo_dict)
print("\n"+"#"*50)
logging.info('p_0: %.3f , total transition: %d\n' % (p_0, total_transit))
print("#"*50)
bias_dict = utils.get_val_bias(spk_dialogs, emo_dict)
print("#"*50+"\n")
trace = []
label = []
org_pred = []
DED = bs.BeamSearch(p_0, args.crp_alpha, args.num_state,
args.beam_size, args.test_iteration, emo_dict, out_dict)
for i, dia in enumerate(dialogs):
logging.info("Decoding dialog: {}/{}, {}".format(i,len(dialogs),dia))
# Apply p_0 estimated from other 4 sessions.
DED.transition_bias = bias_dict[dia[:5]]
# Beam search decoder
out = DED.decode(dialogs[dia])
trace += out
label += [utils.convert_to_index(emo_dict[utt]) for utt in dialogs[dia]]
org_pred += [np.argmax(out_dict[utt]) for utt in dialogs[dia]]
if args.verbosity > 0:
logging.info("Output: {}\n".format(out))
print("#"*50+"\n")
# Print the results of emotino classifier module
uar, acc, conf = utils.evaluate(org_pred, label)
logging.info('Original performance: uar: %.3f, acc: %.3f' % (uar, acc))
# Eval ded outputs
results = vars(args)
uar, acc, conf = utils.evaluate(trace, label)
logging.info('DED performance: uar: %.3f, acc: %.3f' % (uar, acc))
logging.info('Confusion matrix:\n%s' % conf)
# Save results
results['uar'] = uar
results['acc'] = acc
results['conf'] = str(conf)
logging.info('Save results:')
logging.info('\n%s\n', json.dumps(results, sort_keys=False, indent=4))
json.dump(results, open(args.out_dir+'/%s.json' % args.result_file, "w"))
def main():
"""
dialogs: dict, dialogs of the dataset.
emo_dict: dict, emotions of all utterances.
out_dict: dict, output logits of emotion classifier.
"""
dialogs = joblib.load('./data/dialogs.pkl')
#dialogs_edit = joblib.load('./data/dialog_rearrange_4emo_iemocap.pkl')
#dialogs = joblib.load('./data/dialog_iemocap.pkl')
#out_dict = joblib.load('./data/outputs.pkl')
emo_dict = joblib.load('./data/emo_all.pkl')
output_fold1 = joblib.load('./data/original_iaan/utt_logits_outputs_fold1.pkl')
output_fold2 = joblib.load('./data/original_iaan/utt_logits_outputs_fold2.pkl')
output_fold3 = joblib.load('./data/original_iaan/utt_logits_outputs_fold3.pkl')
output_fold4 = joblib.load('./data/original_iaan/utt_logits_outputs_fold4.pkl')
output_fold5 = joblib.load('./data/original_iaan/utt_logits_outputs_fold5.pkl')
out_dict = {}
for utt in output_fold1:
if utt[4] == '1':
out_dict[utt] = output_fold1[utt]
elif utt[4] == '2':
out_dict[utt] = output_fold2[utt]
elif utt[4] == '3':
out_dict[utt] = output_fold3[utt]
elif utt[4] == '4':
out_dict[utt] = output_fold4[utt]
elif utt[4] == '5':
out_dict[utt] = output_fold5[utt]
'''
out_dict = joblib.load('./data/DAG_emo_output_all_fold_multi8_rearrange.pkl')
C2C_emo_dict = joblib.load('./data/C2C_4emo_all_iemocap.pkl')
U2U_emo_dict = joblib.load('./data/U2U_4emo_all_iemocap.pkl')
LogisticRegression_emo_shift_output = joblib.load('./data/LogisticRegression_emo_shift_output.pkl')
RandomForest_emo_shift_output = joblib.load('./data/RandomForest_emo_shift_output.pkl')
iaan_emo_shift_output = joblib.load('./data/iaan_emo_shift_output_rearrange.pkl')
mlp_sklearn_emo_shift_output = joblib.load('./data/MLPSklearn_emo_shift_output.pkl')
svm_emo_shift_output = joblib.load('./data/SVM_emo_shift_output.pkl')
MLPPytorch_emo_shift_output = joblib.load('./data/MLPPytorch_emo_shift_output.pkl')
'''
#DAG_emo_shift_output_all_fold = joblib.load('./data/DAG_emo_shift_output_all_fold_multi4_rearrange.pkl')
# set log
logging.basicConfig(stream=sys.stdout,
format='%(asctime)s %(levelname)s:%(message)s',
level=logging.INFO,
datefmt='%I:%M:%S')
# Split dialogs by speakers
spk_dialogs = utils.split_dialog(dialogs)
#spk_dialogs_edit = utils.split_dialog(dialogs_edit)
logging.info("Average number of speaker's utterances per dialog: %.3f" % np.mean(
[len(i) for i in spk_dialogs.values()]))
# arguments
args = parse_args()
print('=' * 60 + '\n')
logging.info('Parameters are:\n%s\n', json.dumps(vars(args), sort_keys=False, indent=4))
print('=' * 60 + '\n')
if args.transition_bias > 0:
# Use given p_0
p_0 = args.transition_bias
else:
# Estimate p_0 of ALL dialogs.
p_0, total_transit = utils.transition_bias(spk_dialogs, emo_dict)
print("\n"+"#"*50)
logging.info('p_0: %.3f , total transition: %d\n' % (p_0, total_transit))
print("#"*50)
bias_dict = utils.get_val_bias(spk_dialogs, emo_dict)
print("#"*50+"\n")
# rule based ddCRP alpha (ref: https://arxiv.org/pdf/1911.01266.pdf)
#alpha_dict = utils.get_val_alpha(spk_dialogs_edit, emo_dict)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=0, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=3)
trace = []
label = []
org_pred = []
DED = bs.BeamSearch(p_0, {}, args.crp_alpha, args.num_state,
args.beam_size, args.test_iteration, emo_dict, out_dict, {})
model_pred_dict = {}
for i, dia in enumerate(dialogs):
logging.info("Decoding dialog: {}/{}, {}".format(i,len(dialogs),dia))
# Apply p_0 estimated from other 4 sessions.
DED.transition_bias = bias_dict[dia[:5]]
#DED.crp_alpha = alpha_dict[dia[:5]]
# Apply emo_trans_prob_dict estimated from other 4 sessions.
#DED.emo_trans_prob_dict = intra_emo_trans_prob_dict[dia[:5]]
# Beam search decoder
out = DED.decode(dialogs[dia], model_pred_dict)
trace += out
label += [utils.convert_to_index(emo_dict[utt]) for utt in dialogs[dia]]
org_pred += [np.argmax(out_dict[utt]) for utt in dialogs[dia]]
if args.verbosity > 0:
logging.info("Output: {}\n".format(out))
print("#"*50+"\n")
# Print the results of emotino classifier module
uar, acc, conf = utils.evaluate(org_pred, label)
logging.info('Original performance: uar: %.3f, acc: %.3f' % (uar, acc))
# Eval ded outputs
results = vars(args)
uar, acc, conf = utils.evaluate(trace, label)
logging.info('DED performance: uar: %.3f, acc: %.3f' % (uar, acc))
logging.info('Confusion matrix:\n%s' % conf)
# Save results
results['uar'] = uar
results['acc'] = acc
results['conf'] = str(conf)
logging.info('Save results:')
logging.info('\n%s\n', json.dumps(results, sort_keys=False, indent=4))
json.dump(results, open(args.out_dir+'/%s.json' % args.result_file, "w"))
joblib.dump(model_pred_dict, './outputs/preds_4.pkl')
def main():
"""
dialogs: dict, dialogs of the dataset.
emo_dict: dict, emotions of all utterances.
out_dict: dict, output logits of emotion classifier.
"""
train_dialogs = joblib.load('./data/train_dialogs.pkl')
val_dialogs = joblib.load('./data/val_dialogs.pkl')
test_dialogs = joblib.load('./data/test_dialogs.pkl')
train_val_dialogs = dict(train_dialogs)
train_val_dialogs.update(val_dialogs)
#dialogs_edit = joblib.load('./data/dialog_6emo_iemocap.pkl')
train_emo_dict = joblib.load('./data/train_emo_all.pkl')
val_emo_dict = joblib.load('./data/val_emo_all.pkl')
test_emo_dict = joblib.load('./data/test_emo_all.pkl')
train_val_emo_dict = dict(train_emo_dict)
train_val_emo_dict.update(val_emo_dict)
train_out_dict = joblib.load('./data/train_utts_logit.pkl')
val_out_dict = joblib.load('./data/train_utts_logit.pkl')
test_out_dict = joblib.load('./data/test_utts_logit.pkl')
train_val_out_dict = dict(train_out_dict)
train_val_out_dict.update(val_out_dict)
'''
for utt in train_out_dict:
train_out_dict[utt] = softmax(train_out_dict[utt])
for utt in val_out_dict:
val_out_dict[utt] = softmax(val_out_dict[utt])
for utt in test_out_dict:
test_out_dict[utt] = softmax(test_out_dict[utt])
for utt in train_val_out_dict:
test_out_dict[utt] = softmax(train_val_out_dict[utt])
'''
#C2C_emo_dict = joblib.load('./data/C2C_4emo_all_iemocap.pkl')
#U2U_emo_dict = joblib.load('./data/U2U_6emo_all_iemocap.pkl')
# set log
logging.basicConfig(stream=sys.stdout,
format='%(asctime)s %(levelname)s:%(message)s',
level=logging.INFO,
datefmt='%I:%M:%S')
# Split dialogs by speakers
#train_val_spk_dialogs = utils.split_dialog(train_val_dialogs)
train_spk_dialogs = utils.split_dialog(train_dialogs)
#test_spk_dialogs = utils.split_dialog(test_dialogs)
logging.info("Average number of speaker's utterances per dialog: %.3f" % np.mean(
[len(i) for i in train_spk_dialogs.values()]))
# arguments
args = parse_args()
print(args)
print('=' * 60 + '\n')
logging.info('Parameters are:\n%s\n', json.dumps(vars(args), sort_keys=False, indent=4))
print('=' * 60 + '\n')
if args.transition_bias > 0:
# Use given p_0
p_0 = args.transition_bias
else:
# Estimate p_0 of ALL dialogs.
#train_val_p_0, train_val_total_transit = utils.transition_bias(train_val_spk_dialogs, train_val_emo_dict)
train_p_0, train_total_transit = utils.transition_bias(train_spk_dialogs, train_emo_dict)
#test_p_0, test_total_transit = utils.transition_bias(test_spk_dialogs, test_emo_dict)
print("\n"+"#"*50)
logging.info('p_0: %.3f , total transition: %d\n' % (train_p_0, train_total_transit))
print("#"*50)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(emo_dict, dialogs_edit, softmax_or_not=0, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(C2C_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=2)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=1, Bi_or_Tri=3)
#seq_emo_trans_prob_dict, intra_emo_trans_prob_dict = utils.get_val_emo_trans_prob(U2U_emo_dict, dialogs, softmax_or_not=0, Bi_or_Tri=3)
trace = []
label = []
org_pred = []
DED = bs.BeamSearch(train_p_0, {}, args.crp_alpha, args.num_state,
args.beam_size, args.test_iteration, test_emo_dict, test_out_dict)
model_pred_dict = {}
for i, dia in enumerate(test_dialogs):
logging.info("Decoding dialog: {}/{}, {}".format(i,len(test_dialogs),dia))
# Apply p_0 estimated from train_val data
DED.transition_bias = train_p_0
# Apply emo_trans_prob_dict estimated from other 4 sessions.
#DED.emo_trans_prob_dict = seq_emo_trans_prob_dict[dia[:5]]
# Beam search decoder
out = DED.decode(test_dialogs[dia], model_pred_dict)
trace += out
label += [utils.convert_to_index(test_emo_dict[utt]) for utt in test_dialogs[dia]]
org_pred += [np.argmax(test_out_dict[utt]) for utt in test_dialogs[dia]]
if args.verbosity > 0:
logging.info("Output: {}\n".format(out))
print("#"*50+"\n")
# Print the results of emotino classifier module
uar, acc, weighted_f1, conf = utils.evaluate(org_pred, label)
logging.info('Original performance: uar: %.4f, acc: %.4f, weighted f1: %.4f' % (uar, acc, weighted_f1))
print(conf)
# Eval ded outputs
results = vars(args)
uar, acc, weighted_f1, conf = utils.evaluate(trace, label)
logging.info('DED performance: uar: %.4f, acc: %.4f, weighted f1: %.4f' % (uar, acc, weighted_f1))
#logging.info('Confusion matrix:\n%s' % conf)
print(conf)
# Save results
results['uar'] = uar
results['acc'] = acc
results['weighted_f1'] = weighted_f1
results['conf'] = str(conf)
logging.info('Save results:')
logging.info('\n%s\n', json.dumps(results, sort_keys=False, indent=4))
json.dump(results, open(args.out_dir+'/%s.json' % args.result_file, "w"))
joblib.dump(model_pred_dict, './outputs/preds_4.pkl')