)
vocab_size = len(w_ind)
ont_size = len(c_ind)
comp_starttime = time.time()
event_ae = EventAE(
num_args,
vocab_size,
ont_size,
args.hyp_hidden_size,
wc_hidden_sizes,
cc_hidden_sizes,
word_dim=args.word_dim,
concept_dim=args.concept_dim,
word_rep_param=args.change_word_rep,
hyp_model_type=args.hyp_model_type,
wc_pref_model_type=args.wc_pref_model_type,
cc_pref_model_type=args.cc_pref_model_type,
rec_model_type=args.rec_model_type,
init_hyp_strengths=init_hyp_strengths,
relaxed=args.use_relaxation,
no_hyp=args.no_hyp,
wc_lr_wp_rank=args.wc_lr_wp_rank,
cc_lr_wp_rank=args.cc_lr_wp_rank,
)
if use_pretrained_wordrep:
print >> sys.stderr, "Using pretrained word representations from %s" % (args.pt_rep)
num_words_covered = 0
init_wordrep = event_ae.vocab_rep.get_value()
for word in w_ind:
print >>vocab_file, w, ind
vocab_file.close()
ont_file = codecs.open("ont.txt", "w", "utf-8")
for c, ind in c_ind.items():
print >>ont_file, c, ind
ont_file.close()
rev_w_ind = {ind:word for word, ind in w_ind.items()}
rev_c_ind = {ind:concept for concept, ind in c_ind.items()}
train_data = zip(x_data, y_s_data)
sanity_test_data = random.sample(train_data, len(train_data)/10)
vocab_size = len(w_ind)
ont_size = len(c_ind)
event_ae = EventAE(num_args, vocab_size, ont_size, hyp_hidden_size, wc_hidden_sizes, cc_hidden_sizes)
eaes = [copy.deepcopy(event_ae) for _ in range(num_procs)]
part_train_funcs = [eae.get_train_func(learning_rate) for eae in eaes]
#train_func = event_ae.get_train_func(learning_rate)
post_score_func = event_ae.get_posterior_func()
def synchronize_param():
all_repr_params = [ [param.get_value() for param in eae.repr_params] for eae in eaes ]
avg_repr_params = [numpy.mean(param, axis=0) for param in zip(*all_repr_params)]
event_ae.set_repr_params(avg_repr_params)
for eae in eaes:
eae.set_repr_params(avg_repr_params)
all_hyp_params = [ [param.get_value() for param in eae.hyp_model.get_params()] for eae in eaes ]
avg_hyp_params = [numpy.mean(param, axis=0) for param in zip(*all_hyp_params)]
event_ae.hyp_model.set_params(avg_hyp_params)
for y_ind in y_datum:
if y_ind not in test_train_ont_map:
all_present = False
ignored_y_cands += 1
break
fixed_y.append(test_train_ont_map[y_ind])
if all_present:
fixed_y_s.append(fixed_y)
#fixed_y_s = [[test_train_ont_map[ind] for ind in y_datum] for y_datum in y_s_datum]
fixed_data.append((fixed_x, fixed_y_s))
print >>sys.stderr, "Ignored an average of %f y cands per data point"%(float(ignored_y_cands)/len(fixed_data))
print >>sys.stderr, len(x_data), len(y_s_data), len(fixed_data)
event_ae = EventAE(num_args, vocab_size, ont_size, hyp_hidden_size, wc_hidden_sizes, cc_hidden_sizes, relaxed=use_relaxation, no_hyp=args.no_hyp, wc_pref_model_type=args.wc_pref_model_type, cc_pref_model_type=args.cc_pref_model_type, rec_model_type=args.rec_model_type, init_hyp_strengths=init_hyp_strengths)
#for i, param in enumerate(repr_params):
# event_ae.repr_params[i].set_value(param)
event_ae.vocab_rep.set_value(vocab_rep)
event_ae.ont_rep.set_value(ont_rep)
if not args.no_hyp:
event_ae.hyp_model.set_params(hyp_params)
for i, param_dict in enumerate(wcp_params):
for j in param_dict:
event_ae.wc_pref_models[i][j].set_params(param_dict[j])
if not use_relaxation:
for i, params in enumerate(ccp_params):
event_ae.cc_pref_models[i].set_params(params)
event_ae.rec_model.set_params(rec_params)
