ae.update_catalog("ed_"+name,info)
update_status(ae, "history", {"name": name,
"loss": losscomb(0, h),
"acc": losscomb(1, h),
"date": time_str()})
if continue_encdecs:
opt, dr, gbs, gsf, l2 = 'adadelta', 0.001, 0.001, 2, 0.001
name = "o-"+str(opt)+"-dr-"+str(dr)+"-gbs-"+str(gbs)+"-gsf-"+str(gsf)+"-l1-"+str(0)+"-l2-"+str(l2)
ae.enc_opt = opt
ae.drop_rate = dr
ae.sigma_base = gbs
ae.sigma_fact = gsf
ae.l1 = 0
ae.l2 = l2
ae.epochs = 20
ae.load_encdecs(name)
h = ae.pretrain(name=name)
info = {"acc_pre_"+str(i): float(e[-1][1]) for (i,e) in enumerate(h)}
info.update({("loss_pre_"+str(i)): float(e[-1][0]) for (i,e) in enumerate(h)})
ae.update_catalog("ed_"+name,info)
update_status(ae, "history", {"name": name,
"loss": losscomb(0, h),
"acc": losscomb(1, h),
"date": time_str()})
### FINETUNE ###
if finetune:
ae.load_encdecs("best")
encdec_opts = {key: ae.get_from_catalog(key, "ed_best") for key in encdec_opt_strings}
#sink = Sink([ae], lambda r: print(r) if (not r is None) else None)
#print('have sink')
options = [(opt,cl,dr,gbs,gsf,l2)
name = "o-"+str(opt)+"-dr-"+str(dr)+"-gbs-"+str(gbs)+"-gsf-"+str(gsf)+"-l1-"+str(l1)+"-l2-"+str(l2)
ae.enc_opt = opt
ae.drop_rate = dr
ae.sigma_base = gbs
ae.sigma_fact = gsf
ae.l1 = l1
ae.l2 = l2
ae.new_encdecs(use_dropout=(dr != 0), use_noise=(gbs != 0))
print("training encdec "+name)
h = ae.pretrain(name=name)
info = {"acc_pre_"+str(i): float(e[-1][1]) for (i,e) in enumerate(h)}
info.update({("loss_pre_"+str(i)): float(e[-1][0]) for (i,e) in enumerate(h)})
ae.update_catalog("ed_"+name,info)
### FINETUNE ###
cur_left_out = nsm.cross_val_keys[nsm.cross_val_index]
ae.load_encdecs("best")
ae.finetune(name="x-"+cur_left_out+"-val", train_encdecs=True, test_data=nsm.data[cur_left_out])
info = {}
for key in ae.previous_data:
print(key)
arr = np.array(ae.previous_data[key])
phl = len(phase_names)
eye = np.identity(phl)
tc = to_categorical
eva = ae.model.evaluate(arr,
tc(map(lambda n: phase_names.index(n)
,list(repeat(key, len(arr))))
,phl)
,show_accuracy=True)
counts = np.sum(map(lambda p: eye[np.argmax(p)], ae.model.predict(arr)), axis=0)
print(eva)
ae.sigma_fact = gsf
ae.l1 = l1
ae.l2 = l2
ae.new_encdecs(use_dropout=(dr != 0), use_noise=(gbs != 0))
print("training encdec " + name)
h = ae.pretrain(name=name)
info = {
"acc_pre_" + str(i): float(e[-1][1])
for (i, e) in enumerate(h)
}
info.update({("loss_pre_" + str(i)): float(e[-1][0])
for (i, e) in enumerate(h)})
ae.update_catalog("ed_" + name, info)
### FINETUNE ###
cur_left_out = nsm.cross_val_keys[nsm.cross_val_index]
ae.load_encdecs("best")
ae.finetune(name="x-" + cur_left_out + "-val",
train_encdecs=True,
test_data=nsm.data[cur_left_out])
info = {}
for key in ae.previous_data:
print(key)
arr = np.array(ae.previous_data[key])
phl = len(phase_names)
eye = np.identity(phl)
tc = to_categorical
eva = ae.model.evaluate(arr,
tc(
map(lambda n: phase_names.index(n),
list(repeat(key, len(arr)))), phl),
show_accuracy=True)
