def multi_win_during_val(val_fn, inputs, augs, targets):
for idraw in [
100,
]:
for jdraw in [
100,
]:
inpt_multiwin = data_aug(inputs,
mu,
sigma,
deterministic=True,
idraw=idraw,
jdraw=jdraw)
err_pat, output_pat = val_fn(inpt_multiwin, augs, targets)
if 'weight' in locals():
dis = ((idraw / 100.0 - 1.0)**2 +
(jdraw / 100.0 - 1.0)**2)**0.5
wei = np.exp(-np.square(dis) / 2.0 / 0.5**2)
weight += wei
err += err_pat * wei
output += output_pat * wei
else:
dis = ((idraw / 100.0 - 1.0)**2 +
(jdraw / 100.0 - 1.0)**2)**0.5
weight = np.exp(-np.square(dis) / 2.0 / 1.0**2)
err = err_pat * weight
output = output_pat * weight
return err / weight, output / weight
def train_round(num_epochs, network, valid_num, train_fn, val_fn, classn,
X_train, a_train, y_train, X_test, a_test, y_test):
print("Starting training...")
print("TrLoss\t\tVaLoss\t\tVaHamm\t\tAUC\t\tCMatrix\t\t\tEpochs\t\tTime")
batchsize = 100
start_time = time.time()
for epoch in range(num_epochs + 1):
train_err = 0
train_batches = 0
for batch in iterate_minibatches(X_train,
a_train,
y_train,
batchsize,
shuffle=True):
inputs, augs, targets = batch
inputs = data_aug(inputs, mu, sigma)
train_err += train_fn(inputs, augs, targets)
train_batches += 1
train_err = train_err / train_batches
if epoch % 1 == 0:
# And a full pass over the validation data:
test_err, test_ham, _, Pr, Or, Tr = val_fn_epoch(
classn, val_fn, X_test, a_test, y_test)
tpos, tneg, fpos, fneg = confusion_matrix(Pr, Tr)
val_auc = auc_roc(Or, Tr)
# Then we print the results for this epoch:
print(
"{:.4f}\t\t{:.4f}\t\t{:.4f}\t\t{:.4f}\t\t{}/{}/{}/{}\t\t{}/{}\t\t{:.3f}"
.format(train_err, test_err, test_ham, val_auc, tpos, tneg,
fpos, fneg, epoch + 1, num_epochs,
time.time() - start_time))
start_time = time.time()
if epoch % 10 == 0:
param_values = layers.get_all_param_values(network)
pickle.dump(
param_values,
open(model_dump + "_e{}_cv{}.pkl".format(epoch, valid_num),
'w'))
if epoch == 40:
LearningRate.set_value(np.float32(0.10 * LearningRate.get_value()))
if epoch == 100:
LearningRate.set_value(np.float32(0.10 * LearningRate.get_value()))
if epoch == 200:
LearningRate.set_value(np.float32(0.10 * LearningRate.get_value()))
# Return a new set of features.
_, _, _, _, train_Or, _ = val_fn_epoch(classn, val_fn, X_train, a_train,
y_train)
_, _, _, _, test_Or, _ = val_fn_epoch(classn, val_fn, X_test, a_test,
y_test)
return train_Or, test_Or
def train_round(num_epochs, network, valid_num, train_fn, val_fn, classn,
X_train, a_train, y_train, X_test, a_test, y_test):
print("Starting training...")
print("TrLoss\tVaLoss\tAUC\tCMatrix0\tCMatrix1\tCMatrix2\tEpochs\tTime")
start_time = time.time()
for epoch in range(num_epochs + 1):
train_err = 0
train_batches = 0
for batch in iterate_minibatches(X_train,
a_train,
y_train,
BatchSize,
shuffle=True):
inputs, augs, targets = batch
inputs = data_aug(inputs, mu, sigma)
train_err += train_fn(inputs, augs, targets)
train_batches += 1
train_err = train_err / train_batches
if epoch % 1 == 0:
# And a full pass over the validation data:
test_err, _, _, _, Or, Tr = val_fn_epoch(classn, val_fn, X_test,
a_test, y_test)
tpos0, tneg0, fpos0, fneg0 = confusion_matrix(Or, Tr, 0.4)
tpos1, tneg1, fpos1, fneg1 = confusion_matrix(Or, Tr, 0.5)
tpos2, tneg2, fpos2, fneg2 = confusion_matrix(Or, Tr, 0.6)
val_auc = auc_roc(Or, Tr)
# Then we print the results for this epoch:
print(
"{:.4f}\t{:.4f}\t{:.4f}\t{}/{}/{}/{}\t{}/{}/{}/{}\t{}/{}/{}/{}\t{}/{}\t{:.3f}"
.format(train_err, test_err, val_auc, tpos0, tneg0, fpos0,
fneg0, tpos1, tneg1, fpos1, fneg1, tpos2, tneg2, fpos2,
fneg2, epoch + 1, num_epochs,
time.time() - start_time))
start_time = time.time()
if epoch % 10 == 0:
param_values = layers.get_all_param_values(network)
pickle.dump(
param_values,
open(model_dump + "_e{}_cv{}.pkl".format(epoch, valid_num),
'w'))
if epoch == 10:
LearningRate.set_value(np.float32(0.33 * LearningRate.get_value()))
if epoch == 20:
LearningRate.set_value(np.float32(0.33 * LearningRate.get_value()))
if epoch == 30:
LearningRate.set_value(np.float32(0.33 * LearningRate.get_value()))
def multi_win_during_val(val_fn, inputs, augs, targets):
for idraw in [
-1,
]:
for jdraw in [
-1,
]:
inpt_multiwin = data_aug(inputs, mu, sigma)
err_pat, output_pat = val_fn(inpt_multiwin, augs, targets)
if 'weight' in locals():
weight += 1.0
err += err_pat
output += output_pat
else:
weight = 1.0
err = err_pat
output = output_pat
return err / weight, output / weight
