def eval(self, batch_size, out_name, mini_data=True):
val_dataset = DataGen(self.IMG_SIZE, 5, False, mini_data)
val_dataset.generator(batch_size)
print("val data size: ", val_dataset.get_dataset_size())
Y_pred = []
X = []
Y_gt = []
count = 0
for X_batch, y_batch in val_dataset.generator(batch_size):
count += batch_size
print("count:", count)
if count > val_dataset.get_dataset_size():
break
y_pred = self.model.predict(X_batch)
y_pred = np.argmax(y_pred, -1)
y_gt = np.argmax(y_batch, -1)
Y_pred = np.concatenate([Y_pred, y_pred])
Y_gt = np.concatenate([Y_gt, y_gt])
acc = accuracy_score(Y_gt, Y_pred)
print('Eval Accuracy: %2.2f%%' % acc)
# sns.heatmap(confusion_matrix(Y_gt, Y_pred),
# annot=True, fmt="d", cbar=False, cmap=plt.cm.Blues, vmax=Y_pred.shape[0] // 16)
# plt.show()
np_confusion = confusion_matrix(Y_gt, Y_pred)
np.save('confusion_' + str(out_name) + '.npy', np_confusion)
def train(self, batch_size, epoches, out_name, mini_data):
print(mini_data)
# learning rate schedule
def step_decay(epoch):
initial_lrate = 1e-3
drop = 0.5
epochs_drop = 7.0
lrate = initial_lrate * math.pow(
drop, math.floor((1 + epoch) / epochs_drop))
return lrate
train_dataset = DataGen(self.IMG_SIZE, 5, True, mini=mini_data)
train_gen = train_dataset.generator(batch_size, True)
TIMESTAMP = "{0:%Y-%m-%dT%H-%M-%S}".format(datetime.datetime.now())
callbacks_list = [
EvalCallBack(out_name),
EarlyStopping(monitor='val_loss', mode='min', patience=6),
TensorBoard(log_dir='logs/' + TIMESTAMP,
batch_size=batch_size,
update_freq='epoch'),
LearningRateScheduler(step_decay)
]
self.model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['categorical_accuracy'])
self.model.fit_generator(
generator=train_gen,
steps_per_epoch=train_dataset.get_dataset_size() // batch_size,
epochs=epoches,
callbacks=callbacks_list)
def run_eval(self, epoch, batch_size=16):
val_dataset = DataGen(self.IMG_SIZE, 5, False)
val_gen = val_dataset.generator(batch_size)
count = 0
Y_pred = []
Y_gt = []
for X_batch, y_batch in val_gen:
count += batch_size
if count > val_dataset.get_dataset_size():
break
y_pred = self.model.predict(X_batch)
y_pred = np.argmax(y_pred, -1)
y_gt = np.argmax(y_batch, -1)
Y_pred = np.concatenate([Y_pred, y_pred])
Y_gt = np.concatenate([Y_gt, y_gt])
acc = accuracy_score(Y_gt, Y_pred)
print('Eval Accuracy: %2.2f%%' % acc, '@ Epoch ', epoch)
if (epoch+1) % 10 == 0:
print(classification_report(Y_gt, Y_pred))
with open('checkpoints/'+self.out_name+'/val.txt', 'a+') as xfile:
xfile.write('Epoch ' + str(epoch) + ':' + str(acc) + '\n')
def resume_train(self,
batch_size,
model_json,
model_weights,
init_epoch,
epochs,
out_name,
mini_data=True):
self.load_model(model_json, model_weights)
self.model.compile(optimizer=Adam(lr=5e-4),
loss='categorical_crossentropy',
metrics=["categorical_accuracy"])
train_dataset = DataGen(self.IMG_SIZE,
5,
is_train=True,
mini=mini_data)
train_gen = train_dataset.generator(batch_size, True)
model_dir = os.path.dirname(os.path.abspath(model_json))
print(model_dir, model_json)
TIMESTAMP = "{0:%Y-%m-%dT%H-%M-%S}".format(datetime.datetime.now())
callbacks_list = [
EvalCallBack(out_name),
EarlyStopping(monitor='val_loss', mode='min', patience=6),
TensorBoard(log_dir='logs/' + TIMESTAMP,
batch_size=batch_size,
update_freq='epoch')
]
self.model.fit_generator(
generator=train_gen,
steps_per_epoch=train_dataset.get_dataset_size() // batch_size,
initial_epoch=init_epoch,
epochs=epochs,
callbacks=callbacks_list)