def train_msml(feat_model,
model,
train_dataset,
modelpath,
val_dataset=None,
train_flag=True,
aug=True,
models_folder='saved_models',
epochs=120,
batch_size=4,
img_size=(224, 320),
label_smoothing=False,
wlr=False,
BN=False,
CenterLoss=False):
losses_list = [msml(), 'categorical_crossentropy']
if CenterLoss:
print('Using CenterLoss')
losses_list.append(lambda y_true, y_pred: y_pred)
losses_list.append(lambda y_true, y_pred: y_pred)
model.compile(
optimizer=standard_optimizer(),
loss=losses_list,
)
train_dataset = Dataset(train_dataset, to_rgb=False)
modelpath = os.path.join(models_folder, modelpath)
if (os.path.exists(modelpath)):
print('MSML carregada.')
model.load_weights(modelpath)
if train_flag:
checkpoint = ModelCheckpoint(modelpath,
monitor='loss',
verbose=1,
save_best_only=True,
mode='min')
if wlr:
print('Using warmup lr')
lrate = LearningRateScheduler(warmup_lr)
else:
lrate = LearningRateScheduler(step_decay)
callbacks_list = [lrate, checkpoint]
if not CenterLoss:
msml_gen = general_generator(feat_model,
train_dataset,
batch_size=batch_size,
img_size=img_size,
aug=aug,
label_smoothing=label_smoothing)
else:
msml_gen = general_generator_center(
feat_model,
train_dataset,
batch_size=batch_size,
img_size=img_size,
aug=aug,
label_smoothing=label_smoothing)
H = model.fit_generator(msml_gen,
steps_per_epoch=int(train_dataset.ident_num() /
batch_size),
epochs=epochs,
callbacks=callbacks_list,
workers=1)
models_folder = ''
MoRe_path = "MoRe.hdf5"
MoRe = Dataset(MoRe_path, to_bgr=True) # Images were used in BGR format
INPUT_SHAPE = (256, 256)
feat_model = ResNet50_LastStride(input_shape=(INPUT_SHAPE[0], INPUT_SHAPE[1],
3)) # BASELINE STRIDE = 1
feat_model.load_weights(os.path.join(models_folder,
'RESNET50_ORIGINAL_WEIGHTS.h5'),
by_name=True)
feat_model._make_predict_function()
identity_model = classification_layer_baseline(MoRe.ident_num('train'),
BN=True) # Loading BN
clsNet = classification_net(feat_model, identity_model, img_shape=INPUT_SHAPE)
# train_classifier(
# feat_model = feat_model,
# model = clsNet,
# dataset = MoRe,
# modelpath = 'tsting_cls.hdf5',
# models_folder = models_folder,
# epochs = 1,
# batch_size = 32,
# img_size = INPUT_SHAPE,
# label_smoothing = True,
# wlr = True,
# BN = True
