def load_models(self):
cfg = self.cfg
if self.do_rpn:
print('Loading RPN model...')
loss_cls = losses.rpn_loss_cls(cfg.nb_anchors)
loss_rgr = losses.rpn_loss_regr(cfg.nb_anchors)
get_custom_objects().update({"rpn_loss_cls_fixed_num": loss_cls})
get_custom_objects().update({"rpn_loss_regr_fixed_num": loss_rgr})
assert os.path.exists(cfg.my_shared_weights
), 'invalid path: %s' % cfg.my_shared_weights
self.model_rpn = load_model(cfg.my_shared_weights)
self.model_rpn.name = 'rpn'
assert not self.model_rpn.layers[17].get_config()['trainable']
if self.do_det:
print('Loading DET model...')
loss_cls = losses.class_loss_cls
loss_rgr = losses.class_loss_regr(cfg.nb_object_classes - 1)
get_custom_objects().update({"class_loss_cls": loss_cls})
get_custom_objects().update(
{"class_loss_regr_fixed_num": loss_rgr})
get_custom_objects().update(
{"RoiPoolingConv": layers.RoiPoolingConv})
assert os.path.exists(cfg.my_shared_weights
), 'invalid path: %s' % cfg.my_shared_weights
self.model_det = load_model(cfg.my_shared_weights)
self.model_det.name = 'det'
if self.do_hoi:
print('Loading HOI model...')
loss_cls = losses.hoi_loss_cls
get_custom_objects().update({"hoi_loss_cls_fixed_num": loss_cls})
assert os.path.exists(cfg.my_shared_weights
), 'invalid path: %s' % cfg.my_shared_weights
self.model_hoi = load_model(cfg.my_shared_weights)
self.model_hoi.name = 'hoi'
def compile_models(self):
if self.mode == 'test':
return
cfg = self.cfg
print('Compiling models...')
if cfg.optimizer == 'adam':
print(' Opt.:', 'Adam')
opt = Adam(lr=cfg.init_lr)
else:
print(' Opt.:', 'SGD')
opt = SGD(lr=cfg.init_lr, momentum=0.9, decay=0.0, nesterov=False)
print(' Learning rate:', cfg.init_lr)
if self.do_rpn:
if cfg.rpn_uniform_sampling:
print(' Uniform anchor sampling')
else:
print(' Non-Uniform anchor sampling')
model = self.model_rpn
my_losses = [
losses.rpn_loss_cls(cfg.nb_anchors),
losses.rpn_loss_regr(cfg.nb_anchors)
]
my_metrics = None
if self.do_det:
model = self.model_det
my_losses = [
losses.class_loss_cls,
losses.class_loss_regr(cfg.nb_object_classes - 1)
]
my_metrics = {'det_out_class': 'categorical_accuracy'}
if self.do_hoi:
model = self.model_hoi
loss = 'categorical_crossentropy' if cfg.do_categorical_hoi else losses.hoi_loss_cls(
cfg.wp)
print(' -HOI loss', loss)
my_losses = [loss]
my_metrics = ['categorical_accuracy'
] if cfg.do_categorical_hoi else None
model.compile(optimizer=opt, loss=my_losses, metrics=my_metrics)
except:
print(
'Could not load pretrained model weights. Weights can be found in the keras application folder \
https://github.com/fchollet/keras/tree/master/keras/applications')
optimizer = Adam(lr=1e-5)
optimizer_classifier = Adam(lr=1e-5)
model_rpn.compile(
optimizer=optimizer,
loss=[losses.rpn_loss_cls(num_anchors),
losses.rpn_loss_regr(num_anchors)])
model_classifier.compile(
optimizer=optimizer_classifier,
loss=[
losses.class_loss_cls,
losses.class_loss_regr(len(classes_count) - 1)
],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 1000
num_epochs = int(options.num_epochs)
iter_num = 0
losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()
best_loss = np.Inf
model_classifier.load_weights(C.model_path, by_name=True)
# this is a model that holds both the RPN and the classifier, used to load/save weights for the models
model_all = Model([img_input, roi_input], rpn[:2] + classifier)
# print("MODEL RPN")
# model_rpn.summary()
# print("MODEL CLASSIFIER")
# model_classifier.summary()
# print("MODEL ALL")
# model_all.summary()
# Define
optimizer_rpn = optimizers.Adam(lr=1e-5)
optimizer_classifier = optimizers.Adam(lr=1e-5)
model_rpn.compile(optimizer=optimizer_rpn, loss=[losses.rpn_loss_cls(num_anchors), losses.rpn_loss_regr(num_anchors)])
model_classifier.compile(optimizer=optimizer_classifier,
loss=[losses.class_loss_cls, losses.class_loss_regr(len(classes_count) - 1)],
metrics={'dense_class_{}'.format(len(classes_count)): 'accuracy'})
model_all.compile(optimizer='sgd', loss='mae')
epoch_length = 1000
num_epochs = int(options.num_epochs)
iter_num = 0
losses = np.zeros((epoch_length, 5))
rpn_accuracy_rpn_monitor = []
rpn_accuracy_for_epoch = []
start_time = time.time()
best_loss = np.Inf
class_mapping_inv = {v: k for k, v in class_mapping.items()}
#r_loss_rpn_regr2 = record2_df['loss_rpn_regr']
#r_loss_class_cls2 = record2_df['loss_class_cls']
#r_loss_class_regr2 = record2_df['loss_class_regr']
#r_curr_loss2 = record2_df['curr_loss']
#r_elapsed_time2 = record2_df['elapsed_time']
#r_mAP2 = record2_df['mAP']
# print('Already train %dK batches' % (len(record_df)))
# Learning rates for RPN & classifier
optimizer = tf.keras.optimizers.Adam(lr = 1e-5)
optimizer_classifier = tf.keras.optimizers.Adam(lr = 1e-5)
#optimizer_all = tf.keras.optimizers.SGD(lr = 1e-5)
model_rpn.compile(optimizer = optimizer, loss = [losses.rpn_loss_cls(num_anchors), losses.rpn_loss_regr(num_anchors)])
model_classifier.compile(optimizer = optimizer_classifier, loss = [losses.class_loss_cls, losses.class_loss_regr(len(classes_count) - 1)], metrics = {'dense_class_{}'.format(len(classes_count)): 'accuracy'})
#model_all.compile(optimizer = optimizer_all, loss = 'mae')
model_all.compile(optimizer = 'sgd', loss = 'mae')
#total_epochs = len(record_df)
#r_epochs = len(record_df)
#total_epochs2 = len(record2_df)
#r_epochs2 = len(record2_df)
##
##
epoch_length = 1000 # Iterations = (Dataset / Batch-Size = 1)
num_epochs = 65 # Epochs
iter_num = 0 # Index -> iteration
temp_im_test = np.ones((50000, 10, 10))
temp_im_train = np.ones((10000, 10, 10))
temp_bb_train = np.ones((50000, 10, 36))
temp_bb_test = np.ones((10000, 10, 36))
temp_regr_train = np.ones((50000, 2, 2, 36))
temp_regr_test = np.ones((10000, 2, 2, 36))
temp_class_train = np.ones((50000, 2, 2, 9))
temp_class_test = np.ones((10000, 2, 2, 9))
model = ROIModel(10)
model.compile(
optimizer=optimizers.RMSprop(1e-3),
loss={
"output_1": custom_losses.class_loss_cls(),
"output_2": custom_losses.class_loss_regr(9),
"output_3": custom_losses.rpn_loss_regr(num_anchors),
"output_4": custom_losses.rpn_loss_cls(num_anchors),
},
metrics={
"output_1": 'accuracy',
"output_2": 'accuracy',
"output_3": 'accuracy',
"output_4": 'accuracy',
},
)
logdir = "logs/fit/" + datetime.now().strftime("%Y%m%d-%H%M%S")
tensorboard_callback = tf.keras.callbacks.TensorBoard(log_dir=logdir)
# test_image = np.array(train_images[0]).reshape(1,32,32,3)