def __init__(self, test_name):
"""
Init some specify data that is different among sub test classes.
Should be called first by super().__init__ in all sub test classes.
:param test_name: name of test case.
"""
self.test_result = TestResult(test_name)
self.steps = Steps()
self.logger = Logger(test_name)
def init_data_test(self):
"""
Init test data.
If the test case need some extra test date then just override this method.
"""
test_name = os.path.splitext(os.path.basename(inspect.getfile(self.__class__)))[0]
self.test_result = TestResult(test_name)
self.steps = Steps()
self.logger = Logger(test_name)
def main():
setpu_seed(2021)
args = parse_args()
save_path = os.path.join(args.outf, args.save)
device = torch.device(
"cuda" if torch.cuda.is_available() and args.cuda else "cpu")
cudnn.benchmark = True
log = Logger(save_path)
sys.stdout = Print_Logger(os.path.join(save_path, 'train_log.txt'))
print('The computing device used is: ',
'GPU' if device.type == 'cuda' else 'CPU')
# net = models.UNetFamily.U_Net(1,2).to(device)
net = models.LadderNet(inplanes=1, num_classes=2, layers=3,
filters=16).to(device)
print("Total number of parameters: " + str(count_parameters(net)))
log.save_graph(
net,
torch.randn((1, 1, 64, 64)).to(device).to(
device=device)) # Save the model structure to the tensorboard file
# torch.nn.init.kaiming_normal(net, mode='fan_out') # Modify default initialization method
# net.apply(weight_init)
# The training speed of this task is fast, so pre training is not recommended
if args.pre_trained is not None:
# Load checkpoint.
print('==> Resuming from checkpoint..')
checkpoint = torch.load(args.outf +
'%s/latest_model.pth' % args.pre_trained)
net.load_state_dict(checkpoint['net'])
optimizer.load_state_dict(checkpoint['optimizer'])
args.start_epoch = checkpoint['epoch'] + 1
# criterion = LossMulti(jaccard_weight=0,class_weights=np.array([0.5,0.5]))
criterion = CrossEntropyLoss2d() # Initialize loss function
# create a list of learning rate with epochs
# lr_epoch = np.array([50, args.N_epochs])
# lr_value = np.array([0.001, 0.0001])
# lr_schedule = make_lr_schedule(lr_epoch,lr_value)
# lr_scheduler = optim.lr_scheduler.StepLR(optimizer,step_size=10,gamma=0.5)
# optimizer = optim.SGD(net.parameters(),lr=lr_schedule[0], momentum=0.9, weight_decay=5e-4, nesterov=True)
optimizer = optim.Adam(net.parameters(), lr=args.lr)
lr_scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer,
T_max=args.N_epochs,
eta_min=0)
train_loader, val_loader = get_dataloader(args) # create dataloader
if args.val_on_test:
print(
'\033[0;32m===============Validation on Testset!!!===============\033[0m'
)
val_tool = Test(args)
best = {
'epoch': 0,
'AUC_roc': 0.5
} # Initialize the best epoch and performance(AUC of ROC)
trigger = 0 # Early stop Counter
for epoch in range(args.start_epoch, args.N_epochs + 1):
print('\nEPOCH: %d/%d --(learn_rate:%.6f) | Time: %s' % \
(epoch, args.N_epochs,optimizer.state_dict()['param_groups'][0]['lr'], time.asctime()))
# train stage
train_log = train(train_loader, net, criterion, optimizer, device)
# val stage
if not args.val_on_test:
val_log = val(val_loader, net, criterion, device)
else:
val_tool.inference(net)
val_log = val_tool.val()
log.update(epoch, train_log, val_log) # Add log information
lr_scheduler.step()
# Save checkpoint of latest and best model.
state = {
'net': net.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch
}
torch.save(state, os.path.join(save_path, 'latest_model.pth'))
trigger += 1
if val_log['val_auc_roc'] > best['AUC_roc']:
print('\033[0;33mSaving best model!\033[0m')
torch.save(state, os.path.join(save_path, 'best_model.pth'))
best['epoch'] = epoch
best['AUC_roc'] = val_log['val_auc_roc']
trigger = 0
print('Best performance at Epoch: {} | AUC_roc: {}'.format(
best['epoch'], best['AUC_roc']))
# early stopping
if not args.early_stop is None:
if trigger >= args.early_stop:
print("=> early stopping")
break
torch.cuda.empty_cache()
def gpu_memory_in_use(gpu='/device:GPU:0', log=None):
from tensorflow.contrib.memory_stats.python.ops.memory_stats_ops import BytesInUse
with tf.device(gpu):
bytes_in_use = BytesInUse()
with tf.Session() as sess:
val = sess.run(bytes_in_use)
if log is not None:
log.P(" GPU free mem: {:.1f} Gb".format(val / (1024**3)))
return val
if __name__ == '__main__':
cfg = 'config/duplex_config.txt'
log = Logger(lib_name='MGS', config_file=cfg, TF_KERAS=True)
yolo = log.LoadGraphFromModels('01_1712_y_720_1280_c.pb')
face = log.LoadGraphFromModels('20_190301_mob_ssd_faces.pb')
config_proto = tf.compat.v1.ConfigProto()
config_proto.gpu_options.allow_growth = True
yolo_sess = tf.compat.v1.Session(graph=yolo, config=config_proto)
log.P("Created yolo session")
config_proto = tf.compat.v1.ConfigProto()
config_proto.gpu_options.allow_growth = True
face_sess = tf.compat.v1.Session(graph=face, config=config_proto)
log.P("Created ssd session")
tf_face_classes = face_sess.graph.get_tensor_by_name("detection_classes:0")
log.P("*"*50)
log.P("LocalConfig v{} received config data:".format(__VER__))
dct_prev = log.load_data_json(fn_config)
log.P(" Current: {}".format(dct_prev))
log.save_data_json(dct_config, fn_config )
dct_new = log.load_data_json(fn_config)
log.P(" New: {}".format(dct_config))
jresponse = flask.jsonify({
"RECEIVED_CONFIG_UPDATE": dct_new})
jresponse.headers["Access-Control-Allow-Origin"] = "*"
jresponse.headers["Access-Control-Allow-Methods"] = "POST, GET, OPTIONS, DELETE"
jresponse.headers["Access-Control-Allow-Headers"] = "Content-Type"
log.P("*"*50)
return jresponse
if __name__ == '__main__':
config_file = 'config/duplex_config.txt'
log = Logger(lib_name='ODJcfg', config_file=config_file, TF_KERAS=False)
log.P("Starting OmniDJ local config server {}".format(__VER__))
dct_viewer = log.load_data_json(fn_config)
log.P("Currend config:\n{}".format(dct_viewer))
app = flask.Flask('LocalConfigServer')
app.add_url_rule(rule='/config',
endpoint="LocalConfig",
view_func=config_update,
methods = ['GET', 'POST','OPTIONS']
)
app.run(host='127.0.0.1', port=5500)
"""
import tensorflow as tf
from libraries.logger import Logger
def combine_graphs(graph1, graph2, name1='g1', name2='g2'):
gdef_1 = graph1.as_graph_def()
gdef_2 = graph2.as_graph_def()
with tf.Graph().as_default() as g_combined:
tf.import_graph_def(graph_def=gdef_1, name=name1)
tf.import_graph_def(graph_def=gdef_2, name=name2)
return g_combined
if __name__ == '__main__':
log = Logger(lib_name='GC', config_file='config/duplex_config.txt')
g1 = log.LoadGraphFromModels('01_1712_y_720_1280_c.pb')
n1 = 'yolo'
g2 = log.LoadGraphFromModels('20_190301_mob_ssd_faces.pb')
n2 = 'face'
gc = combine_graphs(g1, g2, name1=n1, name2=n2)
config_proto = tf.compat.v1.ConfigProto()
config_proto.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(graph=gc, config=config_proto)
earning_phase = sess.graph.get_tensor_by_name(n1 +
"/keras_learning_phase:0")
tf_classes = sess.graph.get_tensor_by_name(n1 + "/YOLO_OUTPUT_CLASSES" +
":0")
tf_scores = sess.graph.get_tensor_by_name(n1 + "/YOLO_OUTPUT_SCORES" +
