batch_size = 10
base_lr = 4e-5 # 2e-5
momentum = 0.9
weight_decay = 5e-4
lr_policy = "step"
gamma = 0.333
stepsize = 136106 #68053 // after each stepsize iterations update learning rate: lr=lr*gamma
max_iter = 600000
lr_mult_distro = [1.0, 1.0, 4.0, 1]
di = DataIterator("../dataset/val_dataset.h5",
data_shape=(3, 368, 368),
label_shape=(57, 46, 46),
split_point=38,
batch_size=batch_size,
shuffle=True)
x, y1, y2 = di.next()
print("x : ", x.shape)
print("y1 : ", y1.shape)
print("y2 : ", y2.shape)
# as suggested in: https://github.com/fchollet/keras/issues/5920
last_layer_variables = list()
for layer in model.layers:
#print(layer.weights)
if layer.name in ['prediction']:
train_di = train_client.gen()
train_samples = 52597
val_client = DataGeneratorClient(port=pargs.port + 1,
host="localhost",
hwm=160,
batch_size=batch_size,
pstages=paramNumStages)
val_client.start()
val_di = val_client.gen()
val_samples = 2645
else:
train_di = DataIterator("../dataset/train_dataset.h5",
data_shape=(3, 368, 368),
mask_shape=(1, 46, 46),
label_shape=(57, 46, 46),
vec_num=38,
heat_num=19,
batch_size=batch_size,
shuffle=True)
train_samples = train_di.N
val_di = DataIterator("../dataset/val_dataset.h5",
data_shape=(3, 368, 368),
mask_shape=(1, 46, 46),
label_shape=(57, 46, 46),
vec_num=38,
heat_num=19,
batch_size=batch_size,
shuffle=True)
val_samples = val_di.N
# setup lr multipliers for conv layers