def get_graph_init(train=False):
g = Graph('YOLOv2-Test: 16-bit',
dataset='imagenet',
log_level=logging.INFO)
batch_size = 1
with g.as_default():
with g.name_scope('inputs'):
i = get_tensor(shape=(batch_size, 416, 416, 3),
name='data',
dtype=FQDtype.FXP16,
trainable=False)
with g.name_scope('conv0'):
conv0 = yolo_convolution(i,
filters=16,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 9),
bn_dtype=FixedPoint(16, 8))
with g.name_scope('pool0'):
pool0 = maxPool(conv0,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv1'):
conv1 = yolo_convolution(pool0,
filters=32,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 8),
s_dtype=FixedPoint(16, 14),
bn_dtype=FixedPoint(16, 8))
with g.name_scope('pool1'):
pool1 = maxPool(conv1,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv2'):
conv2 = yolo_convolution(
pool1,
filters=64,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
# batch_normalize=False, act='linear',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 10),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 9))
with g.name_scope('pool2'):
pool2 = maxPool(conv2,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv3'):
conv3 = yolo_convolution(pool2,
filters=128,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 10),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 10))
with g.name_scope('pool3'):
pool3 = maxPool(conv3,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv4'):
conv4 = yolo_convolution(pool3,
filters=256,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 10))
with g.name_scope('pool4'):
pool4 = maxPool(conv4,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 2, 2, 1),
pad='VALID')
with g.name_scope('conv5'):
conv5 = yolo_convolution(pool4,
filters=512,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 13),
bn_dtype=FixedPoint(16, 11))
with g.name_scope('pool5'):
pool5 = maxPool(conv5,
pooling_kernel=(1, 2, 2, 1),
stride=(1, 1, 1, 1),
pad=((0, 0), (0, 1), (0, 1), (0, 0)))
with g.name_scope('conv6'):
conv6 = yolo_convolution(pool5,
filters=1024,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 12),
s_dtype=FixedPoint(16, 11),
bn_dtype=FixedPoint(16, 9))
with g.name_scope('conv7'):
conv7 = yolo_convolution(conv6,
filters=1024,
kernel_size=3,
batch_normalize=True,
act='leakyReLU',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11),
s_dtype=FixedPoint(16, 14),
bn_dtype=FixedPoint(16, 12))
with g.name_scope('conv8'):
conv8 = yolo_convolution(conv7,
filters=125,
kernel_size=1,
batch_normalize=False,
act='linear',
w_dtype=FixedPoint(16, 14),
c_dtype=FixedPoint(16, 11))
return g
def _get_output_dtype(self):
return FixedPoint(32, self.data.dtype.frac_bits + self.scale.dtype.frac_bits)
def dnnweaver_init_weight(g, scope, *args, **kwargs):
fname = "cmstack.codegen.dnnweavergen.dnnweaver2.get_tensor"
with g.as_default():
with g.name_scope(scope):
return get_func(fname)(*args, dtype=FixedPoint(16, 14), **kwargs)