def test_quantize(net, data, epsilon=EPSILON):
x_list = [
data["input"][0, :, :], data["layer1_bn_out"][0, :, :, :],
data["layer2_pool_out"][0, :, 0, :],
data["layer3_conv_out"][0, :, 0, :], data["layer4_pool_out"][0, :,
0, :]
]
y_exp_list = [
data["input_quant"][0, 0, :, :], data["layer1_activ"][0, :, :, :],
data["layer2_activ"][0, :, 0, :], data["layer3_activ"][0, :, 0, :],
data["layer4_activ"][0, :, 0, :]
]
scale_list = [
convert.ste_quant(net, "quant1"),
convert.ste_quant(net, "quant2"),
convert.ste_quant(net, "quant3"),
convert.ste_quant(net, "quant4"),
convert.ste_quant(net, "quant5")
]
casename_list = ["input", "layer1", "layer2", "layer3", "layer4"]
ret = {}
for casename, x, y_exp, scale in zip(casename_list, x_list, y_exp_list,
scale_list):
y = F.quantize(x, scale)
l1_error = np.abs(y - y_exp).mean()
success = l1_error < epsilon
ret[casename] = {"result": success, "l1_error": l1_error}
return ret
def __init__(self,
net,
T,
F2,
reorder_bn=True,
clip_balanced=True,
**params):
self.name = "Layer 4: Point Convolution + Batch Norm + ReLU + Pooling"
self.T = T
self.F2 = F2
self.input_shape = ((F2, T // 8))
self.output_shape = ((F2, T // 64))
self.clip_balanced = clip_balanced
self.reorder_bn = reorder_bn
# fetch weights
self.weights, self.weight_scale = convert.inq_conv2d(net, "sep_conv2")
assert self.weights.shape == (self.F2, self.F2, 1, 1)
self.weights = np.reshape(self.weights, (self.F2, self.F2))
# fetch batch norm offset and scale
self.input_scale = convert.ste_quant(net, "quant4")
self.output_scale = convert.ste_quant(net, "quant5")
self.bn_scale, self.bn_offset = convert.batch_norm(net, "batch_norm3")
self.factor, self.bias = convert.div_factor_batch_norm(
self.input_scale,
self.weight_scale,
self.output_scale,
self.bn_scale,
self.bn_offset,
pool=8)
def __init__(self, net, T, F2, N, clip_balanced=True, **params):
self.name = "Layer 5: Linear Layer"
self.T = T
self.F2 = F2
self.N = N
self.flatten_dim = self.F2 * (self.T // 64)
self.input_shape = ((F2, T // 64))
self.output_shape = ((N, ))
self.clip_balanced = clip_balanced
# fetch weights
self.weights, self.bias, self.weight_scale = convert.inq_linear(
net, "fc")
assert self.weights.shape == (self.N, self.flatten_dim)
self.input_scale = convert.ste_quant(net, "quant5")
self.output_scale = convert.ste_quant(net, "quant6")
self.factor = convert.div_factor(self.input_scale, self.weight_scale,
self.output_scale)
def __init__(self, net, T, F2, clip_balanced=True, **params):
self.name = "Layer 3: Convolution in Time"
self.T = T
self.F2 = F2
self.input_shape = ((F2, T // 8))
self.output_shape = ((F2, T // 8))
self.clip_balanced = clip_balanced
# fetch weights
self.weights, self.weight_scale = convert.inq_conv2d(net, "sep_conv1")
assert self.weights.shape == (self.F2, 1, 1, 16)
self.weights = np.reshape(self.weights, (self.F2, 16))
# fetch batch norm offset and scale
self.input_scale = convert.ste_quant(net, "quant3")
self.output_scale = convert.ste_quant(net, "quant4")
self.factor = convert.div_factor(self.input_scale, self.weight_scale,
self.output_scale)
def __init__(self, net, C, T, F1, F2, clip_balanced=True, **params):
self.name = "Layer 1: Convolution in Time + Batch Norm"
self.C = C
self.T = T
self.F1 = F1
self.F2 = F2
self.input_shape = ((C, T))
self.output_shape = ((F2, T // 8))
self.clip_balanced = clip_balanced
# fetch weights
self.weights_1, self.weight_scale_1 = convert.inq_conv2d(net, "conv1")
assert self.weights_1.shape == (self.F1, 1, 1, 64)
self.weights_1 = np.reshape(self.weights_1, (self.F1, 64))
self.weights_2, self.weight_scale_2 = convert.inq_conv2d(net, "conv2")
assert self.weights_2.shape == (self.F2, 1, self.C, 1)
self.weights_2 = np.reshape(self.weights_2, (self.F2, self.C))
# fetch batch norm offset and scale
self.input_scale = convert.ste_quant(net, "quant1")
self.intermediate_scale = convert.ste_quant(net, "quant2")
self.output_scale = convert.ste_quant(net, "quant3")
self.bn_scale_1, self.bn_offset_1 = convert.batch_norm(
net, "batch_norm1")
self.bn_scale_2, self.bn_offset_2 = convert.batch_norm(
net, "batch_norm2")
self.factor_1, self.bias_1 = convert.div_factor_batch_norm(
self.input_scale, self.weight_scale_1, self.intermediate_scale,
self.bn_scale_1, self.bn_offset_1)
self.factor_2, self.bias_2 = convert.div_factor_batch_norm(
self.intermediate_scale,
self.weight_scale_2,
self.output_scale,
self.bn_scale_2,
self.bn_offset_2,
pool=8)
# update the factors and scales
for k in range(16):
self.factor_2[k] *= self.factor_1[k // 2]
self.bias_2[k] *= self.factor_1[k // 2]
def __init__(self, net, C, T, F1, clip_balanced=True, **params):
self.name = "Layer 1: Convolution in Time + Batch Norm"
self.C = C
self.T = T
self.F1 = F1
self.input_shape = ((C, T))
self.output_shape = ((F1, C, T))
self.clip_balanced = clip_balanced
# fetch weights
self.weights, self.weight_scale = convert.inq_conv2d(net, "conv1")
assert self.weights.shape == (self.F1, 1, 1, 64)
self.weights = np.reshape(self.weights, (self.F1, 64))
# fetch batch norm offset and scale
self.input_scale = convert.ste_quant(net, "quant1")
self.output_scale = convert.ste_quant(net, "quant2")
self.bn_scale, self.bn_offset = convert.batch_norm(net, "batch_norm1")
self.factor, self.bias = convert.div_factor_batch_norm(
self.input_scale, self.weight_scale, self.output_scale,
self.bn_scale, self.bn_offset)
def __init__(self,
net,
C,
T,
F1,
F2,
reorder_bn=True,
clip_balanced=True,
**params):
self.name = "Layer 2: Convolution in Space + Batch Norm + ReLU + Pooling"
self.C = C
self.T = T
self.F1 = F1
self.F2 = F2
self.input_shape = ((F1, C, T))
self.output_shape = ((F2, T // 8))
self.clip_balanced = clip_balanced
self.reorder_bn = reorder_bn
# fetch weights
self.weights, self.weight_scale = convert.inq_conv2d(net, "conv2")
self.float_weights = np.reshape(net["conv2.weightFrozen"],
(self.F2, self.C))
self.float_weights = np.flip(self.float_weights, (-1))
assert self.weights.shape == (self.F2, 1, self.C, 1)
self.weights = np.reshape(self.weights, (self.F2, self.C))
# fetch batch norm offset and scale
self.input_scale = convert.ste_quant(net, "quant2")
self.output_scale = convert.ste_quant(net, "quant3")
self.bn_scale, self.bn_offset = convert.batch_norm(net, "batch_norm2")
self.factor, self.bias = convert.div_factor_batch_norm(
self.input_scale,
self.weight_scale,
self.output_scale,
self.bn_scale,
self.bn_offset,
pool=8)
def gen_input_header(net, net_params, data, output_file):
# only allow nets with 255 levels
assert net_params["weightInqNumLevels"] == 255
assert net_params["actSTENumLevels"] == 255
# extract and prepare the input data
scale_factor = convert.ste_quant(net, "quant1")
input_quant = F.quantize_to_int(data, scale_factor)
input_quant_align = align_array(input_quant)
# also generate the padded input vector
_, C, T = input_quant.shape
T_pad = T + 63
assert T_pad % 4 == 0
input_pad = np.zeros((C, T_pad), dtype=np.int)
input_pad[:, 31:31 + T] = input_quant[0]
# generate the header file
header = HeaderFile(output_file, "__INPUT_H__", with_c=True)
header.add(HeaderArray("input_data", "int8_t", input_quant_align.ravel()))
header.add(HeaderArray("input_data_pad", "int8_t", input_pad.ravel()))
header.write()
def gen_net_header(net_file, config_file, output_file):
# load network
net = np.load(net_file)
# load configuration file
with open(config_file, "r") as _f:
config = json.load(_f)
# we only need the network parameters
net_params = config["indiv"]["net"]["params"]
# only allow nets with 255 levels
assert net_params["weightInqNumLevels"] == 255
assert net_params["actSTENumLevels"] == 255
assert net_params["F2"] % 4 == 0
assert net_params["N"] == 4
# prepare params
if net_params["F2"] is None:
net_params["F2"] = net_params["F1"] * net_params["D"]
# only allow F2 = F1 * D
assert net_params["F2"] == net_params["F1"] * net_params["D"]
# start the header file
header = HeaderFile(output_file, "__NET_NET_H__", with_c=True)
# add network dimensions
header.add(HeaderComment("Network Dimensions", blank_line=False))
header.add(HeaderConstant("NET_F1", net_params["F1"], blank_line=False))
header.add(HeaderConstant("NET_F2", net_params["F2"], blank_line=False))
header.add(HeaderConstant("NET_D", net_params["D"], blank_line=False))
header.add(HeaderConstant("NET_C", net_params["C"], blank_line=False))
header.add(
HeaderConstant("NET_C_ALIGN",
align_array_size(net_params["C"]),
blank_line=False))
header.add(HeaderConstant("NET_T", net_params["T"], blank_line=False))
header.add(
HeaderConstant("NET_T_ALIGN",
align_array_size(net_params["T"]),
blank_line=False))
header.add(HeaderConstant("NET_T8", net_params["T"] // 8,
blank_line=False))
header.add(
HeaderConstant("NET_T8_ALIGN",
align_array_size(net_params["T"] // 8),
blank_line=False))
header.add(
HeaderConstant("NET_T64", (net_params["T"] // 8) // 8,
blank_line=False))
header.add(
HeaderConstant("NET_T64_ALIGN",
align_array_size((net_params["T"] // 8) // 8),
blank_line=False))
header.add(HeaderConstant("NET_N", net_params["N"], blank_line=True))
# Layer 1
input_scale = convert.ste_quant(net, "quant1")
weight, weight_scale = convert.inq_conv2d(net, "conv1")
weight = weight.reshape(net_params["F1"], 64)
weight_reverse, _ = convert.inq_conv2d(net, "conv1", store_reversed=True)
weight_reverse = weight_reverse.reshape(net_params["F1"], 64)
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm1")
output_scale = convert.ste_quant(net, "quant2")
factor, offset = convert.div_factor_batch_norm(input_scale, weight_scale,
output_scale, bn_scale,
bn_offset)
# add padding to the weight vector of 4
if WEIGHT_L1_PAD > 0:
weight_reverse_pad = np.zeros((net_params["F1"], 64 + WEIGHT_L1_PAD))
weight_reverse_pad[:, :-WEIGHT_L1_PAD] = weight_reverse
else:
weight_reverse_pad = weight_reverse
header.add(
HeaderComment(
"Layer 1\n"
"=======\n"
"Convolution + BN\n\n"
"Input: [C, T]\n"
"Weight: [F1, 64]\n"
"Output: [F1, C, T]",
mode="/*"))
header.add(HeaderConstant("NET_L1_PAD_START", 31))
header.add(HeaderConstant("NET_L1_PAD_END", 32))
header.add(
HeaderConstant("NET_L1_PAD_INPUT_LEN", net_params["T"] + 31 + 32))
header.add(
HeaderConstant("NET_L1_PAD_INPUT_LEN_ALIGN",
align_array_size(net_params["T"] + 31 + 32)))
header.add(HeaderArray("net_l1_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l1_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L1_WEIGHT_LEN", weight.shape[-1]))
header.add(
HeaderConstant("NET_L1_WEIGHT_LEN_ALIGN",
weight_reverse_pad.shape[-1]))
header.add(HeaderArray("net_l1_weight", "int8_t", weight.ravel()))
header.add(
HeaderArray("net_l1_weight_reverse", "int8_t", weight_reverse.ravel()))
header.add(
HeaderArray("net_l1_weight_reverse_pad", "int8_t",
weight_reverse_pad.ravel()))
# layer2
input_scale = convert.ste_quant(net, "quant2")
weight, weight_scale = convert.inq_conv2d(net,
"conv2",
store_reversed=True)
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm2")
output_scale = convert.ste_quant(net, "quant3")
factor, offset = convert.div_factor_batch_norm(input_scale,
weight_scale,
output_scale,
bn_scale,
bn_offset,
pool=8)
weight = weight.reshape(net_params["F2"], net_params["C"])
weight = align_array(weight)
header.add(
HeaderComment(
"Layer 2\n"
"=======\n"
"Convolution + BN + ReLU + Pooling\n\n"
"Input: [F1, C, T]\n"
"Weight: [F2, C] (aligned to [F2, 24]\n"
"Output: [F2, T // 8]",
mode="/*"))
header.add(HeaderArray("net_l2_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l2_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L2_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l2_weight", "int8_t", weight.ravel()))
header.add(HeaderArray("net_l2_weight_32", "int32_t", weight.ravel()))
# layer3
input_scale = convert.ste_quant(net, "quant3")
weight, weight_scale = convert.inq_conv2d(net, "sep_conv1")
output_scale = convert.ste_quant(net, "quant4")
factor = convert.div_factor(input_scale, weight_scale, output_scale)
weight = weight.reshape(net_params["F2"], 16)
header.add(
HeaderComment(
"Layer 3\n"
"=======\n"
"Convolution\n\n"
"Input: [F2, T // 8]\n"
"Weight: [F2, 16]\n"
"Output: [F2, T // 8]",
mode="/*",
blank_line=False))
header.add(HeaderConstant("NET_L3_PAD_START", 7))
header.add(HeaderConstant("NET_L3_PAD_END", 8))
header.add(
HeaderConstant("NET_L3_PAD_INPUT_LEN", net_params["T"] // 8 + 7 + 8))
header.add(
HeaderConstant("NET_L3_PAD_INPUT_LEN_ALIGN",
align_array_size(net_params["T"] // 8 + 7 + 8)))
header.add(HeaderConstant("NET_L3_FACTOR", factor))
header.add(HeaderConstant("NET_L3_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l3_weight", "int8_t", weight.ravel()))
# layer4
input_scale = convert.ste_quant(net, "quant4")
weight, weight_scale = convert.inq_conv2d(net, "sep_conv2")
output_scale = convert.ste_quant(net, "quant5")
bn_scale, bn_offset = convert.batch_norm(net, "batch_norm3")
factor, offset = convert.div_factor_batch_norm(input_scale,
weight_scale,
output_scale,
bn_scale,
bn_offset,
pool=8)
weight = weight.reshape(net_params["F2"], net_params["F2"])
header.add(
HeaderComment(
"Layer 4\n"
"=======\n"
"Convolution + BN + ReLU + Pooling\n\n"
"Input: [F2, T // 8]\n"
"Weight: [F2, F2]\n"
"Output: [F2, T // 64]",
mode="/*"))
header.add(HeaderArray("net_l4_factor", "int32_t", factor.ravel()))
header.add(HeaderArray("net_l4_offset", "int32_t", offset.ravel()))
header.add(HeaderConstant("NET_L4_WEIGHT_LEN", weight.shape[-1]))
header.add(HeaderArray("net_l4_weight", "int8_t", weight.ravel()))
# layer5
input_scale = convert.ste_quant(net, "quant5")
output_scale = convert.ste_quant(net, "quant6")
weight, bias, weight_scale = convert.inq_linear(net, "fc")
weight = weight.reshape(net_params["N"],
net_params["F2"] * (net_params["T"] // 64))
#weight = align_array(weight)
# we want to align, not for the product F2*T//64, but for T//64 itself.
t64 = net_params["T"] // 64
t64_align = align_array_size(t64)
weight_align = np.zeros((net_params["N"], net_params["F2"] * t64_align),
dtype=int)
for i in range(net_params["F2"]):
weight_align[:, i * t64_align:i * t64_align +
t64] = weight[:, i * t64:(i + 1) * t64]
factor = convert.div_factor(input_scale, weight_scale, output_scale)
header.add(
HeaderComment(
"Layer 5\n"
"=======\n"
"Linear Layer (without scaling in the end)\n\n"
"Input: [F2, T // 64]\n"
"Weight: [N, F2 * (T // 64)]\n"
"Bias: [N]\n"
"Output: [N]",
mode="/*"))
header.add(HeaderConstant("NET_L5_FACTOR", factor))
header.add(HeaderArray("net_l5_bias", "int8_t", bias.ravel()))
header.add(HeaderConstant("NET_L5_WEIGHT_LEN", weight_align.shape[-1]))
header.add(HeaderArray("net_l5_weight", "int8_t", weight_align.ravel()))
# store the header file
header.write()
