def test_scaling_parameter_from_stats(self):
shape = [8, 3, 64, 64]
collect_stats_steps = 100
stats_act = QuantReLU(
bit_width=BIT_WIDTH,
quant_type=QuantType.INT,
scaling_impl_type=ScalingImplType.PARAMETER_FROM_STATS,
scaling_stats_permute_dims=None,
scaling_stats_op=StatsOp.PERCENTILE,
collect_stats_steps=collect_stats_steps,
scaling_min_val=None,
percentile_q=99.0)
stats_act.train()
tensor_quant = stats_act.act_quant.fused_activation_quant_proxy.tensor_quant
scaling_value = tensor_quant.scaling_impl.value
for i in range(collect_stats_steps):
inp = torch.randn(shape)
out = stats_act(inp)
out.requires_grad_(True) # i need something to require a grad
out.sum().backward()
assert scaling_value.grad is None
inp = torch.randn(shape)
out = stats_act(inp)
out.sum().backward()
assert scaling_value.grad is not None
def __init__(self, input_width, weight_width, act_width):
super(QuantLeNet, self).__init__()
self.quant_inp = QuantIdentity(bit_width=input_width,
min_val=-1.0,
max_val=1.0)
self.conv1 = QuantConv2d(1, 6, 5, weight_bit_width=weight_width)
self.conv2 = QuantConv2d(6, 16, 5, weight_bit_width=weight_width)
self.fc1 = QuantLinear(16 * 4 * 4,
120,
bias=True,
weight_bit_width=weight_width)
self.fc2 = QuantLinear(120,
84,
bias=True,
weight_bit_width=weight_width)
self.fc3 = QuantLinear(84,
10,
bias=False,
weight_bit_width=weight_width)
self.relu1 = QuantReLU(bit_width=act_width, max_val=6)
self.relu2 = QuantReLU(bit_width=act_width, max_val=6)
self.relu3 = QuantReLU(bit_width=act_width, max_val=6)
self.relu4 = QuantReLU(bit_width=act_width, max_val=6)
def __init__(self, model_config):
super(JetSubstructureNeqModel, self).__init__()
self.model_config = model_config
self.num_neurons = [model_config["input_length"]] + model_config["hidden_layers"] + [model_config["output_length"]]
layer_list = []
for i in range(1, len(self.num_neurons)):
in_features = self.num_neurons[i-1]
out_features = self.num_neurons[i]
bn = nn.BatchNorm1d(out_features)
if i == 1:
bn_in = nn.BatchNorm1d(in_features)
input_bias = ScalarBiasScale(scale=False, bias_init=-0.25)
input_quant = QuantBrevitasActivation(QuantHardTanh(model_config["input_bitwidth"], max_val=1., narrow_range=False, quant_type=QuantType.INT, scaling_impl_type=ScalingImplType.PARAMETER), pre_transforms=[bn_in, input_bias])
output_quant = QuantBrevitasActivation(QuantReLU(bit_width=model_config["hidden_bitwidth"], max_val=1.61, quant_type=QuantType.INT, scaling_impl_type=ScalingImplType.PARAMETER), pre_transforms=[bn])
mask = RandomFixedSparsityMask2D(in_features, out_features, fan_in=model_config["input_fanin"])
layer = SparseLinearNeq(in_features, out_features, input_quant=input_quant, output_quant=output_quant, sparse_linear_kws={'mask': mask})
layer_list.append(layer)
elif i == len(self.num_neurons)-1:
output_bias_scale = ScalarBiasScale(bias_init=0.33)
output_quant = QuantBrevitasActivation(QuantHardTanh(bit_width=model_config["output_bitwidth"], max_val=1.33, narrow_range=False, quant_type=QuantType.INT, scaling_impl_type=ScalingImplType.PARAMETER), pre_transforms=[bn], post_transforms=[output_bias_scale])
mask = RandomFixedSparsityMask2D(in_features, out_features, fan_in=model_config["output_fanin"])
layer = SparseLinearNeq(in_features, out_features, input_quant=layer_list[-1].output_quant, output_quant=output_quant, sparse_linear_kws={'mask': mask}, apply_input_quant=False)
layer_list.append(layer)
else:
output_quant = QuantBrevitasActivation(QuantReLU(bit_width=model_config["hidden_bitwidth"], max_val=1.61, quant_type=QuantType.INT, scaling_impl_type=ScalingImplType.PARAMETER), pre_transforms=[bn])
mask = RandomFixedSparsityMask2D(in_features, out_features, fan_in=model_config["hidden_fanin"])
layer = SparseLinearNeq(in_features, out_features, input_quant=layer_list[-1].output_quant, output_quant=output_quant, sparse_linear_kws={'mask': mask}, apply_input_quant=False)
layer_list.append(layer)
self.module_list = nn.ModuleList(layer_list)
self.is_verilog_inference = False
self.latency = 1
self.verilog_dir = None
self.top_module_filename = None
self.dut = None
self.logfile = None
def __init__(self):
super().__init__()
self.conv1 = QuantConv2d(kernel_size=KERNEL_SIZE,
in_channels=CHANNELS,
out_channels=CHANNELS,
weight_quant=DPUv1WeightQuantInjector,
bias_quant=None,
output_quant=DPUv1OutputQuantInjector,
bias=False,
return_quant_tensor=True)
self.act1 = QuantReLU(act_quant=DPUv1ActQuantInjector,
return_quant_tensor=True)
self.conv2 = QuantConv2d(kernel_size=KERNEL_SIZE,
in_channels=CHANNELS,
out_channels=CHANNELS,
weight_quant=DPUv1WeightQuantInjector,
bias_quant=None,
output_quant=DPUv1OutputQuantInjector,
bias=False,
return_quant_tensor=True)
self.act2 = QuantReLU(act_quant=DPUv1ActQuantInjector,
return_quant_tensor=True)
self.conv3 = QuantConv2d(kernel_size=KERNEL_SIZE,
in_channels=CHANNELS,
out_channels=CHANNELS,
weight_quant=DPUv1WeightQuantInjector,
bias_quant=None,
output_quant=DPUv1OutputQuantInjector,
bias=False,
return_quant_tensor=True)
self.act3 = QuantReLU(act_quant=DPUv1ActQuantInjector,
return_quant_tensor=False)
self.linear = nn.Linear(FC_IN_SIZE, CHANNELS)
def make_qrelu(no_quant=True, **kwargs) -> QuantReLU:
if no_quant:
return QuantReLU(input_quant=None,
act_quant=None,
output_quant=None,
update_iqi=None,
update_aqi=None)
else:
return QuantReLU(bit_width=kwargs['bit_width'])
def thresholds(module: QuantReLU):
num_distinct_values = 2**int(module.quant_act_bit_width().item())
num_thresholds = num_distinct_values - 1
flat_scale = module.quant_act_scale().view(-1)
num_scale_channels = flat_scale.shape[0]
step = torch.abs(flat_scale)
min_threshold = step / 2
thresholds = torch.empty(num_scale_channels, num_thresholds)
for c in range(num_scale_channels):
for t in range(num_thresholds):
thresholds[c][t] = min_threshold[c] + step[c] * t
return thresholds
def test_scaling_parameter_grad(self):
stats_act = QuantReLU(bit_width=BIT_WIDTH,
max_val=MAX_VAL,
quant_type=QuantType.INT,
scaling_impl_type=ScalingImplType.PARAMETER)
stats_act.train()
for i in range(RANDOM_ITERS):
inp = torch.randn([8, 3, 64, 64])
stats_act(inp)
out = stats_act(inp)
out.sum().backward()
tensor_quant = stats_act.act_quant.fused_activation_quant_proxy.tensor_quant
scaling_value = tensor_quant.scaling_impl.value
assert scaling_value.grad is not None
def __init__(self):
super().__init__()
self.act1 = QuantIdentity(
bit_width=7,
act_quant=ShiftedUint8ActPerTensorFloat,
return_quant_tensor=True)
self.act2 = QuantReLU(act_quant=Uint8ActPerTensorFloat)
def __init__(self,
in_channels,
out_channels,
kernel_size,
weight_bit_width,
act_bit_width,
stride=1,
padding=0,
groups=1,
bn_eps=1e-5,
activation_scaling_per_channel=False):
super(ConvBlock, self).__init__()
self.conv = QuantConv2d(in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=False,
weight_quant=CommonIntWeightPerChannelQuant,
weight_bit_width=weight_bit_width)
self.bn = nn.BatchNorm2d(num_features=out_channels, eps=bn_eps)
self.activation = QuantReLU(
act_quant=CommonUintActQuant,
bit_width=act_bit_width,
per_channel_broadcastable_shape=(1, out_channels, 1, 1),
scaling_per_channel=activation_scaling_per_channel,
return_quant_tensor=True)
def __init__(self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
quant_type,
weight_bit_width,
act_bit_width,
act_scaling_per_channel,
weight_scaling_impl_type,
bias,
compute_micronet_cost,
dilation=1,
groups=1,
bn_eps=1e-5,
shared_act=None):
super(ConvBlock, self).__init__()
self.compute_micronet_cost = compute_micronet_cost
self.conv = QuantConv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
dilation=dilation,
groups=groups,
bias=bias,
weight_quant_type=quant_type,
weight_bit_width=weight_bit_width,
weight_scaling_impl_type=weight_scaling_impl_type,
weight_restrict_scaling_type=RestrictValueType.LOG_FP,
weight_narrow_range=True,
weight_scaling_stats_op=StatsOp.MAX,
weight_scaling_min_val=MIN_SCALING_VALUE,
compute_output_bit_width=
True, # Compute the number of bits in the output accumulator
return_quant_tensor=
True, # Return a quantized tensor that represents the quantized accumulator
weight_scaling_per_output_channel=True)
self.bn = nn.BatchNorm2d(num_features=out_channels, eps=bn_eps)
if shared_act is None and quant_type == QuantType.FP:
self.activ = nn.ReLU6()
elif shared_act is None and quant_type == QuantType.INT:
self.activ = QuantReLU(
quant_type=quant_type,
bit_width=act_bit_width,
max_val=RELU_MAX_VAL,
scaling_per_channel=act_scaling_per_channel,
scaling_impl_type=ScalingImplType.PARAMETER,
scaling_min_val=MIN_SCALING_VALUE,
restrict_scaling_type=RestrictValueType.LOG_FP,
per_channel_broadcastable_shape=(1, out_channels, 1, 1),
return_quant_tensor=True)
elif shared_act is not None:
self.activ = shared_act
else:
raise Exception("Activ non recognized.")
def quant_type(
module: QuantReLU,
supported_bit_width: Tuple[int,...] = (2, 4, 8, 16, 32)):
bit_width = int(module.quant_act_bit_width().item())
if bit_width in list(supported_bit_width):
return f"UINT{bit_width}"
else:
raise RuntimeError(f"Unsupported input bit width {bit_width} for export")
def quant_type(
module: QuantReLU,
supported_int_bit_width_range: Tuple[int,...] = (2, 33)):
bit_width = int(module.quant_act_bit_width().item())
if bit_width in range(*supported_int_bit_width_range):
return f"UINT{bit_width}"
else:
raise RuntimeError(f"Unsupported input bit width {bit_width} for export")
def thresholds(module: QuantReLU, extend_tensor_to_channels=True):
num_distinct_values = 2 ** int(module.quant_act_bit_width().item())
num_thresholds = num_distinct_values - 1
flat_scale = module.quant_act_scale().view(-1)
num_scale_channels = flat_scale.shape[0]
step = torch.abs(flat_scale)
min_threshold = step / 2
thresholds = torch.empty(num_scale_channels, num_thresholds)
for c in range(num_scale_channels):
for t in range(num_thresholds):
thresholds[c][t] = min_threshold[c] + step[c] * t
if extend_tensor_to_channels:
output_channels = module._cached_inp.shape[1]
final_shape = (output_channels, num_thresholds)
if thresholds.shape != final_shape:
thresholds = thresholds.expand(final_shape)
return thresholds
def test_brevitas_act_export_relu(abits, max_val, scaling_impl_type,
QONNX_export):
min_val = -1.0
ishape = (1, 15)
b_act = QuantReLU(
bit_width=abits,
max_val=max_val,
scaling_impl_type=scaling_impl_type,
restrict_scaling_type=RestrictValueType.LOG_FP,
quant_type=QuantType.INT,
)
if scaling_impl_type == ScalingImplType.PARAMETER:
checkpoint = {
"act_quant_proxy.fused_activation_quant_proxy.tensor_quant.\
scaling_impl.learned_value":
torch.tensor(0.49).type(torch.FloatTensor)
}
b_act.load_state_dict(checkpoint)
if QONNX_export:
m_path = export_onnx_path
BrevitasONNXManager.export(b_act, ishape, m_path)
qonnx_cleanup(m_path, out_file=m_path)
model = ModelWrapper(m_path)
model = model.transform(ConvertQONNXtoFINN())
model.save(m_path)
else:
bo.export_finn_onnx(b_act, ishape, export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
inp_tensor = np.random.uniform(low=min_val, high=max_val,
size=ishape).astype(np.float32)
idict = {model.graph.input[0].name: inp_tensor}
odict = oxe.execute_onnx(model, idict, True)
produced = odict[model.graph.output[0].name]
inp_tensor = torch.from_numpy(inp_tensor).float()
b_act.eval()
expected = b_act.forward(inp_tensor).detach().numpy()
if not np.isclose(produced, expected, atol=1e-3).all():
print(abits, max_val, scaling_impl_type)
print("scale: ",
b_act.quant_act_scale().type(torch.FloatTensor).detach())
if abits < 5:
print(
"thres:",
", ".join(["{:8.4f}".format(x)
for x in b_act.export_thres[0]]),
)
print("input:",
", ".join(["{:8.4f}".format(x) for x in inp_tensor[0]]))
print("prod :", ", ".join(["{:8.4f}".format(x) for x in produced[0]]))
print("expec:", ", ".join(["{:8.4f}".format(x) for x in expected[0]]))
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
def test_brevitas_act_export_relu_imagenet(abits, max_val,
scaling_per_channel):
out_channels = 32
ishape = (1, out_channels, 1, 1)
min_val = -1.0
b_act = QuantReLU(
bit_width=abits,
quant_type=QuantType.INT,
scaling_impl_type=ScalingImplType.PARAMETER,
scaling_per_channel=scaling_per_channel,
restrict_scaling_type=RestrictValueType.LOG_FP,
scaling_min_val=2e-16,
max_val=6.0,
return_quant_tensor=True,
per_channel_broadcastable_shape=(1, out_channels, 1, 1),
)
if scaling_per_channel is True:
rand_tensor = (2) * torch.rand((1, out_channels, 1, 1))
else:
rand_tensor = torch.tensor(1.2398)
checkpoint = {
"act_quant_proxy.fused_activation_quant_proxy.tensor_quant.\
scaling_impl.learned_value":
rand_tensor.type(torch.FloatTensor)
}
b_act.load_state_dict(checkpoint)
bo.export_finn_onnx(b_act, ishape, export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(InferShapes())
inp_tensor = np.random.uniform(low=min_val, high=max_val,
size=ishape).astype(np.float32)
idict = {model.graph.input[0].name: inp_tensor}
odict = oxe.execute_onnx(model, idict, True)
produced = odict[model.graph.output[0].name]
inp_tensor = torch.from_numpy(inp_tensor).float()
b_act.eval()
expected = b_act.forward(inp_tensor).tensor.detach().numpy()
if not np.isclose(produced, expected, atol=1e-3).all():
print(abits, max_val)
print("scale: ",
b_act.quant_act_scale().type(torch.FloatTensor).detach())
if abits < 5:
print(
"thres:",
", ".join(["{:8.4f}".format(x)
for x in b_act.export_thres[0]]),
)
print("input:",
", ".join(["{:8.4f}".format(x) for x in inp_tensor[0]]))
print("prod :", ", ".join(["{:8.4f}".format(x) for x in produced[0]]))
print("expec:", ", ".join(["{:8.4f}".format(x) for x in expected[0]]))
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
def __init__(self, cfg, batch_norm, bit_width=8, num_classes=1000):
super(QuantVGG, self).__init__()
self.features = make_layers(cfg, batch_norm, bit_width)
self.avgpool = QuantAvgPool2d(kernel_size=(7, 7), stride=1, bit_width=bit_width)
self.classifier = nn.Sequential(
QuantLinear(
512 * 7 * 7, 4096, bias=True,
weight_quant=CommonIntWeightPerChannelQuant, weight_bit_width=bit_width),
QuantReLU(act_quant=CommonUintActQuant, bit_width=bit_width),
nn.Dropout(),
QuantLinear(
4096, 4096, bias=True,
weight_quant=CommonIntWeightPerChannelQuant, weight_bit_width=bit_width),
QuantReLU(act_quant=CommonUintActQuant, weight_bit_width=bit_width),
nn.Dropout(),
QuantLinear(
4096, num_classes, bias=False,
weight_quant=CommonIntWeightPerTensorQuant, weight_bit_width=bit_width),
)
self._initialize_weights()
def __init__(self):
super().__init__()
self.conv1 = QuantConv2d(
out_channels=OUT_CH,
in_channels=IN_CH,
kernel_size=KERNEL_SIZE,
bias=False,
weight_quant=Int8WeightPerTensorFixedPoint,
input_quant=Int8ActPerTensorFixedPoint,
output_quant=Int8ActPerTensorFixedPoint,
return_quant_tensor=True)
self.relu = QuantReLU(act_quant=None, return_quant_tensor=False)
self.conv1.weight.data.uniform_(-0.01, 0.01)
def make_layers(cfg, batch_norm, bit_width):
layers = []
in_channels = 3
for v in cfg:
if v == 'M':
layers += [nn.MaxPool2d(kernel_size=2, stride=2)]
else:
conv2d = QuantConv2d(
in_channels, v, kernel_size=3, stride=1, padding=1, groups=1, bias=not batch_norm,
weight_bit_width=bit_width, weight_quant=CommonIntWeightPerChannelQuant)
act = QuantReLU(
act_quant=CommonUintActQuant, bit_width=bit_width, return_quant_tensor=True)
if batch_norm:
layers += [conv2d, nn.BatchNorm2d(v), act]
else:
layers += [conv2d, act]
in_channels = v
return nn.Sequential(*layers)
def __init__(
self,
in_channels,
out_channels,
kernel_size,
stride,
padding,
weight_bit_width,
act_bit_width,
act_scaling_per_channel,
bias,
groups=1,
bn_eps=1e-5,
shared_act=None,
return_quant_tensor=False):
super(ConvBlock, self).__init__()
self.conv = QuantConv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
groups=groups,
bias=bias,
weight_bit_width=weight_bit_width,
weight_quant=CommonIntWeightPerChannelQuant)
self.bn = nn.BatchNorm2d(num_features=out_channels, eps=bn_eps)
if shared_act is None:
self.activ = QuantReLU(
act_quant=CommonUintActQuant,
bit_width=act_bit_width,
scaling_per_channel=act_scaling_per_channel,
per_channel_broadcastable_shape=(1, out_channels, 1, 1),
return_quant_tensor=return_quant_tensor)
else:
self.activ = shared_act
def __init__(self):
super(QuantLeNet, self).__init__()
self.conv1 = QuantConv2d(1,
6,
5,
weight_quant=None,
input_quant=None,
bias_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
self.relu1 = QuantReLU(input_quant=None,
act_quant=None,
output_quant=None,
update_iqi=None,
update_aqi=None)
self.conv2 = QuantConv2d(6,
16,
5,
weight_quant=None,
input_quant=None,
bias_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
self.relu2 = QuantReLU(input_quant=None,
act_quant=None,
output_quant=None,
update_iqi=None,
update_aqi=None)
self.fc1 = QuantLinear(16 * 5 * 5,
120,
bias=True,
weight_quant=None,
input_quant=None,
bias_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
self.relu3 = QuantReLU(input_quant=None,
act_quant=None,
output_quant=None,
update_iqi=None,
update_aqi=None)
self.fc2 = QuantLinear(120,
84,
bias=True,
weight_quant=None,
input_quant=None,
bias_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
self.relu4 = QuantReLU(input_quant=None,
act_quant=None,
output_quant=None,
update_iqi=None,
update_aqi=None)
self.fc3 = QuantLinear(84,
10,
bias=False,
weight_quant=None,
input_quant=None,
bias_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
def quant_act_scale(module: QuantReLU):
quant_act_scale = module.quant_act_scale().type(torch.FloatTensor).detach()
return quant_act_scale
def get_8_bits_quantized_lenet():
model = QuantLeNet()
model.conv1 = QuantConv2d(1,
6,
5,
weight_quant=LSQ_weight_quant_8bits,
bias_quant=None,
input_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
model.relu1 = QuantReLU(input_quant=None,
act_quant=LSQ_input_quant_8bits,
output_quant=None,
update_iqi=None,
update_aqi=None)
model.conv2 = QuantConv2d(6,
16,
5,
weight_quant=LSQ_weight_quant_8bits,
bias_quant=None,
input_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
model.relu2 = QuantReLU(input_quant=None,
act_quant=LSQ_input_quant_8bits,
output_quant=None,
update_iqi=None,
update_aqi=None)
model.fc1 = QuantLinear(16 * 5 * 5,
120,
bias=True,
weight_quant=LSQ_weight_quant_8bits,
bias_quant=None,
input_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
model.relu3 = QuantReLU(input_quant=None,
act_quant=LSQ_input_quant_8bits,
output_quant=None,
update_iqi=None,
update_aqi=None)
model.fc2 = QuantLinear(120,
84,
bias=True,
weight_quant=LSQ_weight_quant_8bits,
bias_quant=None,
input_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
model.relu4 = QuantReLU(input_quant=None,
act_quant=LSQ_input_quant_8bits,
output_quant=None,
update_iqi=None,
update_aqi=None)
model.fc3 = QuantLinear(84,
10,
bias=False,
weight_quant=LSQ_weight_quant_8bits,
bias_quant=None,
input_quant=None,
output_quant=None,
update_wqi=None,
update_bqi=None,
update_iqi=None,
update_oqi=None)
return model
def test_module_init_const_scaling(self):
mod = QuantReLU(max_val=6, scaling_impl_type='CONST')
def test_module_init_default(self):
mod = QuantReLU(max_val=6)
def test_end2end_cybsec_mlp_export(QONNX_export):
assets_dir = pk.resource_filename("finn.qnn-data", "cybsec-mlp/")
# load up trained net in Brevitas
input_size = 593
hidden1 = 64
hidden2 = 64
hidden3 = 64
weight_bit_width = 2
act_bit_width = 2
num_classes = 1
model = nn.Sequential(
QuantLinear(input_size,
hidden1,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden1),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden1,
hidden2,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden2),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden2,
hidden3,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden3),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden3,
num_classes,
bias=True,
weight_bit_width=weight_bit_width),
)
trained_state_dict = torch.load(assets_dir +
"/state_dict.pth")["models_state_dict"][0]
model.load_state_dict(trained_state_dict, strict=False)
W_orig = model[0].weight.data.detach().numpy()
# pad the second (593-sized) dimensions with 7 zeroes at the end
W_new = np.pad(W_orig, [(0, 0), (0, 7)])
model[0].weight.data = torch.from_numpy(W_new)
model_for_export = CybSecMLPForExport(model)
export_onnx_path = get_checkpoint_name("export", QONNX_export)
input_shape = (1, 600)
# create a QuantTensor instance to mark the input as bipolar during export
input_a = np.random.randint(0, 1, size=input_shape).astype(np.float32)
input_a = 2 * input_a - 1
scale = 1.0
input_t = torch.from_numpy(input_a * scale)
input_qt = QuantTensor(input_t,
scale=torch.tensor(scale),
bit_width=torch.tensor(1.0),
signed=True)
if QONNX_export:
# With the BrevitasONNXManager we need to manually set
# the FINN DataType at the input
BrevitasONNXManager.export(model_for_export,
input_shape,
export_path=export_onnx_path)
model = ModelWrapper(export_onnx_path)
model.set_tensor_datatype(model.graph.input[0].name,
DataType["BIPOLAR"])
model.save(export_onnx_path)
qonnx_cleanup(export_onnx_path, out_file=export_onnx_path)
model = ModelWrapper(export_onnx_path)
model = model.transform(ConvertQONNXtoFINN())
model.save(export_onnx_path)
else:
bo.export_finn_onnx(model_for_export,
export_path=export_onnx_path,
input_t=input_qt)
assert os.path.isfile(export_onnx_path)
# fix input datatype
finn_model = ModelWrapper(export_onnx_path)
finnonnx_in_tensor_name = finn_model.graph.input[0].name
assert tuple(finn_model.get_tensor_shape(finnonnx_in_tensor_name)) == (1,
600)
# verify a few exported ops
if QONNX_export:
# The first "Mul" node doesn't exist in the QONNX export,
# because the QuantTensor scale is not exported.
# However, this node would have been unity scale anyways and
# the models are still equivalent.
assert finn_model.graph.node[0].op_type == "Add"
assert finn_model.graph.node[1].op_type == "Div"
assert finn_model.graph.node[2].op_type == "MatMul"
assert finn_model.graph.node[-1].op_type == "MultiThreshold"
else:
assert finn_model.graph.node[0].op_type == "Mul"
assert finn_model.get_initializer(
finn_model.graph.node[0].input[1]) == 1.0
assert finn_model.graph.node[1].op_type == "Add"
assert finn_model.graph.node[2].op_type == "Div"
assert finn_model.graph.node[3].op_type == "MatMul"
assert finn_model.graph.node[-1].op_type == "MultiThreshold"
# verify datatypes on some tensors
assert (finn_model.get_tensor_datatype(finnonnx_in_tensor_name) ==
DataType["BIPOLAR"])
first_matmul_w_name = finn_model.get_nodes_by_op_type("MatMul")[0].input[1]
assert finn_model.get_tensor_datatype(
first_matmul_w_name) == DataType["INT2"]
def test_end2end_cybsec_mlp_export():
assets_dir = pk.resource_filename("finn.qnn-data", "cybsec-mlp/")
# load up trained net in Brevitas
input_size = 593
hidden1 = 64
hidden2 = 64
hidden3 = 64
weight_bit_width = 2
act_bit_width = 2
num_classes = 1
model = nn.Sequential(
QuantLinear(input_size,
hidden1,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden1),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden1,
hidden2,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden2),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden2,
hidden3,
bias=True,
weight_bit_width=weight_bit_width),
nn.BatchNorm1d(hidden3),
nn.Dropout(0.5),
QuantReLU(bit_width=act_bit_width),
QuantLinear(hidden3,
num_classes,
bias=True,
weight_bit_width=weight_bit_width),
)
trained_state_dict = torch.load(assets_dir +
"/state_dict.pth")["models_state_dict"][0]
model.load_state_dict(trained_state_dict, strict=False)
W_orig = model[0].weight.data.detach().numpy()
# pad the second (593-sized) dimensions with 7 zeroes at the end
W_new = np.pad(W_orig, [(0, 0), (0, 7)])
model[0].weight.data = torch.from_numpy(W_new)
model_for_export = CybSecMLPForExport(model)
export_onnx_path = get_checkpoint_name("export")
input_shape = (1, 600)
# create a QuantTensor instance to mark the input as bipolar during export
input_a = np.random.randint(0, 1, size=input_shape).astype(np.float32)
input_a = 2 * input_a - 1
scale = 1.0
input_t = torch.from_numpy(input_a * scale)
input_qt = QuantTensor(input_t,
scale=torch.tensor(scale),
bit_width=torch.tensor(1.0),
signed=True)
bo.export_finn_onnx(model_for_export,
export_path=export_onnx_path,
input_t=input_qt)
assert os.path.isfile(export_onnx_path)
# fix input datatype
finn_model = ModelWrapper(export_onnx_path)
finnonnx_in_tensor_name = finn_model.graph.input[0].name
assert tuple(finn_model.get_tensor_shape(finnonnx_in_tensor_name)) == (1,
600)
# verify a few exported ops
assert finn_model.graph.node[1].op_type == "Add"
assert finn_model.graph.node[2].op_type == "Div"
assert finn_model.graph.node[3].op_type == "MatMul"
assert finn_model.graph.node[-1].op_type == "MultiThreshold"
# verify datatypes on some tensors
assert finn_model.get_tensor_datatype(
finnonnx_in_tensor_name) == DataType.BIPOLAR
first_matmul_w_name = finn_model.graph.node[3].input[1]
assert finn_model.get_tensor_datatype(first_matmul_w_name) == DataType.INT2
def test_scaling_stats_to_parameter(self):
stats_act = QuantReLU(bit_width=BIT_WIDTH,
max_val=MAX_VAL,
quant_type=QuantType.INT,
scaling_impl_type=ScalingImplType.STATS)
stats_act.train()
for i in range(RANDOM_ITERS):
inp = torch.randn([8, 3, 64, 64])
stats_act(inp)
stats_state_dict = stats_act.state_dict()
param_act = QuantReLU(bit_width=BIT_WIDTH,
max_val=MAX_VAL,
quant_type=QuantType.INT,
scaling_impl_type=ScalingImplType.PARAMETER)
param_act.load_state_dict(stats_state_dict)
stats_act.eval()
param_act.eval()
assert(torch.allclose(stats_act.quant_act_scale(), param_act.quant_act_scale()))
def quant_type(module: QuantReLU):
bit_width = module.quant_act_bit_width()
signed = module.is_quant_act_signed
return finn_datatype(bit_width, signed)
def PreQuantizedReLU(config):
return QuantReLU(bit_width=config.activation_bit_width,
max_val=6.0,
quant_type=QuantType.INT)