def __init__(self, dtype=np.float32, reshape_torch=False):
"""
Initializes a new LeNet inspired network
Args:
dtype: Datatype to be used
reshape_torch: set this, if the training parameters came from Pytorch which requires a custom reshape
"""
self.reshape_torch = reshape_torch
r1 = EggNet.ReshapeLayer(newshape=[-1, 28, 28, 1])
cn1 = EggNet.Conv2dLayer(in_channels=1, out_channels=16, kernel_size=3, activation='relu',
dtype=np.float32) # [? 28 28 16]
mp1 = EggNet.MaxPool2dLayer(size=2) # [? 14 14 16]
cn2 = EggNet.Conv2dLayer(in_channels=16, out_channels=32, kernel_size=3, activation='relu') # [? 14 14 32]
mp2 = EggNet.MaxPool2dLayer(size=2) # [? 7 7 32]
r2 = EggNet.FlattenLayer()
fc1 = EggNet.FullyConnectedLayer(input_size=32 * 7 * 7, output_size=32, activation='relu', dtype=np.float32)
fc2 = EggNet.FullyConnectedLayer(input_size=32, output_size=10, activation='softmax')
# Store a reference to each layer
self.r1 = r1
self.cn1 = cn1
self.mp1 = mp1
self.cn2 = cn2
self.mp2 = mp2
self.r2 = r2
self.fc1 = fc1
self.fc2 = fc2
self.lenet_layers = [r1, cn1, mp1, cn2, mp2, r2, fc1, fc2]
super(LeNet, self).__init__(self.lenet_layers)
def __init__(self, weights, options, shifts, real_quant=False):
# Check input
r1 = EggNet.ReshapeLayer(newshape=[-1, 28, 28, 1])
cn1 = EggNet.Conv2dLayer(in_channels=1, out_channels=16, kernel_size=3, activation='relu',
dtype=np.float32) # [? 28 28 16]
mp1 = EggNet.MaxPool2dLayer(size=2) # [? 14 14 16]
cn2 = EggNet.Conv2dLayer(in_channels=16, out_channels=32, kernel_size=3, activation='relu') # [? 14 14 32]
mp2 = EggNet.MaxPool2dLayer(size=2) # [? 7 7 32]
r2 = EggNet.FlattenLayer()
fc1 = EggNet.FullyConnectedLayer(input_size=32 * 7 * 7, output_size=32, activation='relu', dtype=np.float32)
if real_quant:
fc2 = EggNet.FullyConnectedLayer(input_size=32, output_size=10, activation=None)
else:
fc2 = EggNet.FullyConnectedLayer(input_size=32, output_size=10, activation='softmax')
if real_quant:
rs1 = EggNet.SimpleShiftLayer(shift=shifts[0], a_min=options['out_min'][0], a_max=options['out_max'][0])
rs2 = EggNet.SimpleShiftLayer(shift=shifts[1], a_min=options['out_min'][1], a_max=options['out_max'][1])
rs3 = EggNet.SimpleShiftLayer(shift=shifts[2], a_min=options['out_min'][2], a_max=options['out_max'][2])
rs4 = EggNet.SimpleShiftLayer(shift=shifts[3], a_min=options['out_min'][3], a_max=options['out_max'][3])
else:
# scales = 2.0 ** (-shifts)
scales = np.ones(shape=(4,))
rs1 = EggNet.ScaleLayer(scale=scales[0], a_min=options['out_min_f'][0], a_max=options['out_max_f'][0])
rs2 = EggNet.ScaleLayer(scale=scales[1], a_min=options['out_min_f'][1], a_max=options['out_max_f'][1])
rs3 = EggNet.ScaleLayer(scale=scales[2], a_min=options['out_min_f'][2], a_max=options['out_max_f'][2])
rs4 = EggNet.ScaleLayer(scale=scales[3], a_min=options['out_min_f'][3], a_max=options['out_max_f'][3])
self.rs1 = rs1
self.rs2 = rs2
self.rs3 = rs3
self.rs4 = rs4
# Store a reference to each layer
self.r1 = r1
self.cn1 = cn1
self.mp1 = mp1
self.cn2 = cn2
self.mp2 = mp2
self.r2 = r2
self.fc1 = fc1
self.fc2 = fc2
self.cn1.weights = weights['cn1.k']
self.cn1.bias = weights['cn1.b']
self.cn2.weights = weights['cn2.k']
self.cn2.bias = weights['cn2.b']
self.fc1.weights = weights['fc1.w']
self.fc1.bias = weights['fc1.b']
self.fc2.weights = weights['fc2.w']
self.fc2.bias = weights['fc2.b']
self.lenet_layers = [r1,
cn1, mp1, rs1,
cn2, mp2, rs2,
r2,
fc1, rs3,
fc2, rs4]
super(FpiLeNet, self).__init__(self.lenet_layers)
def _get_layers(weights_dict, target_bits, fraction_bits):
assert target_bits > fraction_bits
value_bits = target_bits - fraction_bits
a_max = 2 ** (value_bits - 1) - 1
a_min = -2 ** (value_bits - 1)
scale = 1 / 2 ** value_bits
c1_k1 = weights_dict['conv1_k']
c1_b1 = weights_dict['conv1_b']
c2_k2 = weights_dict['conv2_k']
c2_b2 = weights_dict['conv2_b']
fc1_w = weights_dict['fc1_w']
fc1_b = weights_dict['fc1_b']
fc2_w = weights_dict['fc2_w']
fc2_b = weights_dict['fc2_b']
# ni3 = nn_lenet_f64.fc1.input_size
# no3 = nn_lenet_f64.fc1.output_size
# ni4 = nn_lenet_f64.fc2.input_size
# no4 = nn_lenet_f64.fc2.output_size
ni3, no3 = fc1_w.shape
ni4, no4 = fc2_w.shape
qk1 = np.clip(c1_k1 / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qb1 = np.clip(c1_b1 / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qk2 = np.clip(c2_k2 / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qb2 = np.clip(c2_b2 / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qw3 = np.clip(fc1_w / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qb3 = np.clip(fc1_b / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qw4 = np.clip(fc2_w / scale, a_max=a_max, a_min=a_min).astype(np.int8)
qb4 = np.clip(fc2_b / scale, a_max=a_max, a_min=a_min).astype(np.int8)
dfrac_bits = 2 * fraction_bits
layers = [
EggNet.ReshapeLayer(newshape=(-1, 28, 28, 1)),
EggNet.Conv2dLayer(in_channels=1, out_channels=3, kernel_size=3, kernel_init_weights=qk1,
bias_init_weights=qb1, use_bias=True),
EggNet.ShiftLayer(target_bits=target_bits, target_frac_bits=fraction_bits, source_bits=16,
source_frac_bits=dfrac_bits),
EggNet.ReluActivationLayer(),
EggNet.MaxPool2dLayer(),
EggNet.Conv2dLayer(in_channels=3, out_channels=9, kernel_size=3, kernel_init_weights=qk2,
bias_init_weights=qb2, use_bias=True),
EggNet.ShiftLayer(target_bits=target_bits, target_frac_bits=fraction_bits, source_bits=16,
source_frac_bits=dfrac_bits),
EggNet.ReluActivationLayer(),
EggNet.MaxPool2dLayer(),
EggNet.FlattenLayer(),
EggNet.BreakpointLayer(enabled=False),
EggNet.FullyConnectedLayer(input_size=ni3, output_size=no3, dtype=np.int16, weights=qw3, bias=qb3),
EggNet.ShiftLayer(target_bits=target_bits, target_frac_bits=fraction_bits, source_bits=16,
source_frac_bits=dfrac_bits),
EggNet.ReluActivationLayer(),
EggNet.FullyConnectedLayer(input_size=ni4, output_size=no4, dtype=np.int16, weights=qw4, bias=qb4),
EggNet.ShiftLayer(target_bits=target_bits, target_frac_bits=fraction_bits, source_bits=16,
source_frac_bits=dfrac_bits),
EggNet.SoftmaxLayer()
]
return layers