def get_layers(self):
return [
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=784,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense2',
bits=self.bits,
in_units=1024,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=1024,
units=10,
weight_decay=self.weight_decay,
),
]
def get_layers(self):
return [
dfxp.Conv2d_q(
name='conv1',
bits=self.bits,
ksize=[5, 5, 1, 6],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'
),
dfxp.Conv2d_q(
name='conv2',
bits=self.bits,
ksize=[5, 5, 6, 16],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'
),
dfxp.Conv2d_q(
name='conv3',
bits=self.bits,
ksize=[5, 5, 16, 120],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Flatten_q(120),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=120,
units=84,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=84,
units=10,
weight_decay=self.weight_decay,
),
]
def get_layers(self):
self.channels = 16
return [
dfxp.Conv2d_pq(
name='conv1',
bits=self.bits,
ksize=[3, 3, 3, 16],
strides=[1, 1, 1, 1],
padding='SAME',
use_bias=False,
weight_decay=self.weight_decay,
),
# dfxp.GradientBuffer_q(
# name='grad_buffer',
# bits=self.bits,
# shape=[32, 32, 32, 16],
# ),
dfxp.BatchNorm_q(
name='conv1-bn',
bits=self.bits,
num_features=16,
training=self.training,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
] + self._build_blocks(16, self.num_blocks[0], 1) \
+ self._build_blocks(32, self.num_blocks[1], 2) \
+ self._build_blocks(64, self.num_blocks[2], 2) \
+ [
dfxp.AvgPool_q(
ksize=[1, 8, 8, 1],
strides=[1, 1, 1, 1],
padding='VALID',
),
dfxp.Flatten_q(64),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=64,
units=10,
use_bias=False,
weight_decay=self.weight_decay,
),
# dfxp.GradientBuffer_q(
# name='gradient_buffer',
# bits=self.bits,
# shape=[32, 10], # TODO use batch size
# ),
# dfxp.BatchNorm_q(
# name='softmax-bn',
# bits=self.bits,
# num_features=10,
# training=self.training,
# weight_decay=self.weight_decay,
# ),
]
def get_layers(self):
return [
dfxp.Conv2d_q(
name='conv',
bits=self.bits,
training=self.training,
ksize=[5, 5, 1, 20],
strides=[1, 1, 1, 1],
padding='VALID',
weight_decay=self.weight_decay,
),
dfxp.BatchNorm_q(
name='batch_normolization',
bits=self.bits,
num_features=20,
training=self.training,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID'),
dfxp.Flatten_q(12 * 12 * 20),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
training=self.training,
in_units=12 * 12 * 20,
units=100,
weight_decay=self.weight_decay,
),
dfxp.Dense_q(
name='dense2',
bits=self.bits,
training=self.training,
in_units=100,
units=10,
weight_decay=self.weight_decay,
)
]
def get_layers(self):
return [
# conv1-1
dfxp.Conv2d_q(
name='conv1-1',
bits=self.bits,
ksize=[3, 3, 3, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv1-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv1-2',
bits=self.bits,
ksize=[3, 3, 128, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool1
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv2-1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2-1',
bits=self.bits,
ksize=[3, 3, 128, 256],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv2-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2-2',
bits=self.bits,
ksize=[3, 3, 256, 256],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool2
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv3-1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3-1',
bits=self.bits,
ksize=[3, 3, 256, 512],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# conv3-2
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3-2',
bits=self.bits,
ksize=[3, 3, 512, 512],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# pool3
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
dfxp.Flatten_q(512*4*4),
# dense1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=512*4*4,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# dense2
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense2',
bits=self.bits,
in_units=1024,
units=1024,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# softmax
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=1024,
units=10,
weight_decay=self.weight_decay,
),
]
def get_layers(self):
return [
# conv1
# dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv1',
bits=self.bits,
ksize=[5, 5, 3, 64],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv2
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv2',
bits=self.bits,
ksize=[5, 5, 64, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
# conv3
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Conv2d_q(
name='conv3',
bits=self.bits,
ksize=[5, 5, 128, 128],
strides=[1, 1, 1, 1],
padding='SAME',
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
dfxp.MaxPool_q(
ksize=[1, 3, 3, 1],
strides=[1, 2, 2, 1],
padding='SAME'
),
dfxp.Flatten_q(128*4*4),
# dense1
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='dense1',
bits=self.bits,
in_units=128*4*4,
units=400,
weight_decay=self.weight_decay,
),
dfxp.ReLU_q(),
# softmax
dfxp.Dropout_q(self.dropout, self.training),
dfxp.Dense_q(
name='softmax',
bits=self.bits,
in_units=400,
units=10,
weight_decay=self.weight_decay,
),
]
def __init__(self, model, dataset, dataset_name, logger, params):
self.n_epoch = params.n_epoch
self.exp_path = params.exp_path
self.logger = logger
self.graph = tf.Graph()
with self.graph.as_default():
global_step = tf.train.get_or_create_global_step()
self.lr_scheduler = LearningRateScheduler(params.lr,
params.lr_decay_epoch,
params.lr_decay_factor)
optimizer = tf.train.MomentumOptimizer(params.lr, params.momentum)
tower_grads, tower_loss = [], []
with tf.variable_scope(tf.get_variable_scope()):
images = avatar.batch_data()
images = tf.cast(tf.reshape(images, [-1, 32, 32, 3]),
dtype=tf.float32)
lables = avatar.batch_lable()
conv1 = dfxp.Conv2d_q(name='conv1',
bits=params.bits,
training=False,
ksize=[3, 3, 3, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.conv_in_1 = images
self.batch_in_1, _, _ = conv1.forward(self.conv_in_1)
batch1 = dfxp.Normalization_q(name='batch1',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_1 = batch1.forward(self.batch_in_1)
scale1 = dfxp.Rescale_q(name='scale1',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_1, self.g, self.b = scale1.forward(
self.scale_in_1)
relu1 = dfxp.ReLU_q()
self.conv_in_2 = relu1.forward(self.relu_in_1)
conv2 = dfxp.Conv2d_q(name='conv2',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_2, self.conv_in_2_q, _ = conv2.forward(
self.conv_in_2)
batch2 = dfxp.Normalization_q(name='batch2',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_2 = batch2.forward(self.batch_in_2)
scale2 = dfxp.Rescale_q(name='scale2',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_2, _, _ = scale2.forward(self.scale_in_2)
relu2 = dfxp.ReLU_q()
self.conv_in_3 = relu2.forward(self.relu_in_2)
conv3 = dfxp.Conv2d_q(name='conv3',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_3, _, _ = conv3.forward(self.conv_in_3)
batch3 = dfxp.Normalization_q(name='batch3',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_3 = batch3.forward(self.batch_in_3)
scale3 = dfxp.Rescale_q(name='scale3',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_32, _, _ = scale3.forward(self.scale_in_3)
self.relu_in_3 = self.relu_in_32 + self.conv_in_2_q
relu3 = dfxp.ReLU_q()
self.conv_in_4 = relu3.forward(self.relu_in_3)
conv4 = dfxp.Conv2d_q(name='conv4',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_4, self.conv_in_4_q, _ = conv4.forward(
self.conv_in_4)
batch4 = dfxp.Normalization_q(name='batch4',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_4 = batch4.forward(self.batch_in_4)
scale4 = dfxp.Rescale_q(name='scale4',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_4, _, _ = scale4.forward(self.scale_in_4)
relu4 = dfxp.ReLU_q()
self.conv_in_5 = relu4.forward(self.relu_in_4)
conv5 = dfxp.Conv2d_q(name='conv5',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_5, _, _ = conv5.forward(self.conv_in_5)
batch5 = dfxp.Normalization_q(name='batch5',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_5 = batch5.forward(self.batch_in_5)
scale5 = dfxp.Rescale_q(name='scale5',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_52, _, _ = scale5.forward(self.scale_in_5)
self.relu_in_5 = self.relu_in_52 + self.conv_in_4_q
relu5 = dfxp.ReLU_q()
self.conv_in_6 = relu5.forward(self.relu_in_5)
conv6 = dfxp.Conv2d_q(name='conv6',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_6, self.conv_in_6_q, _ = conv6.forward(
self.conv_in_6)
batch6 = dfxp.Normalization_q(name='batch6',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_6 = batch6.forward(self.batch_in_6)
scale6 = dfxp.Rescale_q(name='scale6',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_6, _, _ = scale6.forward(self.scale_in_6)
relu6 = dfxp.ReLU_q()
self.conv_in_7 = relu6.forward(self.relu_in_6)
conv7 = dfxp.Conv2d_q(name='conv7',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 16],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_7, _, _ = conv7.forward(self.conv_in_7)
batch7 = dfxp.Normalization_q(name='batch7',
bits=params.bits,
num_features=16,
training=True)
self.scale_in_7 = batch7.forward(self.batch_in_7)
scale7 = dfxp.Rescale_q(name='scale7',
bits=params.bits,
training=False,
num_features=16)
self.relu_in_72, _, _ = scale7.forward(self.scale_in_7)
self.relu_in_7 = self.relu_in_72 + self.conv_in_6_q
relu7 = dfxp.ReLU_q()
self.conv_in_8 = relu7.forward(self.relu_in_7)
conv8 = dfxp.Conv2d_q(name='conv8',
bits=params.bits,
training=False,
ksize=[3, 3, 16, 32],
strides=[1, 2, 2, 1],
padding='VALID')
self.conv_in_8_1 = tf.pad(self.conv_in_8,
[[0, 0], [1, 0], [1, 0], [0, 0]],
'CONSTANT')
self.batch_in_8, _, _ = conv8.forward(self.conv_in_8_1)
batch8 = dfxp.Normalization_q(name='batch8',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_8 = batch8.forward(self.batch_in_8)
scale8 = dfxp.Rescale_q(name='scale8',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_8, _, _ = scale8.forward(self.scale_in_8)
relu8 = dfxp.ReLU_q()
self.conv_in_9 = relu8.forward(self.relu_in_8)
conv9 = dfxp.Conv2d_q(name='conv9',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 32],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_9, _, _ = conv9.forward(self.conv_in_9)
batch9 = dfxp.Normalization_q(name='batch9',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_9 = batch9.forward(self.batch_in_9)
scale9 = dfxp.Rescale_q(name='scale9',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_9, _, _ = scale9.forward(self.scale_in_9)
conv10 = dfxp.Conv2d_q(name='conv10',
bits=params.bits,
training=False,
ksize=[1, 1, 16, 32],
strides=[1, 2, 2, 1],
padding='VALID')
self.batch_in_10, _, self.w = conv10.forward(self.conv_in_8)
batch10 = dfxp.Normalization_q(name='batch10',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_10 = batch10.forward(self.batch_in_10)
scale10 = dfxp.Rescale_q(name='scale10',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_102, _, _ = scale10.forward(self.scale_in_10)
self.relu_in_10 = self.relu_in_102 + self.relu_in_9
relu10 = dfxp.ReLU_q()
self.conv_in_11 = relu10.forward(self.relu_in_10)
conv11 = dfxp.Conv2d_q(name='conv11',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 32],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_11, self.conv_in_11_q, _ = conv11.forward(
self.conv_in_11)
batch11 = dfxp.Normalization_q(name='batch11',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_11 = batch11.forward(self.batch_in_11)
scale11 = dfxp.Rescale_q(name='scale11',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_11, _, _ = scale11.forward(self.scale_in_11)
relu11 = dfxp.ReLU_q()
self.conv_in_12 = relu11.forward(self.relu_in_11)
conv12 = dfxp.Conv2d_q(name='conv12',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 32],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_12, _, _ = conv12.forward(self.conv_in_12)
batch12 = dfxp.Normalization_q(name='batch12',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_12 = batch12.forward(self.batch_in_12)
scale12 = dfxp.Rescale_q(name='scale12',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_122, _, _ = scale12.forward(self.scale_in_12)
self.relu_in_12 = self.relu_in_122 + self.conv_in_11_q
relu12 = dfxp.ReLU_q()
self.conv_in_13 = relu12.forward(self.relu_in_12)
conv13 = dfxp.Conv2d_q(name='conv13',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 32],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_13, self.conv_in_13_q, _ = conv13.forward(
self.conv_in_13)
batch13 = dfxp.Normalization_q(name='batch13',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_13 = batch13.forward(self.batch_in_13)
scale13 = dfxp.Rescale_q(name='scale13',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_13, _, _ = scale13.forward(self.scale_in_13)
relu13 = dfxp.ReLU_q()
self.conv_in_14 = relu13.forward(self.relu_in_13)
conv14 = dfxp.Conv2d_q(name='conv14',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 32],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_14, _, _ = conv14.forward(self.conv_in_14)
batch14 = dfxp.Normalization_q(name='batch14',
bits=params.bits,
num_features=32,
training=True)
self.scale_in_14 = batch14.forward(self.batch_in_14)
scale14 = dfxp.Rescale_q(name='scale14',
bits=params.bits,
training=False,
num_features=32)
self.relu_in_142, _, _ = scale14.forward(self.scale_in_14)
self.relu_in_14 = self.relu_in_142 + self.conv_in_13_q
relu14 = dfxp.ReLU_q()
self.conv_in_15 = relu14.forward(self.relu_in_14)
conv15 = dfxp.Conv2d_q(name='conv15',
bits=params.bits,
training=False,
ksize=[3, 3, 32, 64],
strides=[1, 2, 2, 1],
padding='VALID')
self.conv_in_15_1 = tf.pad(self.conv_in_15,
[[0, 0], [1, 0], [1, 0], [0, 0]],
'CONSTANT')
self.batch_in_15, _, _ = conv15.forward(self.conv_in_15_1)
batch15 = dfxp.Normalization_q(name='batch15',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_15 = batch15.forward(self.batch_in_15)
scale15 = dfxp.Rescale_q(name='scale15',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_15, _, _ = scale15.forward(self.scale_in_15)
relu15 = dfxp.ReLU_q()
self.conv_in_16 = relu15.forward(self.relu_in_15)
conv16 = dfxp.Conv2d_q(name='conv16',
bits=params.bits,
training=False,
ksize=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_16, _, _ = conv16.forward(self.conv_in_16)
batch16 = dfxp.Normalization_q(name='batch16',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_16 = batch16.forward(self.batch_in_16)
scale16 = dfxp.Rescale_q(name='scale16',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_16, _, _ = scale16.forward(self.scale_in_16)
conv17 = dfxp.Conv2d_q(name='conv17',
bits=params.bits,
training=False,
ksize=[1, 1, 32, 64],
strides=[1, 2, 2, 1],
padding='VALID')
self.batch_in_17, _, _ = conv17.forward(self.conv_in_15)
batch17 = dfxp.Normalization_q(name='batch17',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_17 = batch17.forward(self.batch_in_17)
scale17 = dfxp.Rescale_q(name='scale17',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_172, _, _ = scale17.forward(self.scale_in_17)
self.relu_in_17 = self.relu_in_172 + self.relu_in_16
relu17 = dfxp.ReLU_q()
self.conv_in_18 = relu17.forward(self.relu_in_17)
conv18 = dfxp.Conv2d_q(name='conv18',
bits=params.bits,
training=False,
ksize=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_18, self.conv_in_18_q, _ = conv18.forward(
self.conv_in_18)
batch18 = dfxp.Normalization_q(name='batch18',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_18 = batch18.forward(self.batch_in_18)
scale18 = dfxp.Rescale_q(name='scale18',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_18, _, _ = scale18.forward(self.scale_in_18)
relu18 = dfxp.ReLU_q()
self.conv_in_19 = relu18.forward(self.relu_in_18)
conv19 = dfxp.Conv2d_q(name='conv19',
bits=params.bits,
training=False,
ksize=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_19, _, _ = conv19.forward(self.conv_in_19)
batch19 = dfxp.Normalization_q(name='batch19',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_19 = batch19.forward(self.batch_in_19)
scale19 = dfxp.Rescale_q(name='scale19',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_192, _, _ = scale19.forward(self.scale_in_19)
self.relu_in_19 = self.relu_in_192 + self.conv_in_18_q
relu19 = dfxp.ReLU_q()
self.conv_in_20 = relu19.forward(self.relu_in_19)
conv20 = dfxp.Conv2d_q(name='conv20',
bits=params.bits,
training=False,
ksize=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_20, self.conv_in_20_q, _ = conv20.forward(
self.conv_in_20)
batch20 = dfxp.Normalization_q(name='batch20',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_20 = batch20.forward(self.batch_in_20)
scale20 = dfxp.Rescale_q(name='scale20',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_20, _, _ = scale20.forward(self.scale_in_20)
relu20 = dfxp.ReLU_q()
self.conv_in_21 = relu20.forward(self.relu_in_20)
conv21 = dfxp.Conv2d_q(name='conv21',
bits=params.bits,
training=False,
ksize=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME')
self.batch_in_21, _, _ = conv21.forward(self.conv_in_21)
batch21 = dfxp.Normalization_q(name='batch21',
bits=params.bits,
num_features=64,
training=True)
self.scale_in_21 = batch21.forward(self.batch_in_21)
scale21 = dfxp.Rescale_q(name='scale21',
bits=params.bits,
training=False,
num_features=64)
self.relu_in_212, _, _ = scale21.forward(self.scale_in_21)
self.relu_in_21 = self.relu_in_212 + self.conv_in_20_q
relu21 = dfxp.ReLU_q()
self.conv_in_22 = relu21.forward(self.relu_in_21)
pool = dfxp.AvgPool_q(ksize=[1, 8, 8, 1],
strides=[1, 1, 1, 1],
padding='VALID')
self.fc_in_1 = pool.forward(self.conv_in_22)
flat = dfxp.Flatten_q(1 * 1 * 64)
self.flat = flat.forward(self.fc_in_1)
fc1 = dfxp.Dense_q(name='dense1',
bits=params.bits,
training=False,
in_units=64,
units=10)
self.softmax_in, _ = fc1.forward(self.flat)
self.loss = tf.nn.softmax_cross_entropy_with_logits(
labels=lables, logits=self.softmax_in)
self.indiff = tf.gradients(self.loss, self.conv_in_2)
self.train_step = optimizer.minimize(self.loss)
self.init_op = tf.global_variables_initializer()
self.saver = tf.train.Saver()
self.summary = tf.summary.merge_all()
self.graph.finalize()
def __init__(self, model, dataset, dataset_name, logger, params):
self.n_epoch = params.n_epoch
self.exp_path = params.exp_path
self.logger = logger
self.graph = tf.Graph()
with self.graph.as_default():
global_step = tf.train.get_or_create_global_step()
self.lr_scheduler = LearningRateScheduler(params.lr,
params.lr_decay_epoch,
params.lr_decay_factor)
optimizer = tf.train.MomentumOptimizer(params.lr, params.momentum)
tower_grads, tower_loss = [], []
with tf.variable_scope(tf.get_variable_scope()):
images = avatar.batch_data()
images = tf.cast(tf.reshape(images, [-1, 28, 28, 1]),
dtype=tf.float64)
lables = avatar.batch_lable()
conv = dfxp.Conv2d_q(name='conv',
bits=params.bits,
training=False,
ksize=[5, 5, 1, 20],
strides=[1, 1, 1, 1],
padding='VALID')
self.conv_in = images
self.batch_in, self.conv_w, self.x_s, self.w_s, self.conv_w_q = conv.forward(
self.conv_in)
batch = dfxp.Normalization_q(name='batch',
bits=params.bits,
num_features=20,
training=True)
self.scale_in = batch.forward(self.batch_in)
scale = dfxp.Rescale_q(name='scale',
bits=params.bits,
training=False,
num_features=20)
self.relu_in, self.scale_w, self.scale_b = scale.forward(
self.scale_in)
relu = dfxp.ReLU_q()
self.pool_in = relu.forward(self.relu_in)
pool = dfxp.MaxPool_q(ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1],
padding='VALID')
flat = dfxp.Flatten_q(12 * 12 * 20)
self.fc1_in = pool.forward(self.pool_in)
self.flat = flat.forward(self.fc1_in)
fc1 = dfxp.Dense_q(name='dense1',
bits=params.bits,
training=False,
in_units=12 * 12 * 20,
units=100)
self.fc2_in, self.w1, self.b1 = fc1.forward(self.flat)
fc2 = dfxp.Dense_q(name='dense2',
bits=params.bits,
training=False,
in_units=100,
units=10)
self.softmax_in, self.w2, self.b2 = fc2.forward(self.fc2_in)
self.loss = tf.nn.softmax_cross_entropy_with_logits(
labels=lables, logits=self.softmax_in)
self.conv_indiff = tf.gradients(self.loss, self.conv_in)
self.batch_indiff = tf.gradients(self.loss, self.batch_in)
self.scale_indiff = tf.gradients(self.loss, self.scale_in)
self.train_step = optimizer.minimize(self.loss)
self.init_op = tf.global_variables_initializer()
self.saver = tf.train.Saver()
self.summary = tf.summary.merge_all()
self.graph.finalize()
