def lp_conv(self, input, k, s_h, s_w, bw, fl, rs, padding):
"""
Low precision convolution.
"""
c = tf.nn.conv2d(input, k, [1, s_h, s_w, 1], padding=padding)
return lp_op.lp(
c, bw, rs
) # Do the right shift here and bit truncation and satturation here !
def lp_deconv(self, input, k, output_shape, strides, bw, fl, rs,rate, padding):
"""
Low precision convolution.
"""
if rate==1:
c = tf.nn.conv2d_transpose(input, k, output_shape, strides=strides, padding=padding) # Tensorflow
else:
c = tf.nn.atrous_conv2d_transpose(input,k,output_shape,rate=rate,padding=padding)
return lp_op.lp(c, bw, rs) # Do the right shift here and bit truncation and satturation here !
def fc(self, input, num_out, name, bw=cfg.WORD_WIDTH, fl=10, rs=0, relu=True):
print("FC DEBUG: input: \n")
print(input)
print("FC DEBUG: done input\n")
with tf.variable_scope(name) as scope:
if isinstance(input, tuple):
input = input[0]
input_shape = input.get_shape()
if input_shape.ndims == 4:
dim = 1
print(" FC input_shape: %s\n" %(str(input_shape)))
#print input
print( input_shape[1:])
for d in input_shape[1:].as_list():
dim = dim * d
feed_in = tf.reshape(tf.transpose(input,[0,3,1,2]), [-1, dim])
else:
feed_in, dim = (input, int(input_shape[-1]))
if name == 'bbox_pred':
init_weights = tf.truncated_normal_initializer(0.0, stddev=0.001)
init_biases = tf.constant_initializer(0.0)
else:
init_weights = tf.truncated_normal_initializer(0.0, stddev=0.01)
init_biases = tf.constant_initializer(0.0)
weights = self.make_var('weights', [dim, num_out], init_weights, self.trainable, \
regularizer=self.l2_regularizer(0.0005))
biases = self.make_var('biases', [num_out], init_biases, self.trainable)
biases_dummy = self.make_var('biases_dummy', [num_out], init_biases, self.trainable) # Dummy bias to take care of saturation
if cfg.ENABLE_TENSORBOARD:
self.variable_summaries(biases)
self.variable_summaries(weights)
fc = tf.nn.xw_plus_b(feed_in, weights, biases_dummy, name=scope.name)
if self.isHardware:
fc = lp_op.lp(fc, bw, rs)
fc_biased = tf.add(fc, biases)
if self.isHardware:
fc_biased_saturate = self.saturate(fc_biased, cfg.WORD_WIDTH)
if relu:
return tf.nn.relu(fc_biased_saturate)
return fc_biased_saturate
if relu:
return tf.nn.relu(fc_biased)
return fc_biased