def deform_conv_2d(img,
num_outputs,
kernel_size=3,
stride=2,
normalizer_fn=ly.batch_norm,
activation_fn=lrelu,
name=''):
img_shape = img.shape.as_list()
assert (len(img_shape) == 4)
N, C, H, W = img_shape
with tf.variable_scope('deform_conv' + '_' + name):
offset = ly.conv2d(img,
num_outputs=2 * kernel_size**2,
kernel_size=3,
stride=2,
activation_fn=None,
data_format='NCHW')
kernel = tf.get_variable(
name='d_kernel',
shape=(num_outputs, C, kernel_size, kernel_size),
initializer=tf.random_normal_initializer(0, 0.02))
res = deform_conv_op(img,
filter=kernel,
offset=offset,
rates=[1, 1, 1, 1],
padding='SAME',
strides=[1, 1, stride, stride],
num_groups=1)
if normalizer_fn is not None:
res = normalizer_fn(res)
if activation_fn is not None:
res = activation_fn(res)
return res
def run_benchmark():
global parameters
timing_entries = []
with tf.Graph().as_default():
# Generate some dummy images.
image_size = 224
kernel_col = 3
kernel_rol = 3
kernel_size = kernel_col * kernel_rol
channel = 64
# Note that our padding definition is slightly different the cuda-convnet.
# In order to force the model to start with the same activations sizes,
# we add 3 to the image_size and employ VALID padding above.
if FLAGS.data_format == 'NCHW':
image_shape = [FLAGS.batch_size, 3, image_size + 3, image_size + 3]
else:
image_shape = [FLAGS.batch_size, image_size + 3, image_size + 3, 3]
offset_shape = [
FLAGS.batch_size, 2 * kernel_size, image_size + 3, image_size + 3
]
kernel_shape = [channel, 3, kernel_col, kernel_rol]
images = tf.Variable(
tf.random_normal(image_shape, dtype=tf.float32, stddev=1e-1))
offset = tf.Variable(
tf.random_normal(offset_shape, dtype=tf.float32, stddev=1e-1))
kernel = tf.Variable(
tf.random_normal(kernel_shape, dtype=tf.float32, stddev=1e-1))
parameters = [kernel]
last_layer = deform_conv_op.deform_conv_op(images,
kernel,
offset,
strides=[1, 1, 1, 1],
rates=[1, 1, 1, 1],
padding="SAME",
num_groups=1,
deformable_group=1)
# Build an initialization operation.
init = tf.global_variables_initializer()
# Start running operations on the Graph.
sess = tf.Session()
sess.run(init)
run_forward = True
run_forward_backward = True
if FLAGS.forward_only and FLAGS.forward_backward_only:
raise ValueError("Cannot specify --forward_only and "
"--forward_backward_only at the same time.")
if FLAGS.forward_only:
run_forward_backward = False
elif FLAGS.forward_backward_only:
run_forward = False
if run_forward:
# Run the forward benchmark.
timing_entries.append(
time_tensorflow_run(sess, last_layer, "Forward"))
if run_forward_backward:
# Add a simple objective so we can calculate the backward pass.
# objective = loss(last_layer, labels)
loss = lambda x: tf.reduce_sum(x)
objective = loss(last_layer)
# Compute the gradient with respect to all the parameters.
grad = tf.gradients(objective, parameters)
# Run the backward benchmark.
timing_entries.append(
time_tensorflow_run(sess, grad, "Forward-backward"))
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # see issue #152
# os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import numpy as np
import tensorflow as tf
import lib.deform_conv_op as deform_conv_op
# load test data, generated by np.random.random((8, 6, 4, 5))
with open("test.npz", 'rb') as f:
arr = np.load(f)
# arr = np.zeros((8, 6, 4, 5))
with tf.Session() as sess:
with tf.device('/gpu:0'):
a = tf.constant(arr, dtype=tf.float32)
b = tf.constant(np.ones((21, 2, 2, 2), dtype=np.float32))
c = tf.constant(np.ones((8, 8, 2, 2), dtype=np.float32))
result = deform_conv_op.deform_conv_op(a,
b,
c,
strides=[1, 1, 2, 2],
rates=[1, 1, 1, 1],
padding="VALID",
num_groups=3)
sm = sess.run(result)
d = tf.constant(np.ones((8, 21, 2, 2), dtype=np.float32))
grad = tf.gradients(result, [a, b, c])
res = [sess.run(g) for g in grad]
print(res[0])
# print(sm)
