def conv_base(inputs, kernel, dilation=(1, 1), downsample=False):
filters = inputs.shape[1]
if downsample:
output = layers.conv(inputs, filters * 2, kernel, (2, 2))
else:
output = layers.conv(inputs, filters, kernel, dilation=dilation)
return output
def pair_base(inputs, kernel, downsample=False):
filters = inputs.shape[1]
if downsample:
output = layers.conv(inputs, filters, (1, kernel), (1, 2))
output = layers.conv(output, filters, (kernel, 1), (2, 1))
output = layers.conv(output, filters * 2, (1, 1))
else:
output = layers.conv(inputs, filters, (1, kernel))
output = layers.conv(output, filters, (kernel, 1))
return output
def block(x, tokens, pre_activation=False, downsample=False):
""" build block. """
if pre_activation:
x = layers.bn_relu(x)
res = x
else:
res = x
x = layers.bn_relu(x)
x = ops[tokens[0]](x, downsample)
print("%s \t-> shape %s" % (ops[0].__name__, x.shape))
for token in tokens[1:]:
x = layers.bn_relu(x)
x = ops[token](x)
print("%s \t-> shape %s" % (ops[token].__name__, x.shape))
if downsample:
filters = res.shape[1]
if FLAGS.downsample == "conv":
res = layers.conv(res, filters * 2, (1, 1), (2, 2))
elif FLAGS.downsample == "pool":
res = layers.avgpool(res, (2, 2), (2, 2))
res = fluid.layers.pad(
res, (0, 0, filters // 2, filters // 2, 0, 0, 0, 0))
else:
raise NotImplementedError
return x + res
def maxpool_base(inputs, kernel, downsample=False):
if downsample:
filters = inputs.shape[1]
output = layers.maxpool(inputs, kernel, (2, 2))
output = layers.conv(output, filters * 2, (1, 1))
else:
output = layers.maxpool(inputs, kernel)
return output
def net(inputs, tokens):
""" build network with skip links """
x = layers.conv(inputs, FLAGS.width, (3, 3))
num_ops = FLAGS.num_blocks * FLAGS.num_ops
x = stage(x, tokens[:num_ops], pre_activation=True)
for i in range(1, FLAGS.num_stages):
x = stage(x, tokens[i * num_ops:(i + 1) * num_ops], downsample=True)
x = layers.bn_relu(x)
x = layers.global_avgpool(x)
x = layers.dropout(x)
logits = layers.fully_connected(x, num_classes)
return fluid.layers.softmax(logits)
def net(inputs, tokens):
depth = len(tokens)
q, r = divmod(depth + 1, FLAGS.num_stages)
downsample_steps = [
i * q + max(0, i + r - FLAGS.num_stages + 1) - 2
for i in range(1, FLAGS.num_stages)
]
x = layers.conv(inputs, FLAGS.width, (3, 3))
x = layers.bn_relu(x)
for i, token in enumerate(tokens):
downsample = i in downsample_steps
x = ops[token](x, downsample)
print("%s \t-> shape %s" % (ops[token].__name__, x.shape))
if downsample:
print("=" * 12)
x = layers.bn_relu(x)
x = layers.global_avgpool(x)
x = layers.dropout(x)
logits = layers.fully_connected(x, num_classes)
return fluid.layers.softmax(logits)