Hello,
Would you like to include my implementation using Tensorflow? I use some cool tricks, like slim's "@argscope" annotation to create a custom tensorflow layer, and mid to high level Tensorflow APIs such as the new tf.data API. Might be a helpful reference point for people trying to write Tensorflow code without a lot of boilerplate.
For example, here's your fire module:
@add_arg_scope
def fire_module(inputs,
squeeze_depth,
expand_depth,
reuse=None,
scope=None):
with tf.variable_scope(scope, 'fire', [inputs], reuse=reuse):
with arg_scope([conv2d, max_pool2d]):
net = _squeeze(inputs, squeeze_depth)
net = _expand(net, expand_depth)
return net
def _squeeze(inputs, num_outputs):
return conv2d(inputs, num_outputs, [1, 1], stride=1, scope='squeeze')
def _expand(inputs, num_outputs):
with tf.variable_scope('expand'):
e1x1 = conv2d(inputs, num_outputs, [1, 1], stride=1, scope='1x1')
e3x3 = conv2d(inputs, num_outputs, [3, 3], scope='3x3')
return tf.concat([e1x1, e3x3], 1)
Hello, Would you like to include my implementation using Tensorflow? I use some cool tricks, like slim's "@argscope" annotation to create a custom tensorflow layer, and mid to high level Tensorflow APIs such as the new tf.data API. Might be a helpful reference point for people trying to write Tensorflow code without a lot of boilerplate.
For example, here's your fire module: