Inception v1 missing 5x5 Convolutions

[JoshVarty]

According to Going Deeper with Convolutions the original inception modules had the following structure:

However, while reviewing models/research/slim/nets/inception_v1.py I noticed that the inception modules were missing 5x5 convolutions and instead had an extra 3x3 convolution in Branch_2:

models/research/slim/nets/inception_v1.py

Lines 84 to 97 in 6263692

	with tf.variable_scope(end_point):
	with tf.variable_scope('Branch_0'):
	branch_0 = slim.conv2d(net, 64, [1, 1], scope='Conv2d_0a_1x1')
	with tf.variable_scope('Branch_1'):
	branch_1 = slim.conv2d(net, 96, [1, 1], scope='Conv2d_0a_1x1')
	branch_1 = slim.conv2d(branch_1, 128, [3, 3], scope='Conv2d_0b_3x3')
	with tf.variable_scope('Branch_2'):
	branch_2 = slim.conv2d(net, 16, [1, 1], scope='Conv2d_0a_1x1')
	branch_2 = slim.conv2d(branch_2, 32, [3, 3], scope='Conv2d_0b_3x3')
	with tf.variable_scope('Branch_3'):
	branch_3 = slim.max_pool2d(net, [3, 3], scope='MaxPool_0a_3x3')
	branch_3 = slim.conv2d(branch_3, 32, [1, 1], scope='Conv2d_0b_1x1')
	net = tf.concat(
	axis=3, values=[branch_0, branch_1, branch_2, branch_3])

Would you like me to submit a PR to correct this? If so, how would you like to handle the pre-trained networks that you have available for download?

/cc @nealwu @sguada

[nealwu]

@derekjchow can you take a look? Looks like the model needs a fix, which means we'll want to retrain a new checkpoint.

[JoshVarty]

Also there are a few other things I've noticed that might be mistakes or things I'm misunderstanding.

1. There are no LocalResponseNorm layers as indicated in the paper.

2. I'm not sure if the 1x1 Convolution at the front of the network is using the correct weights. However, the paper is missing the 1x1 Convolution from its weight table, so it's hard to say.

The images (in both the paper and code) are indicated to be of size 224x224.

type	Output Size	Note
Input	224x224x3
Conv 7x7	112x112x64
Max Pool	56x56x64
Conv 1x1	56x56x64	I think this is unnecessary since there is no dimensionality reduction
Conv 3x3	56x56x192

The code for these layers is available here:

models/research/slim/nets/inception_v1.py

Lines 62 to 71 in 6263692

	end_point = 'Conv2d_1a_7x7'
	net = slim.conv2d(inputs, 64, [7, 7], stride=2, scope=end_point)
	end_points[end_point] = net
	if final_endpoint == end_point: return net, end_points
	end_point = 'MaxPool_2a_3x3'
	net = slim.max_pool2d(net, [3, 3], stride=2, scope=end_point)
	end_points[end_point] = net
	if final_endpoint == end_point: return net, end_points
	end_point = 'Conv2d_2b_1x1'
	net = slim.conv2d(net, 64, [1, 1], scope=end_point)

[tensorflowbutler]

Hi There,
We are checking to see if you still need help on this, as this seems to be considerably old issue. Please update this issue with the latest information, code snippet to reproduce your issue and error you are seeing.
If we don't hear from you in the next 7 days, this issue will be closed automatically. If you don't need help on this issue any more, please consider closing this.

[amahendrakar]

Automatically closing due to lack of recent activity. Please update the issue when new information becomes available, and we will reopen the issue. Thanks!