The output problem of sharing weights.

[LynnHo]

I use the code like below,

...
resnet = partial(tensornets.ResNet50, reuse=tf.AUTO_REUSE)
outputs_1 = resnet(x_1).get_outputs()
outputs_2 = resnet(x_2).get_outputs() 
...

outputs_2 appends the outputs of x_2 to outputs_1.

[LynnHo]

@taehoonlee Do you have any idea? Is it a bug?

[taehoonlee]

@LynnHo, Sorry for the late reply. This is because the two models have the same prefix for the output tensors. Currently, the following is one of the workarounds:

import tensorflow as tf
import tensornets as nets
from functools import partial

x_1 = tf.placeholder(tf.float32, [None, 224, 224, 3], name='x1')
x_2 = tf.placeholder(tf.float32, [None, 224, 224, 3], name='x2')

resnet = partial(nets.ResNet50, scope='resnet50', reuse=tf.AUTO_REUSE)
model1 = resnet(x_1)
with tf.name_scope('xx'):
    model2 = resnet(x_2)

assert len(model1.get_outputs()) == 161
assert len(model2.get_outputs()) == 161
assert len(model1.get_weights()) == 320
assert len(model2.get_weights()) == 320

If I find the elegant methods or make weight sharing more seamless, I'll let you know. Thank you for the reports!

[taehoonlee]

The recent updates address the issue you reported. Currently, I recommend the following pattern:

with tf.name_scope('clone0'):
    model1 = nets.ResNet50(x1, reuse=tf.AUTO_REUSE)
with tf.name_scope('clone1'):
    model2 = nets.ResNet50(x2, reuse=tf.AUTO_REUSE)

These are the materials you might be interested in:

1. Fix: ae8e431
2. Notes: Notes for variable_scope, name_scope, and weight sharing #43
3. Test codes: b1c2533

If you want to use the latest version, you can try pip install -U tensornets. Please feel free to open it again at any time if you have additional comments :)

[LynnHo]

@taehoonlee Thanks for your effort！ Tensornets really helps in my works！