Exported fcn_resnet50_voc model has duplicate layer names about gluon-cv

[wfus]

I was able to fix it by just directly modifying the .json file containing the model definition after exporting it. Then after reloading the new model file everything works.

[zhanghang1989]

This looks wired. I will take a look

[ragavvenkatesan]

Any updates on this?

[Ishitori]

There was another encounter with similar issue here. I decided to reproduce it, and it seems that the issue is still there.

I have latest version of MXNet (1.3.1) and gluoncv (0.3.0) and I confirm that it acts weird - there are some non-unique names are used. Here is the example to reproduce it:

import gluoncv
import mxnet as mx
from mxnet import gluon

pretrained_net = gluoncv.model_zoo.get_fcn(dataset='pascal_voc', backbone='resnet50', pretrained=False)
pretrained_net.hybridize()

item = mx.random.uniform(shape=(1, 3, 512, 512))
pretrained_net(item)

pretrained_net.export("/tmp/fcn_resnet50_v1b")

deserialized_net = gluon.nn.SymbolBlock.imports("/tmp/fcn_resnet50_v1b-symbol.json", ['data'],
                                                "/tmp/fcn_resnet50_v1b-0000.params")

So I basically get the model, hybridize and do a forward pass, then export it to file system, and then load it back. If I do that I get:

ValueError: There are multiple outputs with name "fcn0_resnetv1b0_layers1_relu0_fwd_output"

I wrote a custom python script to parse json file and there are multiple nodes with same name. The full list I receive is:

fcn0_resnetv1b0_layers1_relu0_fwd
fcn0_resnetv1b0_layers1_relu0_fwd
fcn0_resnetv1b0_layers1_relu1_fwd
fcn0_resnetv1b0_layers1_relu1_fwd
fcn0_resnetv1b0_layers1_relu2_fwd
fcn0_resnetv1b0_layers1_relu2_fwd
fcn0_resnetv1b0_layers2_relu0_fwd
fcn0_resnetv1b0_layers2_relu0_fwd
fcn0_resnetv1b0_layers2_relu1_fwd
fcn0_resnetv1b0_layers2_relu1_fwd
fcn0_resnetv1b0_layers2_relu2_fwd
fcn0_resnetv1b0_layers2_relu2_fwd
fcn0_resnetv1b0_layers2_relu3_fwd
fcn0_resnetv1b0_layers2_relu3_fwd
fcn0_resnetv1b0_layers3_relu0_fwd
fcn0_resnetv1b0_layers3_relu0_fwd
fcn0_resnetv1b0_layers3_relu1_fwd
fcn0_resnetv1b0_layers3_relu1_fwd
fcn0_resnetv1b0_layers3_relu2_fwd
fcn0_resnetv1b0_layers3_relu2_fwd
fcn0_resnetv1b0_layers3_relu3_fwd
fcn0_resnetv1b0_layers3_relu3_fwd
fcn0_resnetv1b0_layers3_relu4_fwd
fcn0_resnetv1b0_layers3_relu4_fwd
fcn0_resnetv1b0_layers3_relu5_fwd
fcn0_resnetv1b0_layers3_relu5_fwd
fcn0_resnetv1b0_layers4_relu0_fwd
fcn0_resnetv1b0_layers4_relu0_fwd
fcn0_resnetv1b0_layers4_relu1_fwd
fcn0_resnetv1b0_layers4_relu1_fwd
fcn0_resnetv1b0_layers4_relu2_fwd
fcn0_resnetv1b0_layers4_relu2_fwd

Interesting, that they all are RELU.

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.