[Improvement] Model definitions

[d02k01]

Some model definitions are not precise:

• Original MobileNetV2 models dose not have any densely-connected layers.
• Resnet50 weights are initialized by using he_normal.

[taehoonlee]

Thank you for the comments @mktozk.

• For MobileNetV2, can you elaborate on the mismatch you think? The equivalence between tensorflow/models and keras-applications has already been checked.
• For ResNet50, the initializers are he_normal, xavier (the tf-layer default), and glorot (the keras default) in facebook/fb.resnet.torch, tensorflow/models, and keras-applications, respectively. The original one is he_normal as you said, but they didn't show significant differences.

[d02k01]

Thank you for your reply.

• For computing outputs, a fully-connected layer is used in keras-applications while a 1x1 conv2d is used in tensorflow/models. Their operations are not same but provide same results in this case.
• You are right. There are little differences. FYI: he_normal initializers are used in Resnet implementation by chainer. In this implementation, an initialization depends on using pretrained weights.

My comments will be helpful to increase reproducibility and to reduce ambiguity.

[taehoonlee]

Note that all the models in here are for only inference.

• In the very last layer, the Dense(classes) and the Conv2D(classes, 1) produce the same results for 224x224 inputs. This is just a style difference rather than a mismatch. Actually, I also want to replace Dense(classes) with Conv2D(classes, 1) but the replacement requires updates of all the weight files for all the models including MobileNetV2. That is a very big construction.
• OK. I'll update with kernel_initializer='he_normal'.

[taehoonlee]

@mktozk, I'll close the issue for now. Please feel free to open it again at any time if you have additional comments.