Tensorflow parser problem (Slim and tf.layers)
Model: Inception-resnet-v2 and inception-resnet-v1
Source: Tensorflow
Destination: IR
Author: Jiahao
When we convert the above two models from tensorflow to IR, they crashed because of flatten and batchnorm layer.
Our tensorflow parser works on those op from tf.layers. Nevertheless, the op of the above two models comes from tensorflow.slim.
The difficulty in slim module is that there are lots of condition operator in the graph.
In graph definition, we can see the difference in tensorboard.
tf.layers
tf.slim
We can output the inputs of the FusedBatchNorm operator.
tf.layers
tf.slim
For the flatten layer in slim, it is also built up with many small ops.
To conclude, the complexity of the ops adds to the difficulty of conversion.