Using Caffe or TensorFlow networks with DAAL

[Igor_M_]

Hi,
I am interested in using the intel daal library with networks trained in different frameworks. For example would it be possible to load caffe netowork (with prototxt for architecture and weights) and use it in daal for inference and/or finetuning?

thx

[Ruslan_I_Intel]

Hi Igor,

DAAL models are not fully compatible with Caffe or TensorFlow ones, due to different subsets of parameters of the layers and some discrepancies in implementation. So automated conversion from one representation to another with the guaranties of correctness is hard to implement. We are analyzing how to support this use case, depending on the analysis we would provide samples and tools for semi-automated model conversion on one of future releases of the library.

For now I would suggest to reproduce required topology in the following way:

• read Caffe or TensorFlow models via Python;
• extract weights and biases from them;
• set the weights and biases to the DAAL model. 

Please, have a look at the example shows how it can be done with Python https://software.intel.com/en-us/node/682264 (there are similar ones for C++ and Java).

Let us know, if you need any assistance from our side in this conversion.