I am trying to run squeezedet on ncs and i found the following operations are not supported as of now
- tf.unstack and tf.stack
I replaced stridedslice with the operation tf.slice and tf.unstack operation with tf.split.
But for the operation tf.math.greater i could not find any other replacements. Basically i am performing the following op
(tensor > threshold ) which returns boolean tensor
I wrote a another way to calculate greater operation as follows
- convert float to int32 dtype tensor
- right shift the number by 31 bits so that signed bit becomes the LSB
- perform bitwise and of 0x1 , input tensor
- now the result will contain either 1 or 0 as value
- Result one indicates input value is less than my threshold and zero indicated input value is greater than my threshold.
To perform the above operation i need to convert my float tensor to int tensor. But Cast is not supported. So these are my questions
How to get know the operations supported by intel movidius sdk( Not the layers but the operations like unstack, stack, slice, tf.tocast)
I am thinking of splitting the model into multiple graphs. perform the operations not supported by ncs sdk on host machine and take the results and do the remaining
Perform initial operations on movidius. take the result
calculate the unsupported operations available in the graph output on host machine and perform remaining ops on movidius.
So I would like to know how to load more than one graphs on movidius.