[FRONTEND][TENSORFLOW] Add extra_param to from_tensorflow

[jcf94]

In some graph, for example:

    g = tf.Graph()
    a_shape = [None, 30]
    b_shape = [30, 40]
    batch_size = 25
    static_a_shape = [batch_size, 30]
    dtype = "float32"
    with g.as_default():
        A = tf.placeholder(shape=a_shape, dtype=dtype, name="A")
        B = tf.placeholder(shape=b_shape, dtype=dtype, name="B")
        C = tf.matmul(A, B)
        batch = tf.placeholder(shape=(), dtype="int32", name="batch_size")
        batch = tf.expand_dims(batch, axis=0)
        shape = tf.concat([batch, tf.constant([5]), tf.constant([8])], axis=0)
        D = tf.reshape(C, shape)

We have input placeholder batch_size whose value would be used to determine the shape of another op. Even though the shape of tensor_a can be specified through from_tensorflow(shape=(...)) and in this case the batch_size value must be a fixed number, we cannot process this as a static graph.

This PR adds an extra parameter to from_tensorflow, that we can manually fix a tensor to be a constant value, then the graph can be processed as a static graph.

[u99127]

@leandron - is there a need to check this usage via tvmc as well if this changes ?

[junrushao]

CC @yongwww @comaniac

[comaniac]

AFAIK, other frontends also have the similar issue. For example, ONNX graph also put some shapes to the model parameters, but their constant values are maintained in the model artifacts when the model is frozen, so it may just need extra_param_names. As a result, instead of fixing this issue one-by-one, I'm wondering if we could have a unified solution. Specifically, can we do the following to achieve the same goal?

1. Get a dynamic graph via a frontend.
2. Bind parameters (i.e., batch_size) by name.
3. Run DynamicToStatic to get a static graph.

It seems to me that the above flow should result in the same outcome without this PR (assuming DynamicToStatic is doing a good job). If that works, it could be the best practice for not only TF but all frameworks IMHO.

[jcf94]

Thanks to all!
Have discussed with @comaniac before, and seems we need a better solution for this. I'll close this for now, and try to solve with another approach.