Constraint on subset of weight matrix

I’ve got an MLP with an input layer connecting into a Dense layer. I currently have 3 inputs, the last of which must only have non-negative weightts, but for the other 2 covariates I don’t want any restrictions.

Currently, my Dense layer is defined as:

Dense(2, activation='relu',
      kernel_constraint=non_neg(), 
      kernel_initializer=RandomUniform(minval=0, maxval=2))

However, I’d ideally only apply this initialiser and constraint to the last input, i.e. the last row of the weight matrix. Is this possible?

[stulacy]

I had to separate the row of weights that I’m interested in and concatenate it with the rest of the weight matrix, rather than modify it in place.

class NonNegLast(Constraint):
    def __call__(self, w):
        last_row = w[-1, :] * K.cast(K.greater_equal(w[-1, :], 0.), K.floatx())
        last_row = K.expand_dims(last_row, axis=0)
        full_w = K.concatenate([w[:-1, :], last_row], axis=0)
        return full_w

[shaifugpt]

@raghakot How can we define such a custom constraint in a bidirectional wrapper, need two separate custom kernel constraints: one for the forward layer and other for backward layer?