Closed byzhang closed 6 years ago
and also the values in the output? Thanks a lot!
input_calibration_layer accepts features and applies 1-D calibration per each feature and returns them. In this example, we have 15 features, so the input to this layer has a shape, [batch_size, 15], and then the output should be [batch_size, 15].
Initial output value will depend on hparam's output_min and output_max. Since it is configured to [-1, 1], the output should be a linear function of your input that maps [input_min, input_max] to [-1, 1]. But note that output keypoints are trained jointly with your upper layer DNN. So as training proceeds, the output value will change even for the same input value :).
Thank you Seungil! We need to precompute the quantiles? And what's the effect of nodes_per_layer and layers in https://github.com/tensorflow/lattice/blob/master/examples/uci_census.py#L407
Yes. You need to pre-compute quantiles, but that can be done by passing the flag --compute_quantiles :).
nodes_per_layer and layers param is used in here: https://github.com/tensorflow/lattice/blob/master/examples/uci_census.py#L424
to construct DNN.
Thanks!
Thanks, -B
On Sat, Feb 10, 2018 at 4:59 PM, Seungil You notifications@github.com wrote:
Yes. You need to pre-compute quantiles, but that can be done by passing the flag --compute_quantiles :).
nodes_per_layer and layers param is used in here: https://github.com/tensorflow/lattice/blob/master/examples/ uci_census.py#L424
to construct DNN.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/tensorflow/lattice/issues/14#issuecomment-364709583, or mute the thread https://github.com/notifications/unsubscribe-auth/AA1O1WjgIeaS13K355HkvBXw3gZ00LTGks5tTjtbgaJpZM4SA0yw .
c.f. https://github.com/tensorflow/lattice/blob/master/examples/uci_census.py#L419