Drop the manually_create field from RnnState.

Initially, I thought that the shape of RNN state arrays could always be
determined by shape propagation. Then I came across some graphs where this
wasn't so easy to infer, so I introduced manually_create thinking of it
as a hack. Today I took another look at dropping that hack, and had a
"D'oh" moment when I realized that the cyclic nature of RNN graphs makes
it impossible to infer the shapes of all arrays by usual propagation.
For example, in a LSTM cell, the input array is concatenated with
a state array, so if we don't already know the shape of that state array,
shape propagation stops there.

Thus, this change removes manually_create by making toco always behave as
if manually_create=true, i.e. early-creating all RNN state arrays with
the shape explicitly specified by the user. The next TODO item here
(see model_flags.proto) is to introduce a generic 'shape' field, so far
the current 'size' field only allows specifying 1-D shapes.

PiperOrigin-RevId: 183294102

1 files changed, 0 insertions, 3 deletions

diff --git a/tensorflow/contrib/lite/toco/model_cmdline_flags.cc b/tensorflow/contrib/lite/toco/model_cmdline_flags.cc
index 790b3443ce..36520d9c55 100644
--- a/tensorflow/contrib/lite/toco/model_cmdline_flags.cc
+++ b/tensorflow/contrib/lite/toco/model_cmdline_flags.cc
@@ -327,9 +327,6 @@ void ReadModelFlagsFromCommandLineFlags(
         CHECK(absl::SimpleAtoi(value, &size));
         CHECK_GT(size, 0);
         rnn_state_proto->set_size(size);
-      } else if (key == "manually_create") {
-        CHECK_EQ(absl::AsciiStrToLower(value), "true");
-        rnn_state_proto->set_manually_create(true);
       } else {
         LOG(FATAL) << "Unknown key '" << key << "' in --rnn_states";
       }