| Commit message (Collapse) | Author | Age |
|
|
|
|
|
| |
inputs
PiperOrigin-RevId: 199776409
|
|
|
|
|
|
|
|
| |
tensorflow::str_util equivalents.
This will allow the deprecated methods to be removed.
PiperOrigin-RevId: 191627087
|
|
|
|
|
|
| |
sets the maximum size of each created constant.
PiperOrigin-RevId: 191523208
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
1. Tensor names in TF may be in the form of "a:0", "a:1", or "a" as a shorthand
notation of "a:0". FoldConstant library always expected the shorthand notation,
and did not handle the cases where explicit notation was passed to input or
output list. This means that this library could not handle the case when input
or output were not the first output of a node.
2. To match the input nodes in the original graph and the added Recv nodes in
rewritten graph, FoldConstant library used prefix matching. Unfortunately, this
means that when a input name is a prefix of another input name, there is
possibility that wrong Recv node gets matched. For example, if input names were
"placeholder" and "placeholder_1", then it did not handle the case very well.
3. RemoveUnusedNodes() in FoldConstants lib could remove nodes which output
depended on. This happened when an input name points to a node with multiple
outputs and not all outputs of that node were included in the input names.
4. ReplaceSendRecvs() in FoldConstants lib assumed that all input nodes are
removed during rewriting the graph. This assumption is not necessarily true,
and it could add a duplicate node in the graph.
PiperOrigin-RevId: 172641947
|
|
|
|
|
|
|
|
|
| |
tools/graph_transforms.
By setting this option to false, the transformer will not strip off the shape
information stored as attributes.
PiperOrigin-RevId: 172057283
|
|
|
|
| |
PiperOrigin-RevId: 170818644
|
|
|
|
| |
PiperOrigin-RevId: 162328390
|
|
|
|
| |
PiperOrigin-RevId: 155511618
|
|
|
|
| |
Change: 142729497
|
|
It's often useful to preprocess graphs to fold constant expression subgraphs
into single constant nodes, especially as preparation for further processing for
purposes like batch normalization folding. This change uses the built-in constant
folding in the runtime to rewrite frozen models on disk.
Change: 135440279
|