PiperOrigin-RevId: 286401971
Change-Id: I3451b6d464c6cade88830a9dbccead8387988173

PiperOrigin-RevId: 286395088
Change-Id: I91015bcde147b90457a8a20ac4d60e77bede0254

PiperOrigin-RevId: 286394331
Change-Id: I139c64cfc1b4bc262417db4bb27f332e67250dd1

Prior to this change qint*, quint* and resource were defined in Python
as a single-field struct arrays, e.g.

_np_qint8 = np.dtype([("qint8", np.int8)])

Several TensorFlow functions had to special-case struct arrays (which
otherwise have type NPY_VOID) and infer the real dtype from struct fields.

Having these data types defined and handled in C++ allows to minimize magic
on the Python/C++ boundary.

Note 1: The defined data types are *not* registered because NumPy requires
every registered data type to also register casting functions and UFunc
specializations which will require significantly more code.

Note 2: Tensor->NumPy conversion for qint*, quint* and resource dtypes does
not use the defined types, so

tf.constant([42], dtype=tf.qint8).numpy()

returns an array of dtype tf.int8. This behavior is unaffected by this change.

PiperOrigin-RevId: 286392791
Change-Id: I0bdc55c2002195bca6273c94d3965d6620239985

Pulling out kExternalStatePolicy as its own const char instead of re-using the DatasetToGraphOp constant

PiperOrigin-RevId: 286386422
Change-Id: I4432145c3bbe4903ff121b3e0fb9b488af30a610

PiperOrigin-RevId: 286384508
Change-Id: I67a3997ba67a7b474a9e1157f6f3fb12dbe84007

Handling of tag sets via string join and split leads to convert an empty list
into a list with empty string. Instead of fixing the get_meta_graph_def to receive
a list of tags instead of a string and move the issue to the caller of it,
get_meta_graph now ignores tags that are empty.

Additionally change print of the tag sets so use '%r' so that it is clear there is
a tag set (i.e. '') instead of printing an empty line.

PiperOrigin-RevId: 286375150
Change-Id: I8b6b17c8c1a30d76bbebfbfffdf064e367e526f9

PiperOrigin-RevId: 286371061
Change-Id: I2817748cc82f745cae8cc71f6d1f44dd7d7ba6cc

PiperOrigin-RevId: 286368500
Change-Id: Ia83374fbb5b8a205ecd27ff45c59111777d2c703

PiperOrigin-RevId: 286365254
Change-Id: Ia0c2482f4f264f36fe61db5f9c72f24db35faf65

PiperOrigin-RevId: 286358256
Change-Id: Iaef36feff46676492a14a8c5192f5ff724589569

Added a helper method to compute paddings for the ReduceWindow HLO op. This could be moved to the op definition after handling dilations attributes as well.

PiperOrigin-RevId: 286350876
Change-Id: Ibf91eb35887de184853da4ba6498ee8d814907e1

PiperOrigin-RevId: 286350492
Change-Id: Ib3ee849657948b8b197bb40f41376531e8739071

PiperOrigin-RevId: 286346436
Change-Id: Icf22652caed99cbef2cea01a3d0fdd473d2fedc1

PiperOrigin-RevId: 286341322
Change-Id: I8e5003d44210d185e3b3314cff357b7d1cc41628

PiperOrigin-RevId: 286335164
Change-Id: I66793be1e5277425f6da6dd519b256e18d61d071

PiperOrigin-RevId: 286332862
Change-Id: Ifa6d6354f8d775f3ab42f77018e344aff27adee0

PiperOrigin-RevId: 286331612
Change-Id: Ifbf078bffb74b8964699deec9d566e01538f2262

PiperOrigin-RevId: 286330416
Change-Id: I5878ab23b9596ac0c5dbf91394ee6ba8e5683c77

PiperOrigin-RevId: 286325675
Change-Id: Ie6e881ce51dbbfeee62ae80deeed6ad1a211e949

PiperOrigin-RevId: 286321206
Change-Id: Iae1a39ce283cdb9b3c0e09284e4a74eba178d77c

PiperOrigin-RevId: 286320322
Change-Id: I69bc815219e3ee66792ed5ab991461673bc68c06

PiperOrigin-RevId: 286320192
Change-Id: I13385f1cc7a71af835aef5663ccf488748674a96

This is mostly the result of an internal cleanup and formatting pass.

PiperOrigin-RevId: 286318018
Change-Id: I8f9e2f7519070035da73f9f24d2fc90864abc51b

PiperOrigin-RevId: 286317185
Change-Id: I8307567b315bcea4056a37bd9eb374861db20f32

Handle ExternalStatePolicy during Iterator::Save(). This CL adds an optional external_state_policy attr to the SerializeIterator op which is by default set to FAIL thereby keeping the behavior the same as before. Now users can override it if they want something else. This CL also exposes this option in the CheckpointInputPipelineHook.

PiperOrigin-RevId: 286314588
Change-Id: I4f8eae88f42c74b168348afca8b295eb37d845dc

PiperOrigin-RevId: 286313739
Change-Id: I9d4dfb66a8e82c56249bc7f7b70c71d228912e9f

Verify the early stopping condition in for loops before consuming the target, rather than just at the beginning of the iteration.

PiperOrigin-RevId: 286311186
Change-Id: Ie2afdd2a9f26d5045baf4517141517c0d28e3879

PiperOrigin-RevId: 286311054
Change-Id: I1225b7ed8f6ca83368dac2bb1572b6abc7945fd3

PiperOrigin-RevId: 286310575
Change-Id: I5027b42f4a620c15420c591bcc4a0909466da7fc

In all cases, we can't rely on the tensor.device attribute being set. So its better to get the device for a SaveSpec from the device passed in rather. This was an issue with saving iterators because for iterators the resource usually has a device specification but the serialized tensor derived from it might not have it set. As a result, when saving iterators in a sharded fashion all iterators end up on '' device instead which is not what is intended.

Also adding support for saving iterators in a sharded fashion to avoid unnecessary copying during checkpointing.

PiperOrigin-RevId: 286310419
Change-Id: I1a957af783f7f69753992ce220b59eb43df2c02f

Adding a define for TENSORFLOW_LITE_PROTOS; this gets the android build working with the granular target form core/platform:env.

PiperOrigin-RevId: 286308504
Change-Id: I6dd05fbcc62000836bc03866c07677cde33b1652

PiperOrigin-RevId: 286308295
Change-Id: I3df054ffbab15a1db933c6242c68e7fbc0b415e5

PiperOrigin-RevId: 286307709
Change-Id: Ib936279d64e8faf8e452b95a08a9c4a3c3026c4a

PiperOrigin-RevId: 286305667
Change-Id: I4b016ac41492d52f361a121f2c303cdd753fbae8

Use newly added LookUp call.

PiperOrigin-RevId: 286305459
Change-Id: Ibb5294ce98904bac6abbd902b0109715fffa34be

PiperOrigin-RevId: 286305113
Change-Id: I43539ef03478ad19a8329362d74f02c6e6035db8