PiperOrigin-RevId: 296572040
Change-Id: If0e24b44d0917ac4a060761f76e64c207d767774

Add a test for simplified ResNet model that verifies NCHW->NHWC conversion.

PiperOrigin-RevId: 296567494
Change-Id: I14031649ec47015bfcad219cdcea23f7bcffb156

imort foo` work with autocomplete.

PiperOrigin-RevId: 296565836
Change-Id: I6d520018778ee5dd5c32f299e65bc4101db1bf7c

This change avoids constructing `InstantiateOptions` unless we actually need to instantiate the function (on its first use), and rearranges things to avoid copying so many strings and structures.

PiperOrigin-RevId: 296565724
Change-Id: Ieaa5934f0debc7fe2c722e823185d5f958993cc3

PiperOrigin-RevId: 296559221
Change-Id: I07442f9296dd02f2a6ef46969a3efc475c2235e5

PiperOrigin-RevId: 296558792
Change-Id: I393cf5666956ec7185458705d8e166ec97457a51

number of copies by 2.

We do the following in this CL.

1) Get rid of the snappy input and output buffers that were previously being
used to do the compression. This saved one copy.
2) Directly decompress the compressed bytes into the TensorBuffer for simple
types (not string, variant, resource). This saves another copy during Tensor
creation. For complex types, we still continue to use the TensorProto encoding
and pay a copy there.
3) As a result, we end up changing the on-disk format for Snapshot. For a group
of tensors that make up one element of an IteratorGetNext output, we first
write out a metadata proto that describes the types, shapes and sizes of
tensors. After that we lay out the Tensor data (TensorBuffers for simple types
and TensorProtos serialized for complex ones) and compress them via snappy.
4) Add a version to the SnapshotMetadata. If it isn't set its assumed to be 0
and the old code path runs. We now set it to 1 while writing so that all new
snapshots are written in this data format.

PiperOrigin-RevId: 296554803
Change-Id: I2590769f534dc3ca8211058e0b67475af728a3da

Makes Model compile run inside the captured distribution strategy, so that any metrics/optimizers created by compile are created in the distribution strategy scope (e.g. when deserializing strings that are the metric names).

Also adds a correctness test that verifies the model successfully captures the distribution strategy.

It also raises an error if there are metrics that are created outside the scope and not in compile.

So e.g. this will help the following case because the optimizer/metrics get created by compile:
with strategy.scope():
  model = ...
model.compile(optimizer='sgd', metrics=['binary_accuracy'])

And, it will raise an error in the following case because the metrics are created in a different distribution strategy scope than the model:
with strategy.scope():
  model = ...
model.compile(optimizer=tf.keras.optimizers.Blah() metrics=[tf.keras.metrics.BinaryAccuracy()])

PiperOrigin-RevId: 296553610
Change-Id: I988a80c1863de732da45555c444fc5a237ecd425

PiperOrigin-RevId: 296553524
Change-Id: If334f904a3ad375e18a0e33915e289ee101a1b27

PiperOrigin-RevId: 296552557
Change-Id: I88c70037f907339d7df80d0e47de75dda555a86d

PiperOrigin-RevId: 296552480
Change-Id: Ic71122a8f285fc791764562fbc7309320c0c7394

PiperOrigin-RevId: 296550340
Change-Id: I3cb1cd4374804b84f09a64d0c32ddc7f9e1de28e

PiperOrigin-RevId: 296550182
Change-Id: I9e0d6c2ff2c63642bda3724513c06b65bb8b3b15

PiperOrigin-RevId: 296549986
Change-Id: Ieba3dd38868f6fa903adf86c0588baf0e3c7787d

This allows one to write code like this without one_hot raising a TypeError complaining that the float32 inputs don't match the dtype of float16:

tf.one_hot(indices=[0, 1, 2], depth=3, on_value=1., off_value=0., dtype=tf.float16)

Also fix issue with several tests, where previously they were mostly run in float32 even if the test was intended to run in a different dtype.

PiperOrigin-RevId: 296548854
Change-Id: I44881c5a5a007e255671d86808b73015d56dfa94

Simplify unit test code for fused_batch_norm and its gradient.

PiperOrigin-RevId: 296546311
Change-Id: Ieb78b4038a39dd2bcde16302541e733f4e8604bd

PiperOrigin-RevId: 296544701
Change-Id: I49ca867e15a9da92e2bdba6410da5ccb16a57440

PiperOrigin-RevId: 296541416
Change-Id: Ie6c32d0e24c08ad73b2d647b8c7fb65a8869e801

PiperOrigin-RevId: 296541308
Change-Id: I54d33a55791a79b1183f55b13f937c9d65aefd71

PiperOrigin-RevId: 296540786
Change-Id: If8a4c33ac876e92d4c9bfc0a65147d9ac5695143

PiperOrigin-RevId: 296538689
Change-Id: If6c267de658fe8aa44870b14af5b4880ea980a62

PiperOrigin-RevId: 296538507
Change-Id: I6351ce464efd94fd3bb745e78bee8cd1d60bf573

PiperOrigin-RevId: 296538329
Change-Id: Ie5ab37dee86543349fc092ee9130a8539e204202

Since remote workers can be added and removed, using the Device pointer
is not appropriate since it is possible for Device objects to be
deleted. Thus we switch the mirrors to key of the device incarnation.

In addition we remove SetRemoteOpIdAndOutputNumToLocalTensorHandle as it
is unused, and we add some const annotations.

PiperOrigin-RevId: 296538295
Change-Id: Icdffd40a5ed242c2c0f81fbede8a35a03bd79044

Currently we move a default-constructed tstring to the back of the parsed list before resizing and parsing into that object. This change avoids the move by "constructing" (where necessary) the string in place and mutating it directly.

PiperOrigin-RevId: 296538238
Change-Id: Ic5efdc83a378b198386e726531a121d3b81837fa

PiperOrigin-RevId: 296537652
Change-Id: I8fa58ba6cb3aa31aa79070ec1a2e3ca7b1fed68b

Simplify unit test code for fused_batch_norm and its gradient.

PiperOrigin-RevId: 296537647
Change-Id: I09543d18a68ac9673e1168a1d130554a37cce14e

PiperOrigin-RevId: 296536900
Change-Id: I3394fdf582c259f1cf61fce0130555e01f97706e

PiperOrigin-RevId: 296535607
Change-Id: I304d467a00c4880bd6648c918bdcc39cc455101f

PiperOrigin-RevId: 296532815
Change-Id: Ibacf4929fc689dce87433a4e063b923b24ef9773

PiperOrigin-RevId: 296532794
Change-Id: I466c2c579a73390e8de1f69fadf7d3442d169e55

PiperOrigin-RevId: 296531864
Change-Id: I6310a939ac43ffabb356196a14df58f887adae3b

PiperOrigin-RevId: 296530052
Change-Id: Ieab047fb73d7b15c26d716bd43c18b8187acc3e0

This adds support for 'num_cores_per_replica' > 1. Device assignment is determined, in this case, with additional attributes 'topology' (TopologyProto) and 'device_assignment' (list of device coordinates), and creating a mapping between the two alongside available TPU devices. A test case in TPU rewrite is removed and will be updated once TPU rewrite pass is updated to support model parallelism properly.

PiperOrigin-RevId: 296528720
Change-Id: Ic59fc9b0337fdb47ca87d08c978ed89c1bb0d1ba

 - Shuffles the accelerometer data axes so that they match the training data orientation with the Arduino mounted as desired.

 - Replaced the run-based post-processing with averaging over time.

 - Reverted the model data back to the original version.

 - Replaced the output buffer resetting with a detection suppression window after a gesture is found.

 - Removed the wait on the serial port, and added LED indicators for 'ring' and 'slope' (which seem to be the most accurately-detected gestures).

For an Arduino library zip file based off this change, see https://storage.googleapis.com/download.tensorflow.org/deps/tflite/arduino_library/tensorflow_lite_2020_02_09.zip

PiperOrigin-RevId: 296528242
Change-Id: I5219484de8f003bdac5fdd75bd6ec65ee51006e2

Fix the tenses used.

PiperOrigin-RevId: 296526030
Change-Id: I1eace9e6a477d303f7c2edef0cfdaf3460279e98

$TPU_NAME and $TPU_ZONE aren't set until after ctpu_up is run.

PiperOrigin-RevId: 296524196
Change-Id: I2d8bf8885e5d6cd236db234244fb08a8f0373c5c

to avoid cyclic dependencies

PiperOrigin-RevId: 296522872
Change-Id: I39922b5e92bfe748cbb2a3c93b9c96756cf5c527

PiperOrigin-RevId: 296521494
Change-Id: Ife15f953f44a36cb7e2b5cb5e4430f78d3460f94

PiperOrigin-RevId: 296520517
Change-Id: I11b94874383c60cdb8fd39cea3e208931cdafcf4