|View source on GitHub|
TPU related configuration required by
iterations_per_loop: This is the number of train steps running in TPU system before returning to CPU host for each
Session.run. This means global step is increased
iterations_per_looptimes in one
Session.run. It is recommended to be set as number of global steps for next checkpoint. Note that in evaluation don't use this value, instead we run total eval
stepson TPU for a single
iterations_per_loopcan be specified as a time interval. To specify N seconds in one
Session.run, one can specify it as
Nsand substitute the N with the N with the number of desired seconds. Alternatively, the unit of time can also be specified in minutes or hours, e.g.
num_shards: (Deprecated, ignored by TPUEstimator). The number of model replicas in the system. For non-model-parallelism case, this number equals the total number of TPU cores. For model-parallelism, the total number of TPU cores equals num_cores_per_replica * num_shards.
num_cores_per_replica: Defaults to
None, which disables model parallelism. An integer which describes the number of TPU cores per model replica. This is required by model-parallelism which enables partitioning the model to multiple cores. Currently num_cores_per_replica must be 1, 2, 4, or 8.
input_fnis invoked once on each host. With the per-core input pipeline configuration, it is invoked once for each core. With a global batch size
TPUEstimatorconstructor, the batch size for each shard is
train_batch_size// #hosts in the
PER_HOST_V2mode, it is
train_batch_size// #cores. In
input_fnis only invoked once on host 0 and the tensors are broadcasted to all other replicas. The batch size equals to
train_batch_size. With the per-core input pipeline configuration, the shard batch size is also
train_batch_size// #cores. Note: per_host_input_for_training==PER_SHARD_V1 only supports mode.TRAIN.
tpu_job_name: The name of the TPU job. Typically, this name is auto-inferred within TPUEstimator, however when using ClusterSpec propagation in more esoteric cluster configurations, you may need to specify the job name as a string.
initial_infeed_sleep_secs: The number of seconds the infeed thread should wait before enqueueing the first batch. This helps avoid timeouts for models that require a long compilation time.
input_partition_dims: A nested list to describe the partition dims for all the tensors from input_fn(). The structure of input_partition_dims must match the structure of
labelsfrom input_fn(). The total number of partitions must match
num_cores_per_replica. For example, if input_fn() returns two tensors: images with shape [N, H, W, C] and labels [N]. input_partition_dims = [[1, 2, 2, 1], None] will split the images to 4 pieces and feed into 4 TPU cores. labels tensor are directly broadcasted to all the TPU cores since the partition dims is
None. Current limitations: This feature is only supported with the PER_HOST_V2 input mode.
input_fnis only invoked once on host 0 and the tensors are broadcasted to all other replicas. Unlike per_host_input_for_training=BROADCAST, each replica will only get a slice of the data instead of a whole copy. If
PER_HOST_V1, the behaviour is determined by per_host_input_for_training.
num_cores_per_replicais not 1, 2, 4, 8 or 16.