Available TensorFlow Ops

This page lists the TensorFlow Python APIs and graph operators available on Cloud TPU.

Available Python APIs

The list below is a guide to the set of available TensorFlow Python APIs. This list is not exhaustive. Library functions not on this list may work if they are composed of available primitives.

See the performance guide for recommendations about specific operators.

Module Available Python API Comments
tf tf.abs
tf.acosh
tf.add
tf.add_n
tf.angle
tf.arg_max dimension argument must be a compile-time constant.
tf.arg_min dimension argument must be a compile-time constant.
tf.asinh
tf.assign Available only for resource variable.
tf.assign_add Available only for resource variable.
tf.assign_sub Available only for resource variable.
tf.atan
tf.atan2
tf.atanh
tf.batch_to_space crops and block_shape arguments must be compile-time constant.
tf.batch_to_space_nd crops argument must be a compile-time constant.
tf.broadcast_dynamic_shape
tf.broadcast_static_shape
tf.case Experimental (control-flow). May not work reliably yet.
tf.cast
tf.ceil
tf.cholesky Experimental. May have issues with numerical precision.
tf.cholesky_solve Experimental. May have issues with numerical precision.
tf.clip_by_average_norm
tf.clip_by_global_norm
tf.clip_by_norm
tf.clip_by_value
tf.complex
tf.concat concat_dim must be a compile-time constant.
tf.cond Experimental (control-flow). May not work reliably yet.
tf.conj
tf.constant
tf.convert_to_tensor
tf.cos
tf.cosh
tf.cross
tf.cumprod axis must be a compile-time constant.
tf.cumsum axis must be a compile-time constant.
tf.depth_to_space
tf.diag
tf.diag_part
tf.div int32 division is slower than other types.
tf.divide int32 division is slower than other types.
tf.dynamic_stitch indices must be a compile-time constant.
tf.einsum
tf.equal
tf.erf
tf.erfc
tf.exp
tf.expand_dims dims must be a compile-time constant.
tf.expm1
tf.extract_image_patches
tf.eye
tf.fake_quant_with_min_max_args
tf.fake_quant_with_min_max_args_gradient
tf.fake_quant_with_min_max_vars
tf.fake_quant_with_min_max_vars_gradient
tf.fft
tf.fft2d
tf.fft3d
tf.fill dims argument must be a compile-time constant.
tf.floor
tf.floordiv
tf.floormod
tf.foldl Experimental (control-flow).
tf.foldr Experimental (control-flow).
tf.gather axis must be a compile-time constant.
tf.gather_nd
tf.greater
tf.greater_equal
tf.hessians Experimental (control-flow.
tf.identity
tf.identity_n
tf.ifft
tf.ifft2d
tf.ifft3d
tf.imag
tf.invert_permutation x argument must be a compile-time constant.
tf.is_finite
tf.is_inf
tf.is_nan
tf.is_non_decreasing
tf.is_strictly_increasing
tf.less