可用的 TensorFlow 操作

此页面列出 Cloud TPU 上可用的 TensorFlow Python API 和图运算符。

可用的 Python API

以下列表是可用的 TensorFlow Python API 集合的指南。该列表并不详尽。如果未在此列表中的库函数由可用基元组成，则可能也可以运行。

如需关于特定运算符的建议，请参阅性能指南。

模块	可用的 Python API	注释
`tf`	`tf.abs`
	`tf.acosh`
	`tf.add`
	`tf.add_n`
	`tf.angle`
	`tf.arg_max`	`dimension` 参数必须是编译时常量。
	`tf.arg_min`	`dimension` 参数必须是编译时常量。
	`tf.asinh`
	`tf.assign`	仅适用于 resource 变量。
	`tf.assign_add`	仅适用于 resource 变量。
	`tf.assign_sub`	仅适用于 resource 变量。
	`tf.atan`
	`tf.atan2`
	`tf.atanh`
	`tf.batch_to_space`	`crops` 和 `block_shape` 参数必须是编译时常量。
	`tf.batch_to_space_nd`	`crops` 参数必须是编译时常量。
	`tf.broadcast_dynamic_shape`
	`tf.broadcast_static_shape`
	`tf.case`	实验性（控制流）。目前可能还无法可靠地运行。
	`tf.cast`
	`tf.ceil`
	`tf.cholesky`	实验性。可能存在数值精度问题。
	`tf.cholesky_solve`	实验性。可能存在数值精度问题。
	`tf.clip_by_average_norm`
	`tf.clip_by_global_norm`
	`tf.clip_by_norm`
	`tf.clip_by_value`
	`tf.complex`
	`tf.concat`	`concat_dim` 必须是编译时常量。
	`tf.cond`	实验性（控制流）。目前可能还无法可靠地运行。
	`tf.conj`
	`tf.constant`
	`tf.convert_to_tensor`
	`tf.cos`
	`tf.cosh`
	`tf.cross`
	`tf.cumprod`	`axis` 必须是编译时常量。
	`tf.cumsum`	`axis` 必须是编译时常量。
	`tf.depth_to_space`
	`tf.diag`
	`tf.diag_part`
	`tf.div`	`int32` 除法慢于其他类型。
	`tf.divide`	`int32` 除法慢于其他类型。
	`tf.dynamic_stitch`	`indices` 必须是编译时常量。
	`tf.einsum`
	`tf.equal`
	`tf.erf`
	`tf.erfc`
	`tf.exp`
	`tf.expand_dims`	`dims` 必须是编译时常量。
	`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` 参数必须是编译时常量。
	`tf.floor`
	`tf.floordiv`
	`tf.floormod`
	`tf.foldl`	实验性（控制流）。
	`tf.foldr`	实验性（控制流）。
	`tf.gather`	`axis` 必须是编译时常量。
	`tf.gather_nd`
	`tf.greater`
	`tf.greater_equal`
	`tf.hessians`	实验性（控制流）。
	`tf.identity`
	`tf.identity_n`
	`tf.ifft`
	`tf.ifft2d`
	`tf.ifft3d`
	`tf.imag`
	`tf.invert_permutation`	`x` 参数必须是编译时常量。
	`tf.is_finite`
	`tf.is_inf`
	`tf.is_nan`
	`tf.is_non_decreasing`
	`tf.is_strictly_increasing`
	`tf.less`
	`tf.less_equal`
	`tf.linspace`	`start`、`stop` 和 `num` 参数必须是编译时常量。
	`tf.log`
	`tf.log1p`
	`tf.log_sigmoid`
	`tf.logical_and`
	`tf.logical_or`
	`tf.logical_not`
	`tf.logical_xor`
	`tf.matmul`	使用带有 `float32` 累积的 `bfloat16` matmul 函数。
	`tf.matrix_band_part`
	`tf.matrix_diag`
	`tf.matrix_diag_part`
	`tf.matrix_set_diag`
	`tf.matrix_triangular_solve`	实验性。可能存在数值精度问题。
	`tf.maximum`
	`tf.meshgrid`
	`tf.minimum`
	`tf.mod`
	`tf.multinomial`	`num_samples` 参数必须是编译时常量。
	`tf.multiply`
	`tf.negative`
	`tf.no_op`
	`tf.norm`
	`tf.not_equal`
	`tf.one_hot`	`depth` 必须是编译时常量。
	`tf.ones`
	`tf.ones_like`
	`tf.pad`	`paddings` 参数必须是编译时常量。`REFLECT` 填充的梯度尚不可用。
	`tf.pow`
	`tf.random_normal`	`shape` 必须是编译时常量。
	`tf.random_uniform`	`shape` 必须是编译时常量。
	`tf.range`	`start`、`limit` 和 `delta` 参数必须是编译时常量。
	`tf.rank`
	`tf.real`
	`tf.realdiv`
	`tf.reciprocal`
	`tf.reduce_all`	`axis` 必须是编译时常量。
	`tf.reduce_any`	`axis` 必须是编译时常量。
	`tf.reduce_logsumexp`
	`tf.reduce_max`	`axis` 必须是编译时常量。
	`tf.reduce_min`	`axis` 必须是编译时常量。
	`tf.reduce_prod`	`axis` 必须是编译时常量。
	`tf.reduce_sum`	`axis` 必须是编译时常量。
	`tf.reshape`	`shape` 参数必须是编译时常量。
	`tf.reverse`	`dims` 参数必须是编译时常量。
	`tf.reverse_sequence`
	`tf.reverse_v2`	`axis` 参数必须是编译时常量。
	`tf.rint`
	`tf.round`
	`tf.rsqrt`
	`tf.saturate_cast`
	`tf.scalar_mul`
	`tf.scan`	实验性（控制流）。
	`tf.scatter_nd`
	`tf.sequence_mask`
	`tf.shape`
	`tf.shape_n`
	`tf.sigmoid`
	`tf.sign`
	`tf.sin`
	`tf.sinh`
	`tf.size`
	`tf.slice`	`size` 必须是编译时常量。此外，要么 `begin` 必须是编译时常量，要么 `size` 必须为非负。只有当 `begin` 和 `size` 是编译时常量时才支持反向传播。
	`tf.space_to_batch`	`paddings` 和 `block_shape` 必须是编译时常量。
	`tf.space_to_batch_nd`	`paddings` 必须是编译时常量。
	`tf.space_to_depth`
	`tf.split`	`axis` 必须是编译时常量。
	`tf.sqrt`
	`tf.square`
	`tf.squared_difference`
	`tf.squeeze`
	`tf.stack`
	`tf.stop_gradient`
	`tf.strided_slice`
	`tf.tan`
	`tf.tanh`
	`tf.tensordot`
	`tf.tile`	`multiples` 参数必须是编译时常量。
	`tf.to_bfloat16`
	`tf.to_float`
	`tf.to_int32`
	`tf.to_int64`	`int64` 支持有限。
	`tf.trace`
	`tf.transpose`	`perm` 参数必须是编译时常量。
	`tf.truediv`
	`tf.truncated_normal`	`shape` 必须是编译时常量。
	`tf.truncatediv`
	`tf.truncatemod`
	`tf.unsorted_segment_sum`
	`tf.unstack`
	`tf.where`	`x` 和 `y` 都必须为非 `None`。如果 `x` 和 `y` 均为 `None`，则运算符不具备静态形状。
	`tf.while_loop`	计算 while 循环的梯度需要传递 `maximum_iterations` 参数。
	`tf.zeros`
	`tf.zeros_like`
	`tf.Tensor.__getitem__`	切片的开始、结束和步幅都必须是编译时常量。
`tf.bitwise`	`tf.bitwise_and`
	`tf.bitwise_or`
	`tf.bitwise_invert`
`tf.contrib.stateless`	`tf.contrib.stateless.stateless_random_normal`
	`tf.contrib.stateless.stateless_random_uniform`
`tf.image`	`tf.image.adjust_brightness`
	`tf.image.adjust_contrast`
	`tf.image.adjust_gamma`
	`tf.image.adjust_hue`
	`tf.image.adjust_saturation`
	`tf.image.central_crop`	剪裁系数必须是编译时常量。
	`tf.image.convert_image_dtype`
	`tf.image.flip_left_right`
	`tf.image.flip_up_down`
	`tf.image.grayscale_to_rgb`
	`tf.image.hsv_to_rgb`
	`tf.image.resize_bilinear`	只有 `align_corners=True` 可用。`size` 必须是编译时常量。
	`tf.image.random_brightness`
	`tf.image.random_contrast`
	`tf.image.random_flip_left_right`
	`tf.image.random_flip_up_down`
	`tf.image.random_hue`
	`tf.image.random_saturation`
	`tf.image.rgb_to_hsv`
	`tf.image.rgb_to_grayscale`
	`tf.image.rot90`
	`tf.image.total_variation`
	`tf.image.transpose_image`
`tf.layers`	`tf.layers.average_pooling1d`
	`tf.layers.average_pooling2d`
	`tf.layers.average_pooling1d`
	`tf.layers.batch_normalization`
	`tf.layers.conv1d`
	`tf.layers.conv2d`
	`tf.layers.conv2d_transpose`
	`tf.layers.conv3d`
	`tf.layers.conv3d_transpose`
	`tf.layers.dense`
	`tf.layers.dropout`
	`tf.layers.flatten`
	`tf.layers.max_pooling1d`
	`tf.layers.max_pooling2d`
	`tf.layers.max_pooling3d`
	`tf.layers.separable_conv2d`
`tf.nn`	`tf.nn.atrous_conv2d`
	`tf.nn.atrous_conv2d_transpose`
	`tf.nn.avg_pool`
	`tf.nn.avg_pool3d`
	`tf.nn.batch_normalization`
	`tf.nn.bias_add`
	`tf.nn.conv1d`
	`tf.nn.conv2d`
	`tf.nn.conv2d_backprop_filter`
	`tf.nn.conv2d_backprop_input`
	`tf.nn.conv2d_transpose`
	`tf.nn.conv3d`
	`tf.nn.conv3d_backprop_filter`
	`tf.nn.conv3d_backprop_input`
	`tf.nn.conv3d_transpose`
	`tf.nn.convolution`
	`tf.nn.crelu`
	`tf.nn.depthwise_conv2d`
	`tf.nn.depthwise_conv2d_native`
	`tf.nn.depthwise_conv2d_native_backprop_filter`
	`tf.nn.depthwise_conv2d_native_backprop_input`
	`tf.nn.dropout`
	`tf.nn.dynamic_rnn`	实验性。
	`tf.nn.elu`
	`tf.nn.fused_batch_norm`
	`tf.nn.l2_loss`
	`tf.nn.l2_normalize`
	`tf.nn.leaky_relu`
	`tf.nn.local_response_normalization`
	`tf.nn.log_poisson_loss`
	`tf.nn.log_softmax`
	`tf.nn.max_pool`
	`tf.nn.max_pool3d`
	`tf.nn.moments`
	`tf.nn.normalize_moments`
	`tf.nn.pool`
	`tf.nn.relu`
	`tf.nn.relu6`
	`tf.nn.relu_layer`
	`tf.nn.selu`
	`tf.nn.separable_conv2d`
	`tf.nn.sigmoid_cross_entropy_with_logits`
	`tf.nn.softmax`
	`tf.nn.softmax_cross_entropy_with_logits`
	`tf.nn.softplus`
	`tf.nn.softsign`
	`tf.nn.sparse_softmax_cross_entropy_with_logits`
	`tf.nn.static_bidirectional_rnn`	实验性。
	`tf.nn.static_rnn`	实验性。
	`tf.nn.weighted_cross_entropy_with_logits`	实验性。
	`tf.nn.weighted_moments`
	`tf.nn.with_space_to_batch`
	`tf.nn.xw_plus_b`
	`tf.nn.zero_fraction`
`tf.spectral`	`tf.spectral.fft`
	`tf.spectral.fft2d`
	`tf.spectral.fft3d`
	`tf.spectral.ifft`
	`tf.spectral.ifft2d`
	`tf.spectral.ifft3d`
	`tf.spectral.irfft`	`fft_length` 必须是编译时常量。
	`tf.spectral.irfft2d`	`fft_length` 必须是编译时常量。
	`tf.spectral.irfft3d`	`fft_length` 必须是编译时常量。
	`tf.spectral.rfft`	`fft_length` 必须是编译时常量。
	`tf.spectral.rfft2d`	`fft_length` 必须是编译时常量。
	`tf.spectral.rfft3d`	`fft_length` 必须是编译时常量。

不可用的 Python API

该列表并不详尽。Cloud TPU 上不可用的操作包括以下内容：

模块	不可用的 Python API	注释
`tf`	`tf.accumulate_n`	使用 Ref 变量。
	`tf.acos`
	`tf.asin`
	`tf.betainc`
	`tf.bitcast`
	`tf.add_check_numerics_ops`	包含 check numerics 运算符的程序应当可以运行，但是现在会忽略 check numerics 运算符。
	`tf.assert_...`	包含断言的程序应当可以运行，但会忽略断言。
	`tf.check_numerics`	包含 check numerics 运算符的程序应当可以运行，但是现在会忽略 check numerics 运算符。
	`tf.confusion_matrix`
	`tf.count_nonzero`	使用 `int64` 归约。
	`tf.count_up_to`
	`tf.create_partitioned_variables`
	`tf.dequantize`
	`tf.digamma`
	`tf.dynamic_partition`
	`tf.edit_distance`
	`tf.fake_quant_with_min_max_vars_per_channel`
	`tf.fake_quant_with_min_max_vars_per_channel_gradient`
	`tf.histogram_fixed_width`
	`tf.igamma`
	`tf.igammac`
	`tf.lbeta`
	`tf.lgamma`
	`tf.matrix_determinant`
	`tf.matrix_inverse`
	`tf.matrix_solve`
	`tf.matrix_solve_ls`
	`tf.polygamma`
	`tf.py_func`
	`tf.qr`
	`tf.quantize_v2`
	`tf.quantized_concat`
	`tf.random_crop`
	`tf.random_gamma`
	`tf.random_poisson`
	`tf.random_shuffle`
	`tf.scatter_add`
	`tf.scatter_div`
	`tf.scatter_mul`
	`tf.scatter_nd_add`
	`tf.scatter_nd_sub`
	`tf.scatter_nd_update`
	`tf.segment_mean`
	`tf.segment_max`
	`tf.segment_min`
	`tf.segment_prod`
	`tf.segment_sum`
	`tf.self_adjoint_eig`
	`tf.self_adjoint_eigvals`
	`tf.setdiff1d`
	`tf.sparse_...`
	`tf.string_...`
	`tf.substr`
	`tf.svd`
	`tf.to_double`
	`tf.unique`
	`tf.unsorted_segment_max`
	`tf.zeta`
	`tf.bitwise.bitwise_xor`
	`tf.contrib.stateless.stateless_truncated_normal`

可用的图运算符

运算符	类型约束
`Abs`	`T={bfloat16,float,int32,int64}`
`Acos`	`T={bfloat16,complex64,float,int32,int64}`
`Acosh`	`T={bfloat16,complex64,float}`
`Add`	`T={bfloat16,complex64,float,int32,int64}`
`AddN`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`AdjustContrastv2`	`T={float}`
`AdjustHue`	`T={float}`
`AdjustSaturation`	`T={float}`
`All`	`Tidx={int32,int64}`
`AllToAll`	`T={bfloat16,float}`
`Angle`	`Tout={float}` `T={complex64}`
`Any`	`Tidx={int32,int64}`
`ApproximateEqual`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ArgMax`	`Tidx={int32,int64}` `output_type={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ArgMin`	`Tidx={int32,int64}` `output_type={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`Asin`	`T={bfloat16,complex64,float,int32,int64}`
`Asinh`	`T={bfloat16,complex64,float}`
`Assert`	`T={bfloat16,bool,complex64,float,int32,int64,string,uint32,uint64}`
`AssignAddVariableOp`	`dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`AssignSubVariableOp`	`dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`AssignVariableOp`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Atan`	`T={bfloat16,complex64,float,int32,int64}`
`Atan2`	`T={bfloat16,float}`
`Atanh`	`T={bfloat16,complex64,float}`
`AvgPool`	`T={bfloat16,float}`
`AvgPool3D`	`T={bfloat16,float}`
`AvgPool3DGrad`	`T={bfloat16,float}`
`AvgPoolGrad`	`T={bfloat16,float}`
`BatchMatMul`	`T={bfloat16,complex64,float,int32,int64}`
`BatchToSpace`	`Tidx={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`BatchToSpaceND`	`Tcrops={int32,int64}` `Tblock_shape={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`BiasAdd`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`BiasAddGrad`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`BiasAddV1`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`Bitcast`	`type={bfloat16,complex64,float,int32,int64,uint32,uint64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`BitwiseAnd`	`T={int32,int64,uint32,uint64}`
`BitwiseOr`	`T={int32,int64,uint32,uint64}`
`BitwiseXor`	`T={int32,int64,uint32,uint64}`
`BroadcastArgs`	`T={int32,int64}`
`BroadcastGradientArgs`	`T={int32,int64}`
`BroadcastTo`	`Tidx={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Bucketize`	`T={float,int32,int64}`
`Cast`	`DstT={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `SrcT={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Ceil`	`T={bfloat16,float}`
`CheckNumerics`	`T={bfloat16,float}`
`Cholesky`	`T={float}`
`ClipByValue`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`CollectivePermute`	`T={bfloat16,float}`
`Complex`	`Tout={complex64}` `T={float}`
`ComplexAbs`	`Tout={float}` `T={complex64}`
`Concat`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ConcatOffset`
`ConcatV2`	`Tidx={int32}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Conj`	`T={complex64}`
`ConjugateTranspose`	`Tperm={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Const`	`dtype={bfloat16,bool,complex64,float,int32,int64,string,uint32,uint64}`
`ControlTrigger`
`Conv2D`	`T={bfloat16,float}`
`Conv2DBackpropFilter`	`T={bfloat16,float}`
`Conv2DBackpropInput`	`T={bfloat16,float}`
`Conv3D`	`T={bfloat16,float}`
`Conv3DBackpropFilterV2`	`T={bfloat16,float}`
`Conv3DBackpropInputV2`	`Tshape={int32,int64}` `T={bfloat16,float}`
`Cos`	`T={bfloat16,complex64,float}`
`Cosh`	`T={bfloat16,complex64,float}`
`Cross`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`CrossReplicaSum`	`T={bfloat16,float}`
`Cumprod`	`Tidx={int32,int64}` `T={bfloat16,float,int32}`
`Cumsum`	`Tidx={int32,int64}` `T={bfloat16,float,int32}`
`DataFormatVecPermute`	`T={int32,int64}`
`DepthToSpace`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`DepthwiseConv2dNative`	`T={bfloat16,float}`
`DepthwiseConv2dNativeBackpropFilter`	`T={bfloat16,float}`
`DepthwiseConv2dNativeBackpropInput`	`T={bfloat16,float}`
`Diag`	`T={bfloat16,complex64,float,int32,int64}`
`DiagPart`	`T={bfloat16,complex64,float,int32,int64}`
`Digamma`	`T={bfloat16,float}`
`Div`	`T={bfloat16,complex64,float,int32,int64}`
`DivNoNan`	`T={float}`
`DynamicStitch`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Elu`	`T={bfloat16,float}`
`EluGrad`	`T={bfloat16,float}`
`Empty`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`EmptyTensorList`	`shape_type={int32,int64}` `element_dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Equal`	`T={bfloat16,bool,complex64,float,int32,int64}`
`Erf`	`T={bfloat16,float}`
`Erfc`	`T={bfloat16,float}`
`Exp`	`T={bfloat16,complex64,float}`
`ExpandDims`	`Tdim={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Expm1`	`T={bfloat16,complex64,float}`
`ExtractImagePatches`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`FFT`	`Tcomplex={complex64}`
`FFT2D`	`Tcomplex={complex64}`
`FFT3D`	`Tcomplex={complex64}`
`FakeParam`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`FakeQuantWithMinMaxArgs`
`FakeQuantWithMinMaxArgsGradient`
`FakeQuantWithMinMaxVars`
`FakeQuantWithMinMaxVarsGradient`
`Fill`	`index_type={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Floor`	`T={bfloat16,float}`
`FloorDiv`	`T={bfloat16,complex64,float,int32,int64}`
`FloorMod`	`T={bfloat16,float,int32,int64}`
`FusedBatchNorm`	`T={float}`
`FusedBatchNormGrad`	`T={float}`
`FusedBatchNormGradV2`	`U={float}` `T={bfloat16,float}`
`FusedBatchNormV2`	`U={float}` `T={bfloat16,float}`
`Gather`	`Tindices={int32,int64}` `Tparams={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`GatherNd`	`Tindices={int32,int64}` `Tparams={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`GatherV2`	`Taxis={int32,int64}` `Tindices={int32,int64}` `Tparams={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`GetItem`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Greater`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`GreaterEqual`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`HSVToRGB`	`T={bfloat16,float}`
`IFFT`	`Tcomplex={complex64}`
`IFFT2D`	`Tcomplex={complex64}`
`IFFT3D`	`Tcomplex={complex64}`
`IRFFT`
`IRFFT2D`
`IRFFT3D`
`Identity`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`IdentityN`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`If`	`Tout={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tin={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tcond={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`Imag`	`Tout={float}` `T={complex64}`
`InfeedDequeue`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`InfeedDequeueTuple`	`dtypes={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`InplaceAdd`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`InplaceUpdate`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Inv`	`T={bfloat16,complex64,float,int32,int64}`
`Invert`	`T={int32,int64,uint32,uint64}`
`InvertPermutation`	`T={int32}`
`IsFinite`	`T={bfloat16,float}`
`IsInf`	`T={bfloat16,float}`
`IsNan`	`T={bfloat16,float}`
`L2Loss`	`T={bfloat16,float}`
`LRN`	`T={bfloat16,float}`
`LRNGrad`	`T={bfloat16,float}`
`LeakyRelu`	`T={bfloat16,float}`
`LeakyReluGrad`	`T={bfloat16,float}`
`LeftShift`	`T={int32,int64,uint32,uint64}`
`Less`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`LessEqual`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`Lgamma`	`T={bfloat16,float}`
`LinSpace`	`Tidx={int32,int64}` `T={bfloat16,float}`
`ListDiff`	`out_idx={int32,int64}` `T={int32,int64}`
`Log`	`T={bfloat16,complex64,float}`
`Log1p`	`T={bfloat16,complex64,float}`
`LogSoftmax`	`T={bfloat16,float}`
`LogicalAnd`
`LogicalNot`
`LogicalOr`
`MatMul`	`T={bfloat16,complex64,float}`
`MatrixBandPart`	`Tindex={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`MatrixDiag`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`MatrixDiagPart`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`MatrixSetDiag`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`MatrixTriangularSolve`	`T={complex64,float}`
`Max`	`Tidx={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`MaxPool`	`T={bfloat16,float,int32,int64}`
`MaxPool3D`	`T={bfloat16,float}`
`MaxPool3DGrad`	`TInput={bfloat16,float}` `T={bfloat16,float}`
`MaxPool3DGradGrad`	`T={float}`
`MaxPoolGrad`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`MaxPoolGradGrad`	`T={float}`
`MaxPoolGradGradV2`	`T={float}`
`MaxPoolGradV2`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`MaxPoolV2`	`T={bfloat16,float,int32,int64}`
`Maximum`	`T={bfloat16,float,int32,int64}`
`Mean`	`Tidx={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`Min`	`Tidx={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`Minimum`	`T={bfloat16,float,int32,int64}`
`MirrorPad`	`Tpaddings={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Mod`	`T={bfloat16,float,int32,int64}`
`Mul`	`T={bfloat16,complex64,float,int32,int64}`
`Multinomial`	`output_dtype={int32,int64}` `T={bfloat16,float,int32,int64,uint32,uint64}`
`Neg`	`T={bfloat16,complex64,float,int32,int64}`
`NoOp`
`NonMaxSuppressionV4`	`T={float}`
`NotEqual`	`T={bfloat16,bool,complex64,float,int32,int64}`
`OneHot`	`TI={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`OnesLike`	`T={bfloat16,bool,complex64,float,int32,int64}`
`OutfeedEnqueue`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`OutfeedEnqueueTuple`	`dtypes={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Pack`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Pad`	`Tpaddings={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`PadV2`	`Tpaddings={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ParallelDynamicStitch`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`PlaceholderWithDefault`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Pow`	`T={bfloat16,complex64,float,int32,int64}`
`PreventGradient`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Prod`	`Tidx={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`Qr`	`T={float}`
`QuantizeAndDequantizeV2`	`T={bfloat16,float}`
`QuantizeAndDequantizeV3`	`T={bfloat16,float}`
`RFFT`
`RFFT2D`
`RFFT3D`
`RGBToHSV`	`T={bfloat16,float}`
`RandomShuffle`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`RandomStandardNormal`	`T={int32,int64}` `dtype={bfloat16,float}`
`RandomUniform`	`T={int32,int64}` `dtype={bfloat16,float}`
`RandomUniformInt`	`T={int32,int64}` `Tout={int32,int64}`
`Range`	`Tidx={bfloat16,float,int32,int64}`
`Rank`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ReadVariableOp`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Real`	`Tout={float}` `T={complex64}`
`RealDiv`	`T={bfloat16,complex64,float,int32,int64}`
`Reciprocal`	`T={bfloat16,complex64,float,int32,int64}`
`ReciprocalGrad`	`T={bfloat16,complex64,float}`
`RecvTPUEmbeddingActivations`
`Relu`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`Relu6`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`Relu6Grad`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`ReluGrad`	`T={bfloat16,float,int32,int64,uint32,uint64}`
`Reshape`	`Tshape={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResizeBilinear`	`T={bfloat16,float,int32,int64}`
`ResizeBilinearGrad`	`T={bfloat16,float}`
`ResizeNearestNeighbor`	`T={float,int32,int64}`
`ResourceApplyAdaMax`	`T={bfloat16,float}`
`ResourceApplyAdadelta`	`T={bfloat16,float}`
`ResourceApplyAdagrad`	`T={bfloat16,float}`
`ResourceApplyAdagradDA`	`T={bfloat16,float}`
`ResourceApplyAdam`	`T={bfloat16,float}`
`ResourceApplyAddSign`	`T={bfloat16,float}`
`ResourceApplyCenteredRMSProp`	`T={bfloat16,float}`
`ResourceApplyFtrl`	`T={bfloat16,float}`
`ResourceApplyFtrlV2`	`T={bfloat16,float}`
`ResourceApplyGradientDescent`	`T={bfloat16,float}`
`ResourceApplyKerasMomentum`	`T={bfloat16,float}`
`ResourceApplyMomentum`	`T={bfloat16,float}`
`ResourceApplyPowerSign`	`T={bfloat16,float}`
`ResourceApplyProximalAdagrad`	`T={bfloat16,float}`
`ResourceApplyProximalGradientDescent`	`T={bfloat16,float}`
`ResourceApplyRMSProp`	`T={bfloat16,float}`
`ResourceGather`	`Tindices={int32,int64}` `dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterAdd`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterDiv`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterMax`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterMin`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterMul`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterNdAdd`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterNdSub`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterNdUpdate`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterSub`	`Tindices={int32,int64}` `dtype={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`ResourceScatterUpdate`	`Tindices={int32,int64}` `dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ResourceStridedSliceAssign`	`Index={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Reverse`	`T={bool,complex64,float,int32,int64}`
`ReverseSequence`	`Tlen={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ReverseV2`	`T={bfloat16,bool,complex64,float,int32,int64}` `Tidx={int32,int64}`
`RightShift`	`T={int32,int64,uint32,uint64}`
`Rint`	`T={bfloat16,float}`
`Round`	`T={bfloat16,complex64,float,int32,int64}`
`Rsqrt`	`T={bfloat16,complex64,float}`
`RsqrtGrad`	`T={bfloat16,complex64,float}`
`ScatterNd`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Select`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Selu`	`T={bfloat16,float}`
`SeluGrad`	`T={bfloat16,float}`
`SendTPUEmbeddingGradients`
`Shape`	`out_type={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`ShapeN`	`out_type={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Sigmoid`	`T={bfloat16,complex64,float}`
`SigmoidGrad`	`T={bfloat16,complex64,float}`
`Sign`	`T={bfloat16,complex64,float,int32,int64}`
`Sin`	`T={bfloat16,complex64,float}`
`Sinh`	`T={bfloat16,complex64,float}`
`Size`	`out_type={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Slice`	`Index={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Snapshot`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Softmax`	`T={bfloat16,float}`
`SoftmaxCrossEntropyWithLogits`	`T={bfloat16,float}`
`Softplus`	`T={bfloat16,float}`
`SoftplusGrad`	`T={bfloat16,float}`
`Softsign`	`T={bfloat16,float}`
`SoftsignGrad`	`T={bfloat16,float}`
`SpaceToBatch`	`Tpaddings={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`SpaceToBatchND`	`Tblock_shape={int32,int64}` `Tpaddings={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`SpaceToDepth`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`SparseMatMul`	`Tb={bfloat16,float}` `Ta={bfloat16,float}`
`SparseSoftmaxCrossEntropyWithLogits`	`Tlabels={int32,int64}` `T={bfloat16,float}`
`SparseToDense`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Split`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`SplitV`	`Tlen={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Sqrt`	`T={bfloat16,complex64,float}`
`SqrtGrad`	`T={bfloat16,complex64,float}`
`Square`	`T={bfloat16,complex64,float,int32,int64}`
`SquaredDifference`	`T={bfloat16,complex64,float,int32,int64}`
`Squeeze`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StackCloseV2`
`StackPopV2`	`elem_type={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StackPushV2`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StackV2`	`elem_type={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StatelessIf`	`Tout={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tin={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tcond={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`StatelessMultinomial`	`output_dtype={int32,int64}` `Tseed={int32}` `T={bfloat16,float}`
`StatelessRandomNormal`	`Tseed={int32}` `T={int32,int64}` `dtype={bfloat16,float}`
`StatelessRandomUniform`	`Tseed={int32}` `T={int32,int64}` `dtype={bfloat16,float}`
`StatelessRandomUniformInt`	`Tseed={int32}` `T={int32,int64}` `dtype={int32,int64}`
`StatelessTruncatedNormal`	`Tseed={int32}` `T={int32,int64}` `dtype={bfloat16,float}`
`StatelessWhile`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`StopGradient`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StridedSlice`	`Index={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`StridedSliceGrad`	`Index={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Sub`	`T={bfloat16,complex64,float,int32,int64}`
`Sum`	`Tidx={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`SymbolicGradient`	`Tout={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `Tin={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TPUEmbeddingActivations`
`Tan`	`T={bfloat16,complex64,float,int32,int64}`
`Tanh`	`T={bfloat16,complex64,float}`
`TanhGrad`	`T={bfloat16,complex64,float}`
`TensorArrayCloseV3`
`TensorArrayConcatV3`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArrayGatherV3`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArrayGradV3`
`TensorArrayReadV3`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArrayScatterV3`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArraySizeV3`
`TensorArraySplitV3`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArrayV3`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorArrayWriteV3`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorListElementShape`	`shape_type={int32,int64}`
`TensorListPopBack`	`element_dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorListPushBack`	`element_dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TensorListReserve`	`shape_type={int32,int64}` `element_dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`Tile`	`Tmultiples={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TopKV2`	`T={bfloat16,float,int32,uint32}`
`Transpose`	`Tperm={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`TruncateDiv`	`T={bfloat16,complex64,float,int32,int64}`
`TruncateMod`	`T={bfloat16,float,int32,int64}`
`TruncatedNormal`	`T={int32,int64}` `dtype={float}`
`Unpack`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`UnsortedSegmentMax`	`Tnumsegments={int32,int64}` `Tindices={int32,int64}` `T={bfloat16,float,int32,int64,uint32,uint64}`
`UnsortedSegmentMin`	`Tnumsegments={int32,int64}` `Tindices={int32,int64}` `T={bfloat16,float,int32,int64,uint32,uint64}`
`UnsortedSegmentProd`	`Tnumsegments={int32,int64}` `Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`UnsortedSegmentSum`	`Tnumsegments={int32,int64}` `Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`VarIsInitializedOp`
`VariableShape`	`out_type={int32,int64}`
`While`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`Xdivy`	`T={complex64,float}`
`XlaBroadcastHelper`	`Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaConv`	`Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaDequantize`
`XlaDot`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaDynamicSlice`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaDynamicUpdateSlice`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaHostCompute`	`Toutputs={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `Tinputs={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaIf`	`Tout={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tin={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}` `Tcond={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`XlaKeyValueSort`	`V={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `K={bfloat16,float,int32,int64,uint32,uint64}`
`XlaPad`	`Tindices={int32,int64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaRecv`	`dtype={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaRecvFromHost`	`Toutput={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaReduce`	`T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaReduceWindow`	`Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaSelectAndScatter`	`Tindices={int32,int64}` `T={bfloat16,complex64,float,int32,int64,uint32,uint64}`
`XlaSend`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaSendToHost`	`Tinput={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaSort`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`XlaWhile`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`Xlogy`	`T={complex64,float}`
`ZerosLike`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`_Arg`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`
`_ArrayToList`	`out_types={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`_ListToArray`	`T={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}` `Tin={bfloat16,bool,complex64,float,int32,int64,uint32,uint64}`
`_Retval`	`T={bfloat16,bool,complex64,float,int32,int64,resource,uint32,uint64}`