TensorFlow Lite & TensorFlow Compatibility Guide

TensorFlow Lite supports a number of TensorFlow operations used in common inference models. As they are processed by the TensorFlow Lite Optimizing Converter, those operations may be elided or fused, before the supported operations are mapped to their TensorFlow Lite counterparts.

Since the set of TensorFlow Lite operations is smaller than TensorFlow's, not every model is convertible. Even for supported operations, very specific usage patterns are sometimes expected, for performance reasons. We expect to expand the set of supported operations in future TensorFlow Lite releases.

The best way to understand how to build a TensorFlow model that can be used with TensorFlow Lite is to carefully consider how operations are converted and optimized, along with the limitations imposed by this process.

Supported Types

Most TensorFlow Lite operations target both floating-point (float32) and quantized (uint8) inference, but usually there is little or no support for other types like tf.float16 and strings.

Apart from using different version of the operations, the other difference between floating-point and quantized models lies in the way they are converted. Quantized conversion expect the models to be annotated with “fake quantization” nodes that record the dynamic range of the tensors. Without that information TF Lite is not able to accurately quantize a model, which means that proper quantized training is necessary before conversion.

Data Format and Broadcasting

At the moment TensorFlow Lite supports only TensorFlow's “NHWC” format, and broadcasting is only support in a limited number of ops (tf.add, tf.mul, tf.sub, and tf.div).

Compatible Operations

The following TensorFlow operations are usually mapped to their TensorFlow Lite counterparts:

• tf.batch_to_space_nd - as long as the input tensor is 4D (1 batch + 2 spatial + 1 other) and the crops attribute is not used
• tf.exp
• tf.fake_quant*
• tf.matmul - as long as the second argument is constant and transposition is not used
• tf.nn.avg_pool
• tf.nn.conv2d - as long as the filter is constant
• tf.nn.depthwise_conv2d - as long as the filter is constant and rate is [1,1]
• tf.nn.l2_normalize - as long as normalization is done along the last dimension
• tf.nn.local_response_normalization
• tf.nn.log_softmax - as long as axis is not provided
• tf.nn.max_pool
• tf.nn.softmax - as long as tensors are 2D and axis is the last dimension
• tf.nn.top_k
• tf.one_hot
• tf.pad - as long as mode and constant_values are not used
• tf.reduce_mean - as long as the reduction_indices attribute is not used
• tf.reshape
• tf.sigmoid
• tf.space_to_batch_nd - as long as the input tensor is 4D (1 batch + 2 spatial + 1 other)
• tf.space_to_depth
• tf.split - as long as num is not provided and num_or_size_split contains number of splits as a 0D tensor
• tf.squeeze - as long as axis is not provided
• tf.strided_slice - as long as ellipsis_mask and new_axis_mask are not used
• tf.transpose - as long as conjugate is not used

Straightforward Conversions, Constant-Folding and Fusing

A number of TensorFlow operations can be processed by TensorFlow Lite even though they have no direct equivalent. This is the case for operations that can be simply removed from the graph (tf.identity), replaced by tensors (tf.placeholder), or fused into more complex operations (tf.nn.bias_add). Even some supported operations may sometimes be removed through one of these processes.

Here is a list of TensorFlow operations that are usually removed from the graph:

• tf.add
• tf.check_numerics
• tf.constant
• tf.div
• tf.divide
• tf.fake_quant_with_min_max_args
• tf.fake_quant_with_min_max_vars
• tf.identity
• tf.maximum
• tf.minimum
• tf.multiply
• tf.no_op
• tf.placeholder
• tf.placeholder_with_default
• tf.realdiv
• tf.reduce_max
• tf.reduce_min
• tf.reduce_sum
• tf.rsqrt
• tf.shape
• tf.sqrt
• tf.square
• tf.subtract
• tf.tile
• tf.nn.batch_norm_with_global_normalization
• tf.nn.bias_add
• tf.nn.fused_batch_norm
• tf.nn.relu
• tf.nn.relu6

Note that many of those operations don‘t have TensorFlow Lite equivalents and the corresponding model will not be convertible if they can’t be elided or fused.

Unsupported Operations

TensorFlow operation not listed above are likely unsupported. Notably, the following common ops are not supported at the moment:

• tf.depth_to_space
• tf.image.resize_bilinear
• tf.tanh

TensorFlow Lite Operations

The following TensorFlow Lite operations are fully supported and used in place of the TensorFlow operations listed above:

ADD

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: elementwise sum of the input tensors
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
}

AVERAGE_POOL_2D

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor where each entry is the mean of the input values in the
     corresponding window.
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  padding: SAME|VALID
  stride_w,stride_h: stride of the sliding window
  filter_width,filter_height: size of the sliding window
}

BATCH_TO_SPACE_ND

Inputs {
  0: 4D tensor
  1: 1D tensor
  2: 2D tensor
}
Outputs {
  0: tensor rearranged using block_shape. See tf.batch_to_space_nd for
     details.
}

CONCATENATION

Inputs {
  0-N: any number of tensors
}
Outputs {
  0: concatenation of the input tensors along the given axis.
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  axis: dimension along which the concatenation is performed
}

CONV_2D

Inputs {
  0: 4D tensor
  1: filter
  2: bias (optional)
}
Outputs {
  0: result of 2D convolution of the input tensor
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  padding: SAME|VALID
  stride_w,stride_h: stride of the filter window
}

CONV_2D_TRANSPOSE

Inputs {
  0: output_shape
  1: filter
  2: 4D tensor
}
Outputs {
  0: the transpose (gradient) of conv2d
}
Options {
  padding: SAME|VALID
  stride_w,stride_h: stride of the filter window
}

DEPTHWISE_CONV_2D

Inputs {
  0: 4D tensor
  1: filter
  2: bias (optional)
}
Outputs {
  0: result of a depthwise-2D convolution of the input tensor
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  padding: SAME|VALID
  stride_w,stride_h: stride of the filter window
  depth_multiplier: relation between the last dimension of the input and output
    tensors
}

EQUAL

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is
  equal to the corresponding element of the second tensor.
}

EXP

Inputs {
  0: tensor
}
Outputs {
  0: result of computing element-wise exponential of the input tensor
}

FLOOR

inputs {
  0: tensor
}
outputs: {
  0: result of computing element-wise floor of the input tensor
}

FULLY_CONNECTED

Inputs {
  0: 4D tensor
  1: filter
  2: bias (optional)
}
Outputs {
  0: output of a fully (densely) connected layer, which connects all
     elements in the input tensor with each element in this tensor.
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
}

GATHER

Inputs {
  0: params tensor
  1: indices tensor
  2: axis tensor (optional)
}
Outputs {
  0: a tensor with same type as the params tensor.
}

GREATER

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is
  greater than the corresponding element of the second tensor.
}

GREATER_EQUAL

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is
  greater than or equal to the corresponding element of the second tensor.
}

L2_NORMALIZATION

Inputs {
  0: input tensor
}
Outputs {
  0: normalized tensor (along the last dimension)
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
}

L2_POOL_2D

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to tf.sqrt(tf.nn.ave_pool(tf.square(input))
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  padding: SAME|VALID
  stride_w,stride_h: stride of the sliding window
  filter_width,filter_height: size of the sliding window
}

LESS

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is less
  than the corresponding element of the second tensor.
}

LESS_EQUAL

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is less
  than or equal to the corresponding element of the second tensor.
}

LOCAL_RESPONSE_NORMALIZATION

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to tf.nn.local_response_normalization
}
Options {
  radius
  bias
  alpha
  beta
}

LOGISTIC

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to 1 / (1 + exp(-input))
}

LOG

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to log(input)
}

LOG_SOFTMAX

Inputs {
  0: tensor
}
Outputs {
  0: tensor equivalent to logits - log(reduce_sum(exp(logits), -1))
}

MAX_POOL_2D

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor where each entry is the maximum of the input values in the
     corresponding window.
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
  padding: SAME|VALID
  stride_w,stride_h: stride of the sliding window
  filter_width,filter_height: size of the sliding window
}

MUL

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: elementwise multiplication of the input tensors
}
Options {
  fused_activation_function:  NONE|RELU|RELU6
}

NEG

Inputs {
  0: a tensor
}
Outputs {
  0: elementwise negation of the input tensor
}

PAD

Inputs {
  0: tensor
  1: tensor
}
Outputs {
  0: tensor where additional values are added before and after the contents of
     each dimension
}

MEAN (tf.reduce_mean)

Inputs {
  0: tensor
  1: tensor
}
Outputs {
  0: tensor containing the mean of the elements
}
Options {
  keep_dims: whether to retain reduced dimensions
}

NOT_EQUAL

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: a tensor of type bool, true whenever an element of the first tensor is not
  equal to the corresponding element of the second tensor.
}

RELU

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to max(0, input)
}

RELU_N1_TO_1

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to max(-1, min(input, 1)
}

RELU6

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to max(0, min(input, 6)
}

RESHAPE

Inputs {
  0: a tensor
  1: ignored
}
Outputs {
  0: a tensor with the same elements as the input but with the new shape
}
Options {
  new_shape
}

RSQRT

Inputs {
  0: a tensor
}
Outputs {
  0: result of computing element-wise reciprocal square root of the input tensor
}

SHAPE

Inputs {
  0: a tensor
}
Outputs {
  0: a 1D tensor representing the shape of the input tensor
}
Options {
  out_type: the output type of the op (int32 or int64). Defaults to int32.
}

SLICE

Inputs {
  0: tensor
  1: 1D tensor
  2: 1D tensor
}
Outputs {
  0: slice of the input tensor of the given size from the given begin index.
}

SOFTMAX

Inputs {
  0: a tensor
}
Outputs {
  0: a tensor equivalent to exp(input) / tf.reduce_sum(exp(input * beta), dim),
     where dim is always the last dimension of the input tensor.
}
Options {
  beta
}

SPACE_TO_DEPTH

Inputs {
  0: a 4D tensor
}
Outputs {
  0: a tensor rearranged using block_size. See tf.space_to_depth for details.
}
Options {
  block_size
}

SPACE_TO_BATCH_ND

Inputs {
  0: 4D tensor
  1: 1D tensor
  2: 2D tensor
}
Outputs {
  0: a tensor rearranged using block_shape. See tf.space_to_batch_nd for
     details.
}

SPARSE_TO_DENSE

Inputs {
  0: 0D or 1D or 2D tensor
  1: 1D tensor
  2: 0D or 1D tensor
  3: 0D tensor
  4: a boolean value
}
Outputs {
  0: Dense Tensor of shape output_shape. Has the same type as sparse_values.
}

SPLIT

Inputs {
  0: 0D tensor (axis)
  1: tensor (input)
}
Outputs {
  0-N: subtensors built from the input tensors
}
Options {
  num_splits: Specifies number of outputs
}

SQRT

Inputs {
  0: a tensor
}
Outputs {
  0: result of computing element-wise square root of the input tensor
}

SQUEEZE

Inputs {
  0: tensor
}
Outputs {
  0: tensor without any dimensions of size 1
}
Options {
  squeeze_dims
}

STRIDED_SLICE

Inputs {
  0: tensor
  1: 1D tensor
  2: 1D tensor
  3: 1D tensor
}
Outputs {
  0: slice of the input tensor of the given size
}
Options {
  begin_mask: mask for begin indices
  end_mask: mask for end indices
  shrink_axis_mask: mask that indicates which dimensions to remove
}

TOP_K

Inputs {
  0: tensor
  1: OD tensor
}
Outputs {
  0: k largest element along each last dimensional slice
  1: indices of values within the last dimension of the input ensor
}

TRANSPOSE

Inputs {
  0: tensor
  1: tensor
}
Outputs {
  0: tensor permuted according to perm
}

SELECT

Inputs {
  0: tensor
  1: tensor
  2: tensor
}
Outputs {
  0: tensor that contains the elementwise values of 'tensor 1' if the
  corresponding value of 'tensor 0' is true or the value of 'tensor 2' if false.
}

POW

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: elementwise pow of the input tensors
}

ARG_MAX

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: A tensor of indices of maximum values.
}

ARG_MIN

Inputs {
  0: a tensor
  1: a tensor
}
Outputs {
  0: A tensor of indices of minium values.
}

PACK

Inputs {
  0: a list of tensors.
  1: an integer.
}
Outputs {
  0: A tensor of stacked tensors.
}

LOGICAL_OR

Inputs {
  0: a list of tensors.
  1: a list of tensors.
}
Outputs {
  0: A tensor of logical_or output tensors.
}

UNPACK

Inputs {
  0: a tensor.
  1: an integer.
  2: an integer.
}
Outputs {
  0-N: tensors of unpacked tensor.
}

FLOOR_DIV

Inputs {
  0: a list of tensors.
  1: a list of tensors.
}
Outputs {
  0: A tensor of floor_div output tensors.
}

And these are TensorFlow Lite operations that are present but not ready for custom models yet:

• CALL
• CONCAT_EMBEDDINGS
• CUSTOM
• EMBEDDING_LOOKUP
• EMBEDDING_LOOKUP_SPARSE
• HASHTABLE_LOOKUP
• LSH_PROJECTION
• LSTM
• RESIZE_BILINEAR
• RNN
• SKIP_GRAM
• SVDF
• TANH