TFLite Support¶

Bitweaver accepts TensorFlow Lite (.tflite) flatbuffers for compilation and benchmarking. The file format is detected automatically on upload, so TFLite models go through the same flow as any other Bitweaver model.

This page describes the requirements a .tflite model must meet, the operators currently supported, and the parameter restrictions on each.

Beta

Compiling .tflite models is under active development. The supported operator set and parameter coverage are still expanding, and behavior may change between releases.

Need an op that isn't listed?

Get in touch; new operators are added based on customer demand.

Model requirements¶

The file must be a valid .tflite flatbuffer produced by the TensorFlow Lite converter.
The model must be fully integer-quantized with int8 weights and activations. A subset of element-wise, shape, and pooling ops also accept uint8 activations (noted per operator below). Hybrid (float-activations / int8-weights) models are not supported.
Float32 input and output tensors are allowed when they are bracketed by QUANTIZE / DEQUANTIZE ops, which is the standard TensorFlow Lite full-integer export pattern. Models with int8 input and output tensors are also accepted.
The model must contain a single subgraph. Exports that produce multiple subgraphs (for example, models with multiple signatures, or control-flow ops such as IF / WHILE) are rejected.
The model must have a single input tensor and a single output tensor.
All tensor shapes must be static. Dynamic-shape tensors are not supported.
The batch dimension must be 1 for 4D (NHWC) activations.
Custom operators are not supported.

Supported operators¶

All operators run on int8 activations unless noted otherwise. Fused activations, where applicable, may be NONE, RELU, or RELU6.

Convolution and dense¶

Operator	Restrictions
`CONV_2D`	Per-channel or per-tensor weight quantization. Stride is configurable. Dilation must be `1`. SAME and VALID padding. Fused activation `NONE`, `RELU`, or `RELU6`.
`DEPTHWISE_CONV_2D`	`depth_multiplier` must be `1` (or greater than `1` only when input channels `= 1`). Dilation must be `1`. SAME and VALID padding. Fused activation `NONE`, `RELU`, or `RELU6`.
`FULLY_CONNECTED`	Flat 2D inputs only; a non-trivial leading (batched-matmul) dimension is not supported. Fused activation `NONE`, `RELU`, or `RELU6`.

Pooling and reduction¶

Operator	Restrictions
`AVERAGE_POOL_2D`	Input and output quantization (scale, zero-point) must match. No fused activation. SAME and VALID padding.
`MAX_POOL_2D`	`int8` or `uint8`. Input and output quantization must match. No fused activation. SAME and VALID padding.
`MEAN`	Reduction over the NHWC spatial axes `[1, 2]` only (global average pooling). Input must be 4D.

Element-wise¶

Operator	Restrictions
`ADD`	Two `int8` tensors with NHWC broadcast over size-1 dims; one operand may be constant. Fused activation `NONE`, `RELU`, or `RELU6`.
`MUL`	Two `int8` tensors with NHWC broadcast; one operand may be constant. Activation inputs must be 4D NHWC. No fused activation.
`MINIMUM`	Two `int8` operands that share quantization, with NHWC broadcast.

Shape and data movement¶

Operator	Restrictions
`RESHAPE`	Any shape change permitted by the underlying tensor layout.
`CONCATENATION`	`int8` or `uint8`, along any non-batch axis. Inputs and output must share a dtype. Constant inputs are allowed. No fused activation.
`PAD`	`int8` or `uint8`, rank-4 NHWC. Spatial and channel padding only (no batch-axis padding). Input and output quantization must match.
`MIRROR_PAD`	Rank-4 NHWC, reflect or symmetric mode. No batch-axis padding. Input and output quantization must match.
`SLICE`	Rank-4 NHWC, batch axis kept intact. Input and output quantization must match.
`TRANSPOSE`	`int8` or `uint8`, rank-4 NHWC. The batch axis must stay fixed. Input and output quantization must match.
`RESIZE_BILINEAR`	`int8` or `uint8`, NHWC. Requires `half_pixel_centers = True` and `align_corners = False`. Input and output quantization must match.
`RESIZE_NEAREST_NEIGHBOR`	`int8` or `uint8`, NHWC. Requires `align_corners = False`. Input and output quantization must match.

Activations¶

Operator	Restrictions
`RELU`	`int8` or `uint8`. Input and output quantization must match.
`LEAKY_RELU`	`int8` or `uint8`. `alpha` must be non-negative.
`LOGISTIC`	`int8`, computed along the last axis. Output zero-point must be `-128`.
`SOFTMAX`	Computed along the last axis.

Quantization¶

Operator	Restrictions
`QUANTIZE`	`float32` to `int8` / `uint8` (boundary quantize), or `int8` / `uint8` to `int8` / `uint8` (rescale).
`DEQUANTIZE`	`int8` to `float32` only (network-output boundary).

Shape-computation ops are folded automatically

A SHAPE / STRIDED_SLICE / PACK chain that only recomputes a static reshape target (a common converter artifact) is detected and folded away, so it does not need to appear in the supported list. These ops are not supported as general runtime operators.

Fused activations¶

Fused activations on convolution and fully-connected layers map to operator-level RELU or RELU6. Activations other than RELU and RELU6 (for example TANH or LOGISTIC) are not accepted as fused activations; include them as separate ops only if they appear in the supported list above.

Errors¶

If a model fails the checks above, the benchmark transitions to build_failed and the error is shown in the run details drawer. The most common messages are:

Message	Cause	Suggested fix
Corrupt TFLite file	Flatbuffer header is invalid or the file is truncated.	Re-export the model.
Your TFLite model uses unsupported operators	The model uses an operator or parameter combination not listed above.	Adjust the model to use only supported ops, or rebuild with the DeepGate SDK.
TFLite model contains multiple subgraphs	The exported `.tflite` has more than one subgraph (often from multiple signatures or control-flow ops).	Re-export with a single signature so the converter emits one subgraph.
TFLite model has multiple inputs or outputs	The model has more than one input or output tensor (for example a multi-head detector).	Re-export a single-input, single-output model.