tflite-flops

Roughly calculate FLOPs (floating-point operations) of a TFLite format model. Supported layers:

• Conv2D
• DepthwiseConv2D
• MaxPool2D
• AveragePool2D
• Dense

Install

pip3 install git+https://github.com/gruellan/tflite-flops

Usage

python3 -m tflite_flops example.tflite

Example

wget https://storage.googleapis.com/tf_model_garden/vision/qat/mobilenetv2_ssd_coco/model_int8_qat.tflite
python3 -m tflite_flops ./model_int8_qat.tflite

below lines printed

OP_NAME                 | OUTPUT SHAPE         | FLOPS
---------------------------------------------
QUANTIZE                | [  1 256 256   3]    | <IGNORED>
CONV_2D                 | [  1 128 128  32]    | 28311552
DEPTHWISE_CONV_2D       | [  1 128 128  32]    | 9437184
CONV_2D                 | [  1 128 128  16]    | 16777216
.
.
.
CONCATENATION           | [    1 12276     4]  | <IGNORED>
DEQUANTIZE              | [    1 12276     4]  | <IGNORED>
CUSTOM                  | []                   | <IGNORED>
---------------------------------------------
Total: 1485.7 M FLOPS

How is it calculated?

In the case of Conv2D layer

Multiply-Accumulate (MAC) = output_h * output_w * output_c * kernel_h * kernel_w * input_c 
                         (= output_h * output_w * weight_size)
Floating-point operations (FLOPs) = 2 * MAC