tflite-flops

Roughly calculate FLOPs (floating-point operations) of a tflite format model.

This repository base on https://github.com/lisosia/tflite-flops, I add more ops support.

Support OP

CONV_2D
DEPTHWISE_CONV_2D
FULLY_CONNECTED
BATCH_MATMUL
ADD
SUB
MUL

Install

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

Usage

python3 -m tflite_flops example.tflite

Exapmle

wget https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_1.0_224.tgz
tar xvf mobilenet_v2_1.0_224.tgz
python3 -m tflite_flops ./mobilenet_v2_1.0_224.tflite

below lines printed

OP_NAME            | M FLOPS
------------------------------
CONV_2D            | 21.7
DEPTHWISE_CONV_2D  | 7.2
CONV_2D            | 12.8
.
.
.
CONV_2D            | 2.6
RESHAPE            | <IGNORED>
SOFTMAX            | <IGNORED>
------------------------------
Total: 601.6 M FLOPS

How is it calculated?

In the case of Conv 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

Limitation

Only Conv and DepthwiseConv layers are considered for now. It is enough for most of the time.