Can it run on ESP-CAM with tensorflow lite?

[manjrekarom]

This may be a bit off-topic not sure though. So I am trying to do face recognition on ESP-CAM with 4MB flash.

At the moment the size of weights file for this model is 8MB so I am not able to put it on the device's flash which is only 4MB.
I am thinking about pruning/distilling the model further to reduce its size but I have not done it prior to this.

If possible can you share your ideas how can that be done and if it'll work? My last option is to make a custom network and training it myself. Welcoming other ideas.

Thanks!

[mvhsin]

Hi @manjrekarom ,

You can try with MobileFacenet-S, which model size will be around 1.4MB when quantized with TFlite with the following cmd:

tflite_convert --output_file tf-lite/MobileFacenet_uint8_128.tflite
--graph_def_file tf-lite/MobileFacenet.pb
--input_arrays "input"
--input_shapes "1,112,112,3"
--output_arrays output
--output_format TFLITE
--mean_values 128
--std_dev_values 127.5
--default_ranges_min 0
--default_ranges_max 255
--inference_type QUANTIZED_UINT8
--inference_input_type QUANTIZED_UINT8

But I met the other issue related to accuracy.
The result seems not correct with the quantized model, please kindly share the information if you have the idea. Many thanks!

[manjrekarom]

Hi @mvhsin ! Thanks for your reply.

Can I get a link to pretrained MobileFacenet-S?
The model that we had quantized was also about 1.5MB in size after quantization but after converting it back to .cc file it grew to about 6x - 8x (~8MB).

Sure. I will share if I find something relevant.

[mvhsin]

Hi @manjrekarom ,

The author of MobileFacenet mentioned in the paper.

remove the linear 1 × 1 convolution layer afterthe linear GDConv layer from MobileFaceNet, the resulting network of which is called MobileFaceNet-M. From MobileFaceNet-M, removing the 1 × 1 convolution layer before the linear GDConv layer produces the smallest network called MobileFaceNet-S.
Sorry I am not sure about *.cc file.
Is that done by TF-lite microcontroller?

[manjrekarom]

Oh thanks I'll take a look.

Yes it is for TF-Lite microcontrollers. The FlatBuffer obtained from an initial tf model, is converted to a C byte array.

This C file is taking up ~7x space (1.5 -> ~8MB). So we need an even smaller model. The options I feel I am left with are network pruning, knowledge distillation and training an even smaller custom architecture similar to MobileFaceNet.

[mvhsin]

Same here! Thanks for the info of TF-lite microcontroller.