Maix-EMC: Embedded/Edge Model Convertor, convert NN model for embedded systems
Maix-EMC act as TensorLayer（https://github.com/tensorlayer/tensorlayer）plugin, convert TensorLayer Model to kmodel, see TensorLayer Model Architecture for more information.
Kmodel is an flatten embedded Model format, currently used on Maix K210 boards, but it is also possible to run on normal MCU by using corresponding driver.
We design Maix-EMC as a easy extendable frame, you can easily add new layer type or new MCU or new model convertor for it, detail refer to following sections or this post: https://bbs.sipeed.com/t/topic/916
pip install dissect.cstruct
install TensorLayer, version above 2.0.1: https://github.com/tensorlayer/tensorlayer
Refer to example/cifar
prepare some pics for quant, put into cifar_dataset folder
prepare the model you need convert to kmodel: cifar10.h5
Do like cifar10_to_kmodel.py:
emc.save_kmodel(net, './cifar10.kmodel', './cifar10_dataset', dataset_func='img_0_1', quant_func='minmax', quant_bit=8, version=3)
defination of save_kmodel:
save_kmodel(network, filepath, dataset_dir, dataset_func='img_0_1', quant_func='minmax', quant_bit=8, version=3):
and it will generate cifar10.kmodel
I0628 11:22:10.987743 139933152151360 tl_logging.py:95] generate layer 8: EL_FULLY_CONNECTED
I0628 11:22:10.987811 139933152151360 tl_logging.py:95] buf_map=[40960, 1, 0]
I0628 11:22:10.995115 139933152151360 tl_logging.py:95] buf_size=a000, outsize=28, in_addr=0, out_addr=9fd8
I0628 11:22:10.995439 139933152151360 tl_logging.py:95] max_mem_usage=40960 Byte, 40 KB
I0628 11:22:10.995719 139933152151360 tl_logging.py:99] ===============================end gen_kmodel===================================
I0628 11:22:10.995798 139933152151360 tl_logging.py:99] kmodel size = 283924 Byte, 277 KB
I0628 11:22:10.996294 139933152151360 tl_logging.py:99] [*] Saved TL model into ./cifar10.kmodel, dataset path=./cifar10_dataset, quant func is minmax, quant bit is 8, kmodel version is 3
Now, we use 1.bmp in dataset to compare PC calculation and K210 calculate:
(you can use MaixPy to execute cifar10.kmodel, or use the example project in the example/cifar10)
PC:
1.0859082, -5.415024 , 1.554703 , 0.12335867, 0.46049562,
-0.6250827, -3.6137981, 0.5201204, -0.6762826, -3.7476015
K210:
1.181304, -5.381746, 1.580029, 0.045343, 0.511492,
-0.656629, -3.597826, 0.561016, -0.733270, -3.771462
Their result is close, but still need optimize quant.
add your Layer class in edge_layer.py
make new folder like 'stm32', and write stm32_constant.py, stm32_layer.py(optimize ops for stm32 layers), stm32_xxxmodel_bin.py.
It generate internal network layer list for model bin generator, layers in list have enough information to build anyother model,
and to_xxxmodel method will return bin of this layer, you can implement your self model method by adding to_xxxmodel method.
def gen_edge_model(network, platform, version, dataset_dir, dataset_func, quant_func, quant_bit)
generate edge model file from network, you can add platform in platform_table
def gen_edge_layers_from_network(network, platform, dataset, quant_func, quant_bit=8)
generate layer list from network, corresponding platform layer map is in tl_to_xxx_table
this table map TL layer class to EMC layer generator, and contain TL layer combine information.
def gen_edge_layer_from_network(network, platform, meta_info, idx)
generate single layer from TL layer
all layer type in edge_constant
all layer have init, to_xxxmodel method
all EMC layer types, convert TL layer class to EMC layer types.
load dataset from dir, generate ndarray for quant and infer usage.
you can add load method in loader_func_name_dict
return min,max for quant reference.
you can add your minmax functions in quant_func_name_dict.
now support minmax and kld method.
All common used layer (no kpu/npu/simd/dsp accelerate ops) is implemnet here.
every TL layer correspond to one or more EMC layer, gen_xxx_layer will do this map, it return EMC layer list of current TL layer.
EMC layer type defined in edge_constants.py, and implement in edge_layer.py and xxx_layer.py.
class xxx_Layer have at least two method:
def __init__(self, network, idx):
def to_kmodel(self, arg_oft, eight_bit_mode, buf_map): #generate kmodel bin and other information
all structure in k210's kpu.c
we use dissect.cstruct parse c struct to python objects.
contains kpu accelerate ops.
we transport orignal k210 kpu driver, so this file contains some warpper.
in normal mcu, you can redefine thoss layer to use ARM NN libs.
package every layer's bin to entirety kmodel.