Accuracy drop significantly after deploying model on ESP32 (AIV-603)

[Farzinkh]

Hi, I trained my simple model for grayscale 96*96 images and transferred it to the onnx model, then I tried to deploy the model on the ESP32-CAM board, after several attempts, I finally managed to run my code without errors, but the problem is that the accuracy has decreased from 99% to nearly 10%. In my code I load images from raw files saved by numpy library.

class HANDRECOGNATION : public Model<int16_t>
{
private:
	Reshape<int16_t> l1;
	Conv2D<int16_t> l2;
	Relu<int16_t> l3;
	MaxPool2D<int16_t> l4;
	Transpose<int16_t> l5;
	Reshape<int16_t> l6;
	Conv2D<int16_t> l7;
	Relu<int16_t> l8;
	Conv2D<int16_t> l9;

public:
	Softmax<int16_t> l10;

	HANDRECOGNATION () :
				l1(Reshape<int16_t>({96,96,1},"l1_reshape")),
				l2(Conv2D<int16_t>(-8, get_sequential_2_conv2d_2_biasadd_filter(), get_sequential_2_conv2d_2_biasadd_bias(), get_sequential_2_conv2d_2_biasadd_activation(), PADDING_VALID, {}, 1, 1, "l2")),
				l3(Relu<int16_t>("l3_relu")),
				l4(MaxPool2D<int16_t>({5,5},PADDING_VALID,{}, 5, 5, "l4")),
				l5(Transpose<int16_t>({2,1,0},"l5_transpose")),
				l6(Reshape<int16_t>({1,1,2592},"l6_reshape")),
				l7(Conv2D<int16_t>(-9, get_fused_gemm_0_filter(), get_fused_gemm_0_bias(), get_fused_gemm_0_activation(), PADDING_VALID, {}, 1, 1, "l7")),
				l8(Relu<int16_t>("l8_relu")),
				l9(Conv2D<int16_t>(-9, get_fused_gemm_1_filter(), get_fused_gemm_1_bias(), NULL, PADDING_VALID, {}, 1, 1, "l9")),
				l10(Softmax<int16_t>(-18,"l10")){}
	void build(Tensor<int16_t> &input)
	{
		this->l1.build(input,true);
		this->l2.build(this->l1.get_output(),true);
		this->l3.build(this->l2.get_output(),true);
		this->l4.build(this->l3.get_output(),true);
		this->l5.build(this->l4.get_output(),true);
		this->l6.build(this->l5.get_output(),true);
		this->l7.build(this->l6.get_output(),true);
		this->l8.build(this->l7.get_output(),true);
		this->l9.build(this->l8.get_output(),true);
		this->l10.build(this->l9.get_output(),true);
	}
	void call(Tensor<int16_t> &input)
	{
		this->l1.call(input);
		input.free_element();

		this->l2.call(this->l1.get_output());
		this->l1.get_output().free_element();

		this->l3.call(this->l2.get_output());
		this->l2.get_output().free_element();

		this->l4.call(this->l3.get_output());
		this->l3.get_output().free_element();

		this->l5.call(this->l4.get_output());
		this->l4.get_output().free_element();

		this->l6.call(this->l5.get_output());
		this->l5.get_output().free_element();

		this->l7.call(this->l6.get_output());
		this->l6.get_output().free_element();

		this->l8.call(this->l7.get_output());
		this->l7.get_output().free_element();

		this->l9.call(this->l8.get_output());
		this->l8.get_output().free_element();

		this->l10.call(this->l9.get_output());
		this->l9.get_output().free_element();
	}

};

And for loading images from SDcard:

fseek(f,0,SEEK_END);
long fsize=ftell(f);
fseek(f,0,SEEK_SET);
char *buffer=calloc(1,fsize+1);
fread(buffer,fsize,1,f);
fclose(f);
detect=run_inference((void *) buffer);

int run_inference(void *image){
Tensor<int16_t> input;
input.set_element((int16_t *)image).set_exponent(input_exponent).set_shape({96, 96, 1}).set_auto_free(false);
model.forward(input);
auto *score = model.l10.get_output().get_element_ptr();
auto max_score = score[0];
int max_index = 0;

for (size_t i = 1; i < 10; i++)
{
	if (score[i] > max_score)
	{
		max_score = score[i];
		max_index = i;
	}
}
return (max_index);
}

Also confusion matrix is:

Did I commit errors? I have already tested loading images from SD card with MNIST model without any problems.

[Farzinkh]

Hello everyone, I finally solved this problem here with a lot of experiments and I want to let you know about my repository AI_EDGE which has all the useful information and tools to implement a successful benchmark from SD Card.