esp_nn_conv_s8_esp32s3 returns unespected values #23
Description
Checklist
- Checked the issue tracker for similar issues to ensure this is not a duplicate
- Read the documentation to confirm the issue is not addressed there and your configuration is set correctly
- Tested with the latest version to ensure the issue hasn't been fixed
How often does this bug occurs?
always
Expected behavior
The esp-nn optimized convolution functions are producing incorrect outputs, leading to a significant drop in model accuracy from 92% to below 70%. When using the standard ANSI C implementation, the model performs as expected.
When running the CNN model with esp-nn optimizations enabled, the model should maintain an accuracy similar to the standard C implementation (around 92%). The convolution layers' outputs should be a range of values, not predominantly clamped to -128.
Actual behavior (suspected bug)
Enable esp-nn optimizations. Run the model again. Observe that the accuracy drops significantly, typically below 70%.
Use debugging tools (like printf statements) within the esp_nn_conv_s8_filter_aligned_input_padded_esp32s3 function to inspect the output data. Confirm that the out_data array contains many instances of the value -128.
The attached debug logs show the out_data of multiple convolution layers, consistently containing -128 values, indicating a saturation or clamping issue.
When esp-nn optimization is toggled on, the model's accuracy drops below 70%. Debugging the optimized esp_nn_conv_s8_filter_aligned_input_padded_esp32s3 function shows that the output (out_data) of the convolution operation is filled with an excessive number of -128 values.
This clamping of values suggests a potential issue in the quantization parameters or the implementation of the optimized convolution kernel, causing the output to consistently hit the lowest possible value for an 8-bit signed integer.
Error logs or terminal output
Log Snippets:
--- DEBUG: Funcao otimizada executada! Verificando a saida ---
Primeiros 32 bytes de out_data:
-128 -128 -128 -128 -128 -116 -128 -119 -128 -128 -113 -127 -116 -125 -128 -128 -118 -120 -128 -123 -124 -128 -128 -128 -128 -121 -123 -127 -121 -128 -128 -128
This is the output for the first convolution layer, where many values are already -128.
--- DEBUG: Funcao otimizada executada! Verificando a saida ---
Primeiros 32 bytes de out_data:
-128 -128 -128 -128 -128 -128 -97 -128 -128 -124 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -114 -128 -128 -128 -128 -120 -128 -94 -128 -128 -128
This is the output for the second convolution layer, which also shows widespread -128 values.
--- DEBUG: Funcao otimizada executada! Verificando a saida ---
Primeiros 32 bytes de out_data:
-128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -128 -126 -128 -128 -128 -128 -128 -128 -128 -123 -128 -128 -128 -128 -128 -128 -128
This is the output for the third convolution layer, where almost all values are -128.
The input offset and output offset values from the esp_nn function calls are shown to be -17 and -128, respectively. These values might be contributing to the problem.Steps to reproduce the behavior
Configure a CNN model with the following architecture:
conv2d_46 (Conv2D)
max_pooling2d_46 (MaxPooling2D)
conv2d_47 (Conv2D)
max_pooling2d_47 (MaxPooling2D)
conv2d_48 (Conv2D)
max_pooling2d_48 (MaxPooling2D)
dropout_30 (Dropout)
flatten_15 (Flatten)
dropout_31 (Dropout)
dense_17 (Dense)
Run the model with esp-nn optimizations disabled. Observe that the model achieves an accuracy of approximately 92%.
Enable esp-nn optimizations. Run the model again. Observe that the accuracy drops significantly, typically below 70%.
Use debugging tools (like printf statements) within the esp_nn_conv_s8_filter_aligned_input_padded_esp32s3 function to inspect the output data. Confirm that the out_data array contains many instances of the value -128.
Project release version
Lastest
System architecture
other (details in Additional context)
Operating system
Windows
Operating system version
Windows 11
Shell
PowerShell
Additional context
No response