Accuracy failure

[ntsakoulis]

python command results from model.predict(test_set) with files csim_results.log and rtl_cosim_results.log. The log files are equivalent but vary significantly from the python results. I use mae as accuracy metric

x = layers.Flatten()(inputs)
x = QDense(64, kernel_quantizer = quantized_bits(6,2,1),bias_quantizer = quantized_bits(6,2,1))(x)
x = QActivation("quantized_relu(6,0)")(x)
x = layers.Dropout(0.3)(x)
x = QDense(64, kernel_quantizer = quantized_bits(6,2,1),bias_quantizer = quantized_bits(6,2,1))(x)
x = QActivation("quantized_tanh(6,2)")(x)
x = QDense(64, kernel_quantizer = quantized_bits(6,2,1),bias_quantizer = quantized_bits(6,2,1))(x)
x = QActivation("quantized_tanh(6,2)")(x)
x = layers.Dropout(0.3)(x)
x = QDense(64, kernel_quantizer = quantized_bits(6,2,1),bias_quantizer = quantized_bits(6,2,1))(x)
x = QActivation("quantized_tanh(6,2)")(x)
x = layers.Dropout(0.3)(x)
x = QDense(1, kernel_quantizer = quantized_bits(6,2,1),bias_quantizer = quantized_bits(6,2,1))(x)
outputs = QActivation("quantized_tanh(6,2)")(x)
model = keras.Model(inputs, outputs)

KerasJson: model_Q_ANN_8.json
KerasH5: model_Q_ANN_8_weights.h5
OutputDir: my-hls-q_ann8
ProjectName: project_Q_ANN8
InputData: data_standarized.dat
OutputPredictions: predicted_scaled_Q_A8.dat
XilinxPart: xczu7ev-ffvc1156-2-e
ClockPeriod: 5ns

IOType: io_stream # options: io_stream/io_parallel
HLSConfig:
Model:
Precision: ap_fixed<14,6>
ReuseFactor: 4
Strategy: Resource # options: Latency/Resource

first one is my model and second the configuration file yml,

thanks,

[jmduarte]

@ntsakoulis
Can you share the dataset you are using?

What is the accuracy of the QKeras model compared to the hls4ml model?

[ntsakoulis]

https://www.kaggle.com/datasets/loveall/appliances-energy-prediction

This is the dataset I am using
First 55% training
Next 25% validation
And the remaining as test set

Thanks,