snake model on STM32

[sv1bds]

Hi,
I want to use this model
model = tf.keras.Sequential([
tf.keras.layers.Input(shape=(6,)),
tf.keras.layers.Dense(32, activation='relu'),
tf.keras.layers.Dense(32, activation='relu'),
tf.keras.layers.Dense(4, activation='softmax')
])
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),loss='mse')
to play snake game on an STM32. You mention that only Dense is implemented. Does Input can be implemented as a separate layer or must be included in Dense?

[sv1bds]

Model: "sequential"

---

Layer (type) Output Shape Param #

=================================================================

dense (Dense) (None, 32) 224

dense_1 (Dense) (None, 32) 1056

dense_2 (Dense) (None, 4) 132

=================================================================

Total params: 1412 (5.52 KB)

Trainable params: 1412 (5.52 KB)

Non-trainable params: 0 (0.00 Byte)

---

Traceback (most recent call last):

File "snake4m.py", line 59, in

weights_to_cpp(model)

File "snake4m.py", line 17, in weights_to_cpp

for i in np.array(weights):

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (3,) + inhomogeneous part.

Hi again,
When I try to parse my model it stopped with this message.
I replace the for i in np.array(weights): with for i in weights: and also for biases and seems working ok.
Thanks

[sv1bds]

int[] activations = {SIGMOID, SIGMOID, SIGMOID} //Activations for each layer excluding the input layer
it must be so?
int activations[] = {SIGMOID, SIGMOID, SIGMOID}; //Activations for each layer excluding the input layer

[sv1bds]

void activate(int l, double* z, int activation) {
double sum = 0.0;
if (activation == SOFTMAX) {
for (int j = 0; j < l; j++) {
sum += exp(z[j]);
}
}
for (int i = 0; i < l; i++) {
if (activation == SIGMOID) {
z[i] = 1 / (1 + exp(-z[i]));
} else if (activation == TANH) {
z[i] = tanh(z[i]);
} else if (activation == EXPONENTIAL){
z[i] = exp(z[i]);
} else if (activation == SWISH){
z[i] = z[i] / (1 + exp(-z[i]));
} else if (activation == RELU){
z[i] = fmax(0, z[i]);
} else if (activation == SOFTMAX) {
z[i] = exp(z[i])/sum;
}
}
}

I implement SOFTMAX. It is not verified yet.

[sv1bds]

The snake model works fine on STM32!!! The SOFTMAX is verified. It uses 56% of memory. If you need more info please tell me.
Thanks

[sv1bds]

A complete description of the project can be found here : link