Simple Neural Network (shapes images) Produces Same Outputs

Question

I work on a simple neural network for the detection of the forms, here is all the code on github: https://github.com/stodar/shapes/blob/main/saad.ipynb .

I get the same result even when I change the inputs, I redid the code several times but no result.

please do you have any idea?

corrects, wrongs = 0, 0

print('Target', '     Predicted', ' %')
for i in range(len(x_test)):
    input_vector = np.array(x_test[i], ndmin=2).T
    output_vector = np.dot(wh,  input_vector)
    output_vector = sigmoid(output_vector)
    output_vector = np.dot(wo, output_vector)
    res = sigmoid(output_vector)
    
    #Evaluate Model
    res_max = res.argmax()
    e = np.array(res_max)
    e = np.eye(3)[res_max]
    e.astype('int32')
    if np.array_equal(e,y_test[i]) :
        corrects += 1
    else:
        wrongs += 1
    
    print(y_test[i], ' ',e, '       ', np.max(res))
    
print("accuracy:", corrects / ( corrects + wrongs))
```

Answer 1

The line e.astype('int32') doesn't change the type of e, you need to save the result

Consider changing to:

e = e.astype('int32')

or:

if np.array_equal(e.astype('int32'),y_test[i])

Another thing: e = np.array(res_max) you override this line, you can just remove it.

Try run:

# Hidden Weight
wh = np.random.randn(no_of_hidden_nodes, no_of_in_nodes) * np.sqrt(28.0/no_of_in_nodes)
# Output Weight
wo = np.random.randn(no_of_out_nodes, no_of_hidden_nodes) * np.sqrt(28.0/no_of_hidden_nodes)

corrects, wrongs = 0, 0

print('Target', '     Predicted', ' %')
for i in range(len(x_test)):
    input_vector = np.array(x_test[i], ndmin=2).T
    output_vector = np.dot(wh,  input_vector)
    output_vector = sigmoid(output_vector)
    output_vector = np.dot(wo, output_vector)
    res = sigmoid(output_vector)
    
    #Evaluate Model
    res_max = res.argmax()
    e = np.array(res_max)
    e = np.eye(3)[res_max]
    e = e.astype('int32')
    if np.array_equal(e,y_test[i]) :
        corrects += 1
    else:
        wrongs += 1
    
    print(y_test[i], ' ',e, '       ', np.max(res))
    
print("accuracy:", corrects / ( corrects + wrongs))

And tell me if the result is different.