pred2) # dl_dx5 = 14 x 14 x 3
dl_dw_conv, dl_db_conv = conv_backward(dl_dx5, x, w_conv, b_conv,
pred1)
dL_dw_conv += dl_dw_conv
dL_db_conv += dl_db_conv
dL_dw_fc += dl_dw_fc
dL_db_fc += dl_db_fc
k += 1
if k == num_of_batches:
k = 0
w_conv = w_conv - (learning_rate / batch_size * dL_dw_conv)
b_conv = b_conv - (learning_rate / batch_size * dL_db_conv)
w_fc = w_fc - (learning_rate / batch_size * dL_dw_fc.transpose())
b_fc = b_fc - (learning_rate / batch_size * dL_db_fc.transpose())
print(" CNN == Loading : {:.2f}%".format(iIter / nIters * 100),
end='\r')
print(" CNN == Loading : 100.00%", end='\n')
return w_conv, b_conv, w_fc, b_fc
if __name__ == '__main__':
c1, a1 = main.main_slp_linear()
c2, a2 = main.main_slp()
c3, a3 = main.main_mlp()
c4, a4 = main.main_cnn()
main.visualize_confusion_matrix_all(c1, a1, c2, a2, c3, a3, c4, a4)
mini_batch_dl_db_fc += dl_db_fc
print('Iteration # {}/{} loss = {} \r'.format(iter_i + 1,
NUM_ITERATIONS,
mini_batch_loss),
end='')
loss_values.append(mini_batch_loss)
# Update
# print(np.mean(np.abs(mini_batch_dl_dw_conv)), np.mean(np.abs(mini_batch_dl_db_conv)), np.mean(np.abs(mini_batch_dl_dw_fc)), np.mean(np.abs(mini_batch_dl_db_fc)))
w_conv = w_conv - mini_batch_dl_dw_conv * LEARNING_RATE
# b_conv = b_conv - mini_batch_dl_db_conv * LEARNING_RATE
w_fc = w_fc - mini_batch_dl_dw_fc.reshape(w_fc.shape) * LEARNING_RATE
b_fc = b_fc - mini_batch_dl_db_fc.reshape(b_fc.shape) * LEARNING_RATE
print()
axes = plt.gca()
plt.xlabel('iterations', fontsize=18)
plt.ylabel('training loss', fontsize=16)
plt.plot(loss_values)
plt.show()
return w_conv, b_conv, w_fc, b_fc
if __name__ == '__main__':
np.random.seed(42)
main.main_slp_linear()
main.main_slp()
main.main_mlp()
main.main_cnn()
w_fc -= lr * v_w_fc / np.sqrt(
s_w_fc +
1e-7) # combine momentum and RMSProp to perform update with Adam
v_b_fc = beta1 * v_b_fc + (1 - beta1) * db_fc / batch_size
s_b_fc = beta2 * s_b_fc + (1 - beta2) * (db_fc / batch_size)**2
b_fc -= lr * v_b_fc / np.sqrt(s_b_fc + 1e-7)
#print(L/batch_size)
#L_log.append(L/batch_size)
print("Train Over for SLP")
'''
plt.figure(1)
plt.plot(np.array(L_log))
plt.draw()
plt.pause(5)
plt.clf
'''
return w_conv, b_conv, w_fc, b_fc
if __name__ == '__main__':
retrain_tag = False
for arg in sys.argv[1:]:
if arg == "-r":
retrain_tag = True
main.main_slp_linear(retrain_tag)
main.main_slp(retrain_tag)
main.main_mlp(retrain_tag)
main.main_cnn(retrain_tag)