I write a pocket algorithm in python with numpy and found that if I change
w += (x[:, col_idx] * y[col_idx, 0]).transpose()(wrong answer)
to
w = w + (x[:, col_idx] * y[col_idx, 0]).transpose()(right answer)
the result that the algorithm will print different.
I have printed the w, x, y and the two version are the same.
And w is only being on the left side in above statement and have initialised.
I also have checked the algorithm and ensure that it is right.
What I cannot understand is that why only change w+=() to w=w+() will
change the whole result?(error_rate)
I have tested it that the w value will print the same.
The function part is
def pocket(x, y, update=50):
w = np.zeros(shape=(1, 5))
w_pocket = np.zeros(shape=(1, 5))
update_idx = 0
col_idx = 0
col_number = x.shape[1]
error_pocket = error_rate(x, y, w_pocket)
for update_idx in range(update):
col_range = random.sample(range(col_number), col_number)
for col_idx in col_range:
perceptron_score = w * x[:, col_idx]
if label(perceptron_score) != y[col_idx, 0]:
print "1:", col_idx, ":", x[:, col_idx].transpose(), y[col_idx, 0], w
w = w + (x[:, col_idx] * y[col_idx, 0]).transpose()
# w += (x[:, col_idx] * y[col_idx, 0]).transpose()
print "2:", col_idx, ":", x[:, col_idx].transpose(), y[col_idx, 0], w
error = error_rate(x, y, w)
if error < error_pocket:
w_pocket = w
error_pocket = error
break
return w_pocket
and if you want to test it
you need the file
https://github.com/lypan/Machine_Learning_Foundations
214976eed5abc4514832daa9e6adf6d0250d3371
just unzip and run the script
