def evaluate(A, C):
# accelerator results
acc = A.predict(X)
# classification by C and truly
cls_c = [1 if v[1] > v[0] else 0 for v in C.predict(X)]
cls_t = [accept(Y[i], acc[i]) for i in range(N)] # to bool
# relatvie error for all test data
re = [error.relative_error(Y[i], acc[i]) for i in range(N)]
re_c = [error.relative_error(Y[i], acc[i]) for i in range(N) if cls_c[i]]
# accuracy of C, recall of C
accuracy_of_C = sum([1.0 if cls_t[i] == cls_c[i] else 0 for i in range(N)]) / float(1e-10 + N)
recall_of_C = sum([1.0 if cls_t[i] and cls_c[i] else 0 for i in range(N)]) / float(1e-10 + sum([1 if v else 0 for v in cls_t]))
# invocation of C, invocation truly
invocation_of_C = float(sum([1 if v else 0 for v in cls_c])) / float(1e-10 + N)
invocation_truly = float(sum([1 if v else 0 for v in cls_t])) / float(1e-10 + N)
# re of A, re of A with C
mean_relative_error_of_A = sum(re) / float(1e-10 + len(re))
mean_relative_error_of_A_with_C = sum(re_c) / (1e-10 + len(re_c))
return {
'accuracy_of_C': accuracy_of_C,
'recall_of_C': recall_of_C,
'invocation_of_C': invocation_of_C,
'invocation_truly': invocation_truly,
'mean_relative_error_of_A': mean_relative_error_of_A,
'mean_relative_error_of_A_with_C': mean_relative_error_of_A_with_C
}
def main(benchmark, weights_file):
X0, Y0, X1, Y1 = load_data(benchmark)
A = AcceleratorModel([1, 4, 4, 2])
A.load_weights(weights_file)
Y2 = A.predict(X1)
print A.evaluate(X1, Y1)
re = [error.relative_error(Y1[i], Y2[i]) for i in xrange(len(X1))]
count = [0] * 101
for v in re:
count[int(math.floor(v * 100))] += 1
s = sum(count)
count = [float(v) / s for v in count]
for i in range(101):
print i, count[i]
for i in xrange(len(X1)):
print Y1[i], Y2[i], error.relative_error(Y1[i], Y2[i])
def main(benchmark, weights_file):
X0, Y0, X1, Y1 = load_data(benchmark)
A = AcceleratorModel([6, 8, 8,
1]) #Change this to the model by the network
A.load_weights(weights_file)
Y2 = A.predict(X1)
print(A.evaluate(X1, Y1))
re = [error.relative_error(Y1[i], Y2[i]) for i in range(len(X1))]
count = [0] * 101
for v in re:
count[int(math.floor(v * 100))] += 1
s = sum(count)
count = [float(v) / s for v in count]
for i in range(101):
print(i, count[i])
for i in range(len(X1)):
print(Y1[i], Y2[i], error.relative_error(Y1[i], Y2[i]))
def train_origin(A, X0, Y0, X1, Y1, epoch, batch_size, output_name):
A.fit(X0, Y0, nb_epoch=epoch, batch_size=batch_size)
acc = A.predict(X1)
result_num = [0 for i in range(100)]
result_sum = 0
for i in range(len(X1)):
rate = 0.01
tmp_re = error.relative_error(Y1[i], acc[i])
result_sum += tmp_re
for j in range(100):
if tmp_re <= rate:
result_num[j] += 1
rate += 0.01
result_mre = result_sum / float(len(X1))
result_re_list = [(100.0 * result_num[x] / float(len(X1))) for x in range(len(result_num))]
print result_mre
print result_re_list
f_results = open('../results/origin/{}.csv'.format(output_name), 'w')
f_results.write('mre,{}\n'.format(result_mre))
f_results.write('re_bound,re_percent\n')
for i in range(100):
f_results.write('{}%,{:.3f}%\n'.format(i + 1, result_re_list[i]))
f_results.flush()
f_results.close()
A.save_weights('../weights/origin/{}.weights'.format(output_name), overwrite=True)
keras_to_fann(A, 'fann_sample.nn', '../fann_model/{}.nn'.format(output_name))
'''
def accept(v0, v1):
return error.relative_error(
v0, v1) <= re_bound if eb_type == 1 else error.absolute_error(
v0, v1) <= re_bound
def accept(v0, v1):
return error.relative_error(v0, v1) <= re_bound
