def viz_fit_predict():
actual_fit_str = file_helper.read2mem('../../' + LINE_FILE_PATH)
actual_fit_s = actual_fit_str.split('===\n')
actual_loss = list()
predict_fit_str = file_helper.read2mem('../../' + PREDICT_FILE_PATH)
predict_fit_s = predict_fit_str.split('===\n')
predict_loss = list()
cur_idx = 0
print('actual_fit_s length: %d' % len(actual_fit_s))
print('predict_fit_s length: %d' % len(predict_fit_s))
for loss_data, predict_data in zip(actual_fit_s, predict_fit_s):
loss_es = loss_data.split('\n')
predict_s = predict_data.split('\n')
print('loss_es length: %d' % len(loss_es))
print('predict_s length: %d' % len(predict_s))
for loss, predict in zip(loss_es, predict_s):
if re.match("^\d+?\.\d+?$", loss) is not None:
actual_loss.append(float(loss))
if re.match("^\d+?\.\d+?$", predict) is not None:
predict_loss.append(float(predict))
if len(actual_loss) < 1 or len(predict_loss) < 1:
print('actual_loss length: %d' % len(actual_loss))
print('predict_loss length: %d' % len(predict_loss))
continue
actual_pl, = plt.plot(actual_loss, label='actual')
predict_pl, = plt.plot(predict_loss, label='predict')
plt.legend(handles=[actual_pl, predict_pl])
plt.savefig('data/fit_result%d.png' % cur_idx)
plt.clf()
cur_idx += 1
del actual_loss[:]
del predict_loss[:]
def grad_line():
batch_size_s = list()
depth_s = list()
num_hidden_s = list()
layer_cnt_s = list()
patch_size_s = list()
exp_file = file_helper.read2mem('../../' + GRAD_FILE_PATH)
# print(exp_file)
exp_datas = exp_file.split('\n')
for exp_data in exp_datas:
hps = exp_data[1:-1].split(', ')
if re.match("^(-)\d+?\.\d+?$", hps[0]) is None:
continue
batch_size_s.append(float(hps[0]))
depth_s.append(float(hps[1]))
num_hidden_s.append(float(hps[2]))
layer_cnt_s.append(float(hps[3]))
patch_size_s.append(float(hps[4]))
batch_pl, = plt.plot(batch_size_s, label='batch_size')
depth_pl, = plt.plot(depth_s, label='depth')
num_pl, = plt.plot(num_hidden_s, label='num_hidden')
layer_pl, = plt.plot(layer_cnt_s, label='layer_cnt')
patch_pl, = plt.plot(patch_size_s, label='patch_size')
plt.legend(handles=[batch_pl, depth_pl, num_pl, layer_pl, patch_pl])
plt.savefig('data/grad_change.png')
plt.show()
def viz_hp2trend(dir_pos):
fit_loss_es = file_helper.read2mem(
dir_pos + 'hp2trend_fit_loss.txt').split('\n')[:-1]
stable_loss_es_predict = file_helper.read2mem(
dir_pos + 'hp2trend_stable_loss_predict.txt').split('\n')[:-1]
stable_loss_es_label = file_helper.read2mem(
dir_pos + 'hp2trend_stable_loss_label.txt').split('\n')[:-1]
hps = list()
for i in range(5):
hps.append(
file_helper.read2mem(dir_pos +
'hp2trend_hps%d.txt' % i).split('\n')[:-1])
grads = list()
for i in range(5):
grads.append(
file_helper.read2mem(dir_pos +
'hp2trend_grads%d.txt' % i).split('\n')[:-1])
plt.plot(fit_loss_es, label='fit_loss')
plt.savefig(dir_pos + 'hp2trend_fit_loss.png')
plt.clf()
del fit_loss_es[:]
pl_stable_loss_predict, = plt.plot(stable_loss_es_predict,
label='stable_loss_predict')
pl_stable_loss_label, = plt.plot(stable_loss_es_label,
label='stable_loss_label')
plt.legend(handles=[pl_stable_loss_predict, pl_stable_loss_label])
plt.savefig(dir_pos + 'stable_loss.png')
plt.clf()
del stable_loss_es_predict[:]
del stable_loss_es_label[:]
pl_list = list()
for i in range(5):
pl_hp, = plt.plot(hps[i], label='hp%d' % i)
pl_list.append(pl_hp)
plt.legend(handles=pl_list)
plt.savefig(dir_pos + 'hps.png')
plt.clf()
del pl_list[:]
for i in range(5):
pl_grad, = plt.plot(grads[i], label='grad%d' % i)
pl_list.append(pl_grad)
plt.legend(handles=pl_list)
plt.savefig(dir_pos + 'grads.png')
plt.clf()
del pl_list[:]
def mark():
exp_file = file_helper.read2mem('../../' + LINE_FILE_PATH)
exp_datas = exp_file.split('\n')
cur_line = list()
line_cnt = 0
for exp_data in exp_datas:
if exp_data.startswith('=='):
plt.plot(cur_line)
line_cnt += 1
del cur_line[:]
elif exp_data.startswith('['):
# 超参数转图片名
jpg_name = str([
num_str[0:4].replace("-", "")
for num_str in exp_data[1:-1].split()
])[1:-1].replace("'", "").replace(', ', '-')
plt.savefig('data/' + str(line_cnt / 2) + '-' + jpg_name + '.png')
plt.clf()
continue
else:
cur_line.append(exp_data)
def avg_line():
exp_file = file_helper.read2mem('../../' + LINE_FILE_PATH)
exp_datas = exp_file.split('\n')
cur_sum = 0
cur_cnt = 0
avg_s = list()
is_loss = True
for exp_data in exp_datas:
if exp_data.startswith('=='):
if is_loss:
avg_s.append(cur_sum / cur_cnt)
cur_sum = 0
cur_cnt = 0
is_loss = not is_loss
# 两次一翻转
elif exp_data.startswith('['):
continue
elif re.match("^\d+?\.\d+?$", exp_data) is not None:
if is_loss:
cur_sum += float(exp_data)
cur_cnt += 1
plt.plot(avg_s)
plt.savefig('data/avg_loss_change.png')
plt.show()