utils.py
from torch import nn
import torch.nn.functional as F
import numpy as np
from PIL import Image
import os
import shutil
from config import cfg
import pdb
import math
#初始化權重，models可以是包含多個在內部具有nn.Conv2d的實例的list，也可以是一個實例
#遞歸對models內部的model進行權重的初始化
def weights_normal_init(*models):
for model in models:
dev=0.01
if isinstance(model, list):#如果m是個list
for m in model:
weights_normal_init(m, dev)
else: #如果m是個實例
for m in model.modules():
if isinstance(m, nn.Conv2d):
m.weight.data.normal_(0.0, dev)
if m.bias is not None:
m.bias.data.fill_(0.0)
elif isinstance(m, nn.Linear):
m.weight.data.normal_(0.0, dev)
def initialize_weights(*models):#也是初始化權重和bias的函數，不過這個函數用卷積核和通道數的大小對方差進行調整
for model in models:
if isinstance(model, list):
for m in model:
initialize_weights(m)
else:
for module in model.modules():
if isinstance(module, nn.Conv2d):
n = module.kernel_size[0] * module.kernel_size[1] * module.out_channels
module.weight.data.normal_(0, math.sqrt(2. / n))
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.BatchNorm2d):
module.weight.data.fill_(1)
module.bias.data.zero_()
elif isinstance(module, nn.Linear):
n = module.weight.size(1)
module.weight.data.normal_(0, math.sqrt(2. / n))
module.bias.data.zero_()
def weights_init_kaiming(*models):#用作者名字命名的初始化權重函數，很皮
for model in models:
if isinstance(model, list):
for m in model:
weights_init_kaiming(m)
else:
for module in model.modules():
if isinstance(module, nn.Conv2d):
#kaiming is first name of author whose last name is 'He' lol
nn.init.kaiming_uniform(module.weight)
module.bias.data.zero_()
def adjust_learning_rate(lr, decay, optimizer, cur_epoch, n_epochs):#學習率衰減函數
"""Sets the learning rate to the initially
configured `lr` decayed by `decay` every `n_epochs`"""
new_lr = lr * (decay ** (cur_epoch // n_epochs))#衰減
for param_group in optimizer.param_groups:
param_group['lr'] = new_lr
def calculate_mean_iu(predictions, gts, num_classes):#計算每個類別預測準確度的平均值
sum_iu = 0
class_iu = []
for i in xrange(num_classes):
n_ii = t_i = sum_n_ji = 1e-9
for p, gt in zip(predictions, gts):
n_ii += np.sum(gt[p == i] == i)#好像是預測正確的個數，但是沒太看懂
t_i += np.sum(gt == i)#i等於groundtruth
sum_n_ji += np.sum(p == i)#i等於預測值
sum_iu += float(n_ii) / (t_i + sum_n_ji - n_ii)#第i類的準確度
class_iu.append(float(n_ii) / (t_i + sum_n_ji - n_ii))
mean_iu = sum_iu / num_classes#平均正確率
return mean_iu,class_iu
def calculate_lane_metrics(predictions, gts, num_classes):#計算預測準確率矩陣
sum_iu = 0
class_iu = []
acc = []
rec = []
f1_m = []
for i in xrange(num_classes):
tp = fp = fn = 0.
for p, gt in zip(predictions, gts):
tp += np.sum(gt[p == i] == i)#預測正確
fp += np.sum( (gt[p == i] != i ))#不屬於第i類但被預測爲第i類
fn += np.sum(gt[p != i] == i)#屬於第i類但被預測爲其他類
class_iu.append(tp / (tp + fp + fn + 1e-9))
acc.append(tp/(tp+fp+1e-9))#準確率
rec.append(tp/(tp+fn+1e-9))#召回率
f1_m.append(2*acc[i]*rec[i]/(acc[i]+rec[i]+1e-9))#F1值
sum_iu += tp / (tp + fp + fn + 1e-9)
mean_iu = sum_iu / num_classes
return {'miu':mean_iu,
'ciu':class_iu
},\
{'acc':acc,
'rec':rec,
'f1_m':f1_m
}
class CrossEntropyLoss2d(nn.Module):#交叉熵損失函數
def __init__(self, weight=None, size_average=True):
super(CrossEntropyLoss2d, self).__init__()
self.nll_loss = nn.NLLLoss2d(weight, size_average)
def forward(self, inputs, targets):
return self.nll_loss(F.log_softmax(inputs), targets)
def rmrf_mkdir(dir_name):
if os.path.exists(dir_name):
shutil.rmtree(dir_name)#遞歸刪除dir_name的全部目錄
os.mkdir(dir_name)
def rm_file(path_file):
if os.path.exists(path_file):
os.remove(path_file)
def colorize_mask(mask):#將圖像轉換爲P模式，其中每個像素點由8bit數字定義
# mask: numpy array of the mask
new_mask = Image.fromarray(mask.astype(np.uint8)).convert('P')
new_mask.putpalette(cfg.VIS.PALETTE_LABEL_COLORS)#放置調色板(?)
return new_mask
def _fast_hist(label_pred, label_true, num_classes):
mask = (label_true >= 0) & (label_true < num_classes)
hist = np.bincount(
num_classes * label_true[mask].astype(int) +
label_pred[mask], minlength=num_classes ** 2).reshape(num_classes, num_classes)
return hist
def scores(predictions, gts, num_classes):
hist = np.zeros((num_classes, num_classes))
for lp, lt in zip(predictions, gts):
hist += _fast_hist(lp.flatten(), lt.flatten(), num_classes)#flatten函數能將多維函數轉化爲一維函數
# axis 0: gt, axis 1: prediction
# acc = np.diag(hist).sum() / hist.sum()
# acc_cls = np.diag(hist) / hist.sum(axis=1)
# acc_cls = np.nanmean(acc_cls)
iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
mean_iu = np.nanmean(iu)
# freq = hist.sum(axis=1) / hist.sum()
# fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
# return acc, acc_cls, mean_iu, fwavacc
# pdb.set_trace()
return {'miu':mean_iu,
'ciu':iu
}
def acc(predictions, gts, num_classes=2):
predictions = predictions.data.numpy().astype(np.int64)
gts = gts.data.numpy().astype(np.int64)
predictions[predictions>=0.5]=1
predictions[predictions<0.5]=0
t = predictions==gts
acc = np.sum(t)/float((t.shape[0]*t.shape[1]))
return acc
def streaming_scores(hist,predictions, gts, num_classes):
for lp, lt in zip(predictions, gts):
hist += _fast_hist(lp.flatten(), lt.flatten(), num_classes)
# axis 0: gt, axis 1: prediction
# acc = np.diag(hist).sum() / hist.sum()
# acc_cls = np.diag(hist) / hist.sum(axis=1)
# acc_cls = np.nanmean(acc_cls)
iu = np.diag(hist) / (hist.sum(axis=1) + hist.sum(axis=0) - np.diag(hist))
mean_iu = np.nanmean(iu)
# freq = hist.sum(axis=1) / hist.sum()
# fwavacc = (freq[freq > 0] * iu[freq > 0]).sum()
# return hist, acc, acc_cls, mean_iu, fwavacc
return hist, \
{'miu':mean_iu,
'ciu':iu
}
本文主要供個人學習筆記之用。與諸君共勉