©2020-2024 ioDraw All rights reserved,
Privacy Policy
包括BN层、卷积层、池化层、交叉熵、随机梯度下降法、非极大抑制、k均值聚类等秋招常见的代码实现。
<>1. BN层
import numpy as np def batch_norm(outputs, gamma, beta, epsilon=1e-6, momentum=
0.9, running_mean=0, running_var=1): ''' :param outputs: [B, L] :param gamma:
mean :param beta: :param epsilon: :return: ''' mean = np.mean(outputs, axis=(0,
2, 3), keepdims=True) # 1, C, H, W var = np.var(outputs, axis=(0,2,3), keepdims=
True) # 1, C, H, W # mean = np.mean(outputs, axis=0) # var = np.var(outputs,
axis=0) # 滑动平均用于记录mean和var,用于测试 running_mean = momentum * running_mean + (1-
momentum) * mean running_var = momentum * running_var + (1-momentum) * var res =
gamma* ( outputs - mean ) / np.sqrt(var + epsilon) + beta return res,
running_mean, running_var if __name__ == '__main__': outputs = np.random.random(
(16, 64, 8, 8)) tmp, _, _ = batch_norm(outputs, 1, 1, 1e-6) # print(tmp.shape)
mean= np.mean(tmp[:, 1, :, :]) std = np.sqrt(np.var(tmp[:, 1, :, :])) print(mean
, std)
<>2. 卷积层
import numpy as np def conv_forward_naive(x, w, b, conv_param): ''' :param x:
[N, C_in, H, W] :param w: [C_out, C_in, k1, k2] :param b: [C_out] :param
conv_param: - 'stride': - 'pad': the number of pixels that will be used to
zero-pad the input :return: - 'out': (N, C_out, H', W') - 'cache': (x, w, b,
conv_param) ''' out = None N, C_in, H, W = x.shape C_out, _, k1, k2 = w.shape
stride, padding = conv_param['stride'], conv_param['pad'] H_out = (H-k1+2*
padding) // stride + 1 W_out = (W-k2+2*padding) // stride + 1 out = np.zeros((N,
C_out, H_out, W_out)) x_pad = np.zeros((N, C_in, H+2*padding, W+2*padding))
x_pad[:, :, padding:padding+H, padding:padding+W] = x for i in range(H_out): for
jin range(W_out): x_pad_mask = x_pad[:, :, i*stride:i*stride+k1, j*stride:j*
stride+k2] for c in range(C_out): out[:, c, i, j] = np.sum(x_pad_mask*w[c, :, :,
:], axis=(1,2,3)) out += b[None, :, None, None] cache = (x, w, b, conv_param)
return out, cache
<>3. maxpooling
import numpy as np def maxpooling_forward(feature, kernel, stride): ''' :param
feature: [N, C, H, W] :param kernel: [k1, k2] :param stride: [s1, s2] :return:
''' N, C, H, W = feature.shape k1, k2 = kernel s1, s2 = stride H_out = (H - k1)
// s1 + 1 W_out = (W - k2) // s2 + 1 out = np.zeros((N, C, H_out, W_out)) for i
in range(H_out): for j in range(W_out): feature_mask = feature[:, :, i*s1:i*s1+
k1, j*s2:j*s2+k2] out[:, :, i, j] = np.max(feature_mask, axis=(2,3)) #
注意这里的2,3!!! return out
<>4. cross Entropy
import numpy as np def cross_entropy(label, outputs, reduce=True): ''' :param
label: B x 1 :param outputs: B x c :return: loss ''' loss_list = [] for i in
range(len(label)): y = label[i] output = outputs[i] sum_exp = np.sum([np.exp(k)
for k in output]) prop = np.exp(output[y]) / sum_exp loss_list.append(-np.log(
prop)) if reduce: return np.mean(loss_list) else: return np.sum(loss_list) def
softmax(t): return np.exp(t) / np.sum(np.exp(t), axis=1, keepdims=True) def
softmax2(t): return np.exp(t) / np.sum(np.exp(t), axis=1, keepdims=True) def
cross_entropy_2(y, y_, onehot=True, reduce=True): y = softmax(y) if not onehot:
cates= y.shape y_ = np.eye(cates[-1])[y_] if reduce: return np.mean(-np.sum(y_ *
np.log(y), axis=1)) else: return np.sum(-np.sum(y_ * np.log(y), axis=1)) if
__name__== '__main__': outputs = [[0.5, 0.5], [0, 1], [1, 0]] label = [0, 0, 1]
print(cross_entropy(label, outputs, True)) print(cross_entropy_2(outputs, label,
False))
<>5. sgd
import numpy as np import random class MYSGD: def __init__(self, training_data,
epochs, batch_size, lr, model): self.training_data = training_data self.epochs
= epochs self.batch_size = batch_size self.lr = lr self.weight = [...] self.bias
= [...] def run(self): n = len(self.training_data) for j in range(self.epochs):
random.shuffle(self.training_data) mini_batches = [self.training_data[k*self.
batch_size: (k+1)*self.batch_size] for k in range(n//self.batch_size)] for
mini_batchin mini_batches: self.updata(mini_batch) def update(self, mini_batch):
nabla_b= [np.zeros(b.shape) for b in self.bias] nabla_w = [np.zeros(w.shape)
for w in self.weight] for x, y in mini_batch: delta_nabla_b, delta_nabla_w =
self.backprop(x, y) nabla_b = [nb+dnb for nb, dnb in zip(nabla_b, delta_nabla_b)
] nabla_w = [nw+dnw for nw, dnw in zip(nabla_w, delta_nabla_w)] self.weight = [w
-(self.eta/len(mini_batch))*nw for w, nw in zip(self.weight, nabla_w)] self.bias
= [b-(self.eta/len(mini_batch))*nb for b, nb in zip(self.bias, nabla_b)] def
backprop(self, x, y):
<>6. nms
import numpy as np def iou_calculate(bbox1, bbox2, mode='x1y1x2y2'): # 我的 x11,
y11, x12, y12 = bbox1 x21, y21, x22, y22 = bbox2 area1 = (y12-y11+1)*(x12-x11+1)
area2= (y22-y21+1)*(x22-x21+1) overlap = max(min(y12, y22) - max(y11, y21) + 1,
0) * max(min(x12, x22) - max(x11, x21) + 1, 0) return overlap / (area2 + area1 -
overlap+ 1e-6) def bb_intersection_over_union(boxA, boxB): # 别人的 boxA = [int(x)
for x in boxA] boxB = [int(x) for x in boxB] xA = max(boxA[0], boxB[0]) yA = max
(boxA[1], boxB[1]) xB = min(boxA[2], boxB[2]) yB = min(boxA[3], boxB[3])
interArea= max(0, xB - xA + 1) * max(0, yB - yA + 1) boxAArea = (boxA[2] - boxA[
0] + 1) * (boxA[3] - boxA[1] + 1) boxBArea = (boxB[2] - boxB[0] + 1) * (boxB[3]
- boxB[1] + 1) iou = interArea / float(boxAArea + boxBArea - interArea) return
ioudef nms(outputs, scores, T): ''' :param outputs: bboxes, x1y1x2y2 :param
scores: confidence of each bbox :param T: threshold :return: ''' # 我的 outputs =
np.array(outputs)[np.argsort(-np.array(scores))] saved = [True for _ in range(
outputs.shape[0])] for i in range(outputs.shape[0]): if saved[i]: for j in range
(i+1, outputs.shape[0]): if saved[j]: iou = iou_calculate(outputs[i], outputs[j]
) if iou >= T: saved[j] = False scores = np.sort(-np.array(scores)) return
outputs[saved], -scores[saved] # 别人的 def nms_others(bboxes, scores, iou_thresh):
""" :param bboxes: 检测框列表 :param scores: 置信度列表 :param iou_thresh: IOU阈值 :return:
""" x1 = bboxes[:, 0] y1 = bboxes[:, 1] x2 = bboxes[:, 2] y2 = bboxes[:, 3]
areas= (y2 - y1) * (x2 - x1) # 结果列表 result = [] index = scores.argsort()[::-1]
# 对检测框按照置信度进行从高到低的排序，并获取索引 # 下面的操作为了安全，都是对索引处理 while index.size > 0: #
当检测框不为空一直循环 i = index[0] result.append(i) # 将置信度最高的加入结果列表 # 计算其他边界框与该边界框的IOU x11
= np.maximum(x1[i], x1[index[1:]]) y11 = np.maximum(y1[i], y1[index[1:]]) x22 =
np.minimum(x2[i], x2[index[1:]]) y22 = np.minimum(y2[i], y2[index[1:]]) w = np.
maximum(0, x22 - x11 + 1) # 两个边重叠时,也有1列/行像素点是重叠的 h = np.maximum(0, y22 - y11 + 1
) overlaps = w * h ious = overlaps / (areas[i] + areas[index[1:]] - overlaps) #
只保留满足IOU阈值的索引 idx = np.where(ious <= iou_thresh)[0] index = index[idx + 1] #
处理剩余的边框 bboxes, scores = bboxes[result], scores[result] return bboxes, scores
def mynms(bboxes, scores, iou_T): x1 = bboxes[:, 0] y1 = bboxes[:, 1] x2 =
bboxes[:, 2] y2 = bboxes[:, 3] areas = (y2-y1+1) * (x2-x1+1) ids = np.argsort(
scores)[::-1] res = [] while len(ids) > 0: i = ids[0] res.append(i) x11 = np.
maximum(x1[i], x1[ids[1:]]) x22 = np.minimum(x2[i], x2[ids[1:]]) y11 = np.
maximum(y1[i], y1[ids[1:]]) y22 = np.minimum(y2[i], y1[ids[1:]]) #
np.maximum(X,Y,None) : X与Y逐位取最大者. 最少两个参数 overlap = np.maximum(x22-x11+1, 0) * np
.maximum(y22-y11+1, 0) iou = overlap / (areas[i] +areas[ids[1:]] - overlap) ids
= ids[1:][iou<T] return bboxes[res], scores[res] if __name__ == '__main__':
outputs= [[10, 10, 20, 20], [15, 15, 25, 25], [9, 15, 25, 13]] scores = [0.6,
0.8, 0.7] T = 0.1 print(nms(outputs, scores, T)) print(nms_others(np.array(
outputs), np.array(scores), T)) print(mynms(np.array(outputs), np.array(scores),
T))
<>7. k-means
import numpy as np import copy def check(clusters_last, clusters_center): #
clusters_last.sort() # clusters_center.sort() if len(clusters_last) == 0: return
False for c1, c2 in zip(clusters_last, clusters_center): if np.linalg.norm(c1 -
c2) > 0: return False return True def kMeans(data, k): ''' :param data: [n, c]
:param k: the number of clusters :return: ''' clusters_last = [] clusters_center
= [data[i] for i in range(k)] # random choosed while not check(clusters_last,
clusters_center): clusters_last = copy.deepcopy(clusters_center) clusters = [[]
for _ in range(k)] for i in range(data.shape[0]): min_dis = float('inf') for j,
centerin enumerate(clusters_center): distance = np.linalg.norm(center-data[i])
if distance < min_dis: min_dis = distance idx = j clusters[idx].append(data[i])
clusters_center= [] for i in range(k): clusters_center.append(np.mean(clusters[i
], axis=0)) return clusters_center def kMeans2(data, k): ''' :param data: [n,
c] :param k: the number of clusters :return: ''' clusters_last = []
clusters_center= copy.deepcopy(data[:k]) # random choosed while not check(
clusters_last, clusters_center): clusters_last = copy.deepcopy(clusters_center)
clusters= [[] for _ in range(k)] for i in range(data.shape[0]): distance = np.
linalg.norm(clusters_center - data[i], axis=1) idx = np.argmin(distance)
clusters[idx].append(data[i]) clusters_center = [] for i in range(k):
clusters_center.append(np.mean(clusters[i], axis=0)) clusters_center = np.array(
clusters_center) return clusters_center if __name__ == '__main__': data = np.
random.random((20, 2)) print(kMeans(data, 5)) print(kMeans2(data, 5))