这是在处理特征降维的过程中遇到的问题，要降纬特征向量，首先要读出特征向量。我们的特征存储在字典中，key是类似于这样：CCTV1HD_4m_20180129-112144_1800_3600_2375_2525_37_0_126，但是value有下面两种形式。可以看出来，有一种value还是一个字典，有key 'fea'，而另一种value是一个list。我在代码中加入了判断数据类型的代码段。成功解决了这个问题。

{'fea': array([ 0.31566852,  0.09721687,  0.16430685, ...,  0.14451009,

        0.14061043,  0.66630179], dtype=float32), 'label': 2}

  0.44859198]

附上代码：

# -*- coding:utf-8 -*-
import pickle
import numpy as np
from sklearn.manifold import TSNE
import matplotlib.pyplot as plt
plt.switch_backend('agg')
import collections
import os
import types
import time
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
BTVX = []
a = []
label = []
path = 'D:\PycharmProject\\video_feature\Jan_videos_feature\Jan_videos\cate1'
path1 = 'D:\PycharmProject\\video_feature\Jan_videos_feature\Jan_videos\cate2'
path2 = 'D:\PycharmProject\\video_feature\Jan_videos_feature\Jan_videos\cate6'
path3 = 'D:\PycharmProject\\video_feature\Jan_videos_feature\Jan_videos\cate7'
path4 = 'D:\PycharmProject\\video_feature\Jan_videos_feature\Jan_videos\cate10'

def gen_dataset_from_pkl_feature(path,cate_value):
    count = 0
    x_dataset = []
    y_dataset = []
    x_train = collections.OrderedDict()
    for fpathe, dirs, filenames in os.walk(path):
        for filename in filenames:
            if filename.startswith('CCTV1'):
                if filename.endswith('.pkl'):
                    count += 1
                    pkl_file = os.path.join(fpathe, filename)
                    pkl_fp = open(pkl_file, 'rb')
                    x_train.update(pickle.load(pkl_fp))
                    # print x_train
    x_values = x_train.values()
    for i in range(len(x_values)):
        if type(x_values[i]) == dict:
            x_dataset.append(x_values[i]['fea'])
        else:
            x_dataset.append(x_values[i])
    x_dataset = np.array(x_dataset)
    print x_dataset.shape
    y_dataset = [cate_value for i in range(len(x_train))]
    y_dataset = np.array(y_dataset)
    print y_dataset.shape
        # y_dataset.append(x_train[i]['label'])
    x_dataset = np.array(x_dataset[:500])
    y_dataset = np.array(y_dataset[:500])
    # print(len(x_train))
    # print x_dataset.shape
    return x_dataset, y_dataset,count


x_dataset_cate1,y_dataset_cate1,count1 = gen_dataset_from_pkl_feature(path,1)
x_dataset_cate2,y_dataset_cate2,count2= gen_dataset_from_pkl_feature(path1,2)
x_dataset_cate6,y_dataset_cate6,count6 = gen_dataset_from_pkl_feature(path2,6)
x_dataset_cate7,y_dataset_cate7,count7 = gen_dataset_from_pkl_feature(path3,7)
x_dataset_cate10,y_dataset_cate10,count10 = gen_dataset_from_pkl_feature(path4,10)
# print x_dataset_cate1.shape
# x_dataset = x_dataset_cate2
# y_dataset = y_dataset_cate2
x_dataset = np.concatenate((x_dataset_cate1,x_dataset_cate2,x_dataset_cate6,x_dataset_cate7,x_dataset_cate10),axis=0)
y_dataset = np.concatenate((y_dataset_cate1,y_dataset_cate2,y_dataset_cate6,y_dataset_cate7,y_dataset_cate10),axis=0)
xx = []
for i in range(0,len(x_dataset)): #正则化
    xx.append((x_dataset[i] - np.min(x_dataset)) / (np.max(x_dataset) - np.min(x_dataset)))
x_dataset = xx
tsne = TSNE(n_components=2)
start_time = time.time()
tSNE = tsne.fit_transform(x_dataset)
fig = plt.figure(figsize=(10,10))
ax = plt.subplot(111)
plt.title('CC-Cate1,2,6,7,10:t-SNE(time: %.3fs)'%(time.time() - start_time))
# #画散点图
# cValue = ['r','y','g','b','r','y','g','b','r']
# ax1.scatter(x,y,c=cValue,marker='s')
plt.scatter(tSNE[:, 0], tSNE[:, 1],c=y_dataset)
for i in range(tSNE.shape[0]):
    plt.text(tSNE[i, 0], tSNE[i, 1], str(y_dataset[i]))
plt.savefig('t-SNE.png')