Question

我在将处理后的4D NumPy数组广播到4D数组时遇到问题。我已比较尺寸以检查尺寸是否有问题。我不明白怎么了。

train_path = files
train_file_names = os.listdir(train_path)
train_file_names.sort(key=lambda x: int(x.partition('.')[0]))
seg_num = 60
seg_len = 2
sample_num = len(df)
data = np.zeros((seg_num*100, 496, 64, 1))
label = np.zeros((seg_num * sample_num,))

for i, file_name in enumerate(train_file_names):
        sr, sound_file = wavfile.read(train_path + file_name)
        # print(train_path+file_name)
        length = sr * seg_len           # 5s segment
        range_high = len(sound_file) - length
        print(range_high, length)
        random_start = np.random.randint(range_high, size=seg_num)
        print("i", i)
        for j in range(seg_num):
            cur_wav = sound_file[random_start[j]:random_start[j] + length]
            cur_wav = cur_wav / 32768.0
            cur_spectro = preprocess_sound(cur_wav, sr)
            #print(cur_spectro.shape)
            cur_spectro = np.copy(np.expand_dims(cur_spectro, 3))
            print("cur_spectro",cur_spectro.shape)
            print("data", data.shape)
            print(data[i * seg_num + j, :, :, :].shape)
            data[i * seg_num + j, :, :, :] = cur_spectro
            label[i * seg_num + j] = df['class'][i]

输出

88200 88200
i 0
cur_spectro (0, 496, 64, 1)
data (6000, 496, 64, 1)
(496, 64, 1)
<ipython-input-226-30eefc542ce4> in loading_data(files, df)
      29                 print("data", data.shape)
      30                 print(data[i * seg_num + j, :, :, :].shape)
 ---> 31                 data[i * seg_num + j, :, :, :] = cur_spectro
      32                 label[i * seg_num + j] = df['class'][i]
      33 ValueError: could not broadcast input array from shape (0,496,64,1) into shape (496,64,1)

Answer 1

您的打印输出之一是cur_spectro (0, 496, 64, 1)。

请参见What does a numpy shape starting with zero mean 它包含了如何理解情况的描述数组维度之一为零。实际上，这意味着所讨论的数组为空。

因此，您似乎试图将空数组复制到一个非空数组。由于在这种情况下，这些数组不能对齐（广播），则会引发异常。

就像您读到的评论之一一样，分析原因 preprocess_sound 函数返回为空数组。

也许您应该在代码中添加一个检查 cur_spectro 的尺寸为零，在这种情况下如果您应该跳过有问题的说明，或者将一些“代理”内容复制到目标数组。

广播4D numpy数组

1 个答案: