# 記憶體的資料映射討論
## numpy data mapping
- resize 1張照片成300x300 (dtype uint8)
- 300x300x3 x 1byte = 0.26 MB
- resize 1000張照片成300x300 (dtype uint8)
- 1000x300x300x3 x 1byte = 257 MB
- resize 1000張 照片成300x300 (dtype float32)
- 1000x300x300x3 x 4byte = 1030 MB
- code

    import numpy as np
    import sys
    data_1 = np.ones((1, 300, 300, 3), dtype='uint8')
    data_1000 = np.ones((1000, 300, 300, 3), dtype='uint8')

    data_1_size = getsizeof(data_1)/1024/1024
    data_1000_size = getsizeof(data_1000)/1024/1024
    print('data 1 type: {} \t data 1000 type: {}'.format(data_1.dtype, data_1000.dtype))
    print('data 1 size: {}MB \t data 1000 size: {}MB'.format( round(data_1_size, 2), round(data_1000_size, 2)))

    data_1 = data_1.astype('float32')
    data_1000 = data_1000.astype('float32')
    data_1_size = getsizeof(data_1)/1024/1024
    data_1000_size = getsizeof(data_1000)/1024/1024

    print('\ndata 1 type: {} \t data 1000 type: {}'.format(data_1.dtype, data_1000.dtype))
    print('data 1 size: {}MB \t data 1000 size: {}MB'.format( round(data_1_size, 2), round(data_1000_size, 2)))

- output

## training data mapping
### training input data:
- 1000張x300x300x3x4=1030MB
(4是因為訓練會用float32佔4byte)
## training weights:
- model
- 假設使用Xception (84MB)
- 模型84MBx2(forward&backward 168MB)
- batch operations
- batchsize=1, 4, 16, 32, 64
- 記憶體佔 1x168MB
- 記憶體佔 4x168MB
- 記憶體佔16x168MB=2.6GB
- 記憶體佔32x168MB=5.2GB
- 記憶體佔64x168MB=10 GB