import numpy
def loadMnist() -> (numpy.ndarray,numpy.ndarray,numpy.ndarray,numpy.ndarray):
    """
    :return:  (xTrain,yTrain,xTest,yTest)
    """
    global _TRAIN_SAMPLE_CNT
    global PIC_H
    global PIC_W
    global _TEST_SAMPLE_CNT
    global PIC_HW
    from tensorflow import keras #修改點(diǎn): tensorflow:2.6.2,keras:2.6.0 此版本下,  import keras 換成 from tensorflow import keras
    import tensorflow
    print(f"keras.__version__:{keras.__version__}")#2.6.0
    print(f"tensorflow.__version__:{tensorflow.__version__}")#2.6.2
    # avatar_img_path = "/kaggle/working/data"

    import os
    import cv2
    xTrain:numpy.ndarray; label_train:numpy.ndarray; xTest:numpy.ndarray; label_test:numpy.ndarray
    yTrain:numpy.ndarray; yTest:numpy.ndarray
    #%userprofile%\.keras\datasets\mnist.npz
    (xTrain, label_train), (xTest, label_test) = keras.datasets.mnist.load_data()
    # x_train.shape,y_train.shape, x_test.shape, label_test.shape
    # (60000, 28, 28), (60000,), (10000, 28, 28), (10000,)
    _TRAIN_SAMPLE_CNT,PIC_H,PIC_W=xTrain.shape
    PIC_HW=PIC_H*PIC_W
    xTrain=xTrain.reshape((-1, PIC_H * PIC_W))
    xTest=xTest.reshape((-1, PIC_H * PIC_W))
    _TEST_SAMPLE_CNT=label_test.shape[0]

    from sklearn import preprocessing

    #pytorch 的 y 不需要 oneHot
    #_label_train是1列多行的樣子.  _label_train.shape : (60000, 1)
    yTrain=label_train
    # y_train.shape:(60000) ; y_train.dtype: dtype('int')
    CLASS_CNT=yTrain.shape[0]
    yTest=label_test
    # y_test.shape:(10000) ; y_test.dtype: dtype('int')
    xTrainMinMaxScaler:preprocessing.MinMaxScaler; xTestMinMaxScaler:preprocessing.MinMaxScaler
    xTrainMinMaxScaler=preprocessing.MinMaxScaler()
    xTestMinMaxScaler=preprocessing.MinMaxScaler()
    # x_train.dtype: dtype('uint8') -> dtype('float64')
    xTrain=xTrainMinMaxScaler.fit_transform(xTrain)
    # x_test.dtype: dtype('uint8') -> dtype('float64')
    xTest = xTestMinMaxScaler.fit_transform(xTest)
    return (xTrain,yTrain,xTest,yTest)

xTrain:torch.Tensor;yTrain:torch.Tensor; xTest:torch.Tensor; yTest:torch.Tensor(xTrain,yTrain,xTest,yTest)=loadMnist()

import plotly.express
import plotly.graph_objects
import plotly.subplots
import numpy
xTrain:numpy.ndarray=numpy.random.random((2,28,28))
#xTrain[0].shape:(28,28)
#fig:plotly.graph_objects.Figure=None
fig=plotly.subplots.make_subplots(rows=1,cols=2,shared_xaxes=True,shared_yaxes=True) #共1行2列
fig.add_trace(trace=plotly.express.imshow(img=xTrain[0]).data[0],row=1,col=1) #第1行第1列
fig.add_trace(trace=plotly.express.imshow(img=xTrain[1]).data[0],row=1,col=2) #第1行第2列
fig.show()
#參數(shù)row、col從1開(kāi)始，  不是從0開(kāi)始的

import numpy
xTrain:numpy.ndarray=numpy.random.random((2,28,28))
#xTrain[0].shape:(28,28)
import plotly.express
import plotly.graph_objects
plotly.express.imshow(img=xTrain[0]).show()
#其中plotly.express.imshow(img=xTrain[0]) 的類型是 plotly.graph_objects.Figure

#mnist單樣本分割 分割成4*4小格子顯示出來(lái), 以確認(rèn)分割的對(duì)不對(duì)。 以下代碼是正確的分割。 主要邏輯是：   (7,4,7,4)   [h, :, w, :] 
fig:plotly.graph_objects.Figure=plotly.subplots.make_subplots(rows=7,cols=7,shared_xaxes=True,shared_yaxes=True,vertical_spacing=0,horizontal_spacing=0)
xTrain0Img:torch.Tensor=xTrain[0].reshape((PIC_H,PIC_W))
plotly.express.imshow(img=xTrain0Img).show()
xTrain0ImgCells:torch.Tensor=xTrain0Img.reshape((7,4,7,4))
for h in range(7):
    for w in range(7):
        print(f"h,w:{h},{w}")
        fig.add_trace(trace=plotly.express.imshow(xTrain0ImgCells[h,:,w,:]).data[0],col=h+1,row=w+1)
fig.show()

    """
    mnist單樣本分割 分割成4*4小格子顯示出來(lái), 重點(diǎn)邏輯是： (7, 4, 7, 4)  [h, :, w, :]
    :param xTrain:
    :return:
    """
    fig: plotly.graph_objects.Figure = plotly.subplots.make_subplots(rows=7, cols=7, shared_xaxes=True,  shared_yaxes=True, vertical_spacing=0,  horizontal_spacing=0)
    xTrain0Img: torch.Tensor = xTrain[0].reshape((PIC_H, PIC_W))
    plotly.express.imshow(img=xTrain0Img).show()
    xTrain0ImgCells: torch.Tensor = xTrain0Img.reshape((7, 4, 7, 4))
    xTrain0ImgCells=torch.swapdims(input=xTrain0ImgCells,dim0=1,dim1=2)#交換 (7, 4, 7, 4) 維度1、維度2 即 (0:7, 1:4, 2:7, 3:4)
    for h in range(7):
        for w in range(7):
            print(f"h,w:{h},{w}")
            fig.add_trace(trace=plotly.express.imshow(xTrain0ImgCells[h, w]).data[0], col=h + 1, row=w + 1) # [h, w, :, :] 或 [h, w]
    fig.show()

# mnist單樣本錯(cuò)誤的分拆顯示:
    fig: plotly.graph_objects.Figure = plotly.subplots.make_subplots(rows=7, cols=7, shared_xaxes=True,  shared_yaxes=True, vertical_spacing=0,  horizontal_spacing=0)
    xTrain0Img: torch.Tensor = xTrain[0].reshape((PIC_H, PIC_W))
    plotly.express.imshow(img=xTrain0Img).show()
    xTrain0ImgCells: torch.Tensor = xTrain0Img.reshape((4,7, 4, 7))  #原本是: (7,4,7,4)
    for h in range(7):
        for w in range(7):
            print(f"h,w:{h},{w}")
            fig.add_trace(trace=plotly.express.imshow(xTrain0ImgCells[:, h,  :, w]).data[0], col=h + 1, row=w + 1)  #原本是: [h,:,w,:]
    fig.show()