tf.nn.conv2d()

tf.nn.conv2d(
        input,  Input tensor , The type is float16,float32,float64; The order of tensors depends on data_format To explain 
        filters,  The size of the sliding window , Shape is [filter_height, filter_width, in_channels, out_channels], Window height , wide , Number of channels entered , The number of output channels 
        strides,  Sliding window moves in steps [1, strides, strides, 1]  incorrect batch nucleus channels Calculation ,strides Is the actual moving step 
        padding,  Data fill type {
    'same','valid'},  The former uses 0 fill , The latter does not use 
        data_format='NHWC',{
    "NHWC", "NCHW"}  Respectively corresponding to N: Data volume ,H: Height of the image ,W: The width of the image ,C: Number of image channels 
        dilations=None,  Optional list ,[1,1,1,1], The length is 1-D Tensor ,input Expansion coefficient of each dimension . If k>1, There will be... Between each filter element on this dimension k-1 Skip cells 
        name=None)

tf.nn.max_pool()

tf.nn.max_pool(
       input,  Input tensor (Tensor), The type is float16,float32,float64; The order of tensors depends on data_format To explain 
       ksize,  The size of the pooled window ,[1, height, width, 1],  incorrect batch nucleus channels Pooling ,height and width Is the height and width of the pooled window 
       strides,  The step size of the pooled window movement [1, strides, strides, 1], incorrect batch nucleus channels Calculation ,strides Is the actual moving step 
       padding,  Data fill type (string){
    'same','valid'},  The former uses 0 fill , The latter does not use 
       data_format='NHWC', 'NHWC', Image format (string){
    "NHWC", "NCHW","NCHW_VECT_C"}  Respectively corresponding to N: Data volume ,H: Height of the image ,W: The width of the image ,C: Number of image channels 
       name=None
        )

tf.nn.softmax()

tf.nn.sparse_softmax_cross_entropy_with_logits(
    labels, Label tensor (Tensor), The probability distribution of the tag ,shape by [l0,l1,l2,l3...lr-1],r Is the number of categories of the label , The type is int32 perhaps int64
    logits, Predictor tensor (Tensor), Predicted probability distribution ,shape by [p0,p1,p2,p3...pr-1], among r It's the number of categories , The type is float16,float32 perhaps float64
    name=None
)

Save()

tf.train.Checkpoint().save()

preservation ckpt Model

tf.train.Checkpoint(
				    optimizer=optimizer,
				    encoder=encoder,
				    decoder=decoder
				).save(file_prefix=("./model/ckpt"))#  The untrained model can be saved to modl/ckpt Under the table of contents 

tf.keras.Model().save()

tf.keras.Model(params).save("./model/h5")
#  for example 
model = tf.keras.Model()
model.fit(
			inputs,
			outputs,
			epochs=300,
			vebose=1)
mdoel.save(model_path)

tf.keras.models.save_model()

tf.keras.models.save_model(
				model,  To save  Keras  Model examples 
				modelpath,  route (string), Location to save 
				overwrite=True,  We are any existing model that should cover the target location , You should still use manual prompts to ask the user 
				include_optimizer=True,  If  True, Save the state of the optimization program together 
				save_formate=None, “tf” or “h5”, The indication is to save the model to  Tensorflow SavedModel  still  HDF5. stay  TF 2.X  China and Murdoch think “tf”, stay  TF 1.X  China and Murdoch think “h5”
				signatures=None,  To use  SavedModel  Saved signature . Only applicable to “tf” Format . For more information , see also  tf.saved_model.save  Parameters in .signatures
				options=None （ Applies only to saved model formats ）tf.saved_model.SaveOptions  object , This object specifies to save to  SavedModel  The option to 
)

Load()

model.load_weight()

Import weight data and offset data

state = tf.train.last_checkpoint(modelpath)
model = tf.keras.Model()
model.load_weights(state)# load Weight data and offset data 

tf.keras.models.load_model()

Import the whole model, including data model and structure model

model = tf.keras.models.load_model(modelpath)