sklearn.utils.class_weight.compute_class_weight

 https://blog.csdn.net/FY_2018/article/details/116951278

compute_class_weight這個函數的作用是對於輸入的樣本，平衡類別之間的權重，下面寫段測試代碼測試這個函數：

# coding:utf-8
 
from sklearn.utils.class_weight import compute_class_weight
 
class_weight = 'balanced'
label = [0] * 9 + [1]*1 + [2, 2]
print(label) # [0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 2]
classes=[0, 1, 2]
weight = compute_class_weight(class_weight, classes, label)
print(weight) #[ 0.44444444 4.         2.        ]
print(.44444444 * 9) # 3.99999996
print(4 * 1) # 4
print(2 * 2) # 4

如上圖所示，可以看到這個函數把樣本的平衡後的權重乘積爲4，每個類別均如此。banlanced的計算公式爲：n_samples/n_classes/np.bincount(y)。n_samples表示樣本總數，n_classes表示總類別數量，np.bincount(y)輸出所有類別的每個類別的樣本數量，y是所有樣本的標籤。一個標籤代表一個類別。採用balanced模型時，每種類別的權重爲n_samples/n_classes，即12/3=4；然後根據每種類別中的樣本數量對每個樣本進行平均分配權重，即4/9=0.444, 4/1=4, 4/2=2。0類別有9個樣本，1類別有1個樣本，2類別有2個樣本。

 

 

 

 

#calculate class weights
class_weights = class_weight.compute_class_weight( class_weight ='balanced',
                                                   classes =np.unique(y_train),
                                                  y =y_train.flatten())

 

Type:        module
String form: <module 'sklearn.utils.class_weight' from '/home/software/anaconda3/envs/tf115/lib/python3.7/site-packages/sklearn/utils/class_weight.py'>

 

sklearn.utils.class_weight.compute_class_weight

sklearn.utils.class_weight.compute_class_weight(class_weight, classes, y)[source]

Estimate class weights for unbalanced datasets.

Parameters

class_weight   dict, ‘balanced’ or None

If ‘balanced’, class weights will be given by n_samples / (n_classes * np.bincount(y)). If a dictionary is given, keys are classes and values are corresponding class weights. If None is given, the class weights will be uniform.

classes   ndarray

Array of the classes occurring in the data, as given by np.unique(y_org) with y_org the original class labels.

y  array-like, shape (n_samples,)

Array of original class labels per sample;

Returns

class_weight_vectndarray, shape (n_classes,)

Array with class_weight_vect[i] the weight for i-th class

References

The “balanced” heuristic is inspired by Logistic Regression in Rare Events Data, King, Zen, 2001.

 

REF

https://scikit-learn.org/0.22/modules/generated/sklearn.utils.class_weight.compute_class_weight.html

==================================================================================

import tensorflow as tf

import keras.backend as K
from keras.models import Sequential, Model
from keras.layers import Input, Dense, Dropout, Conv2D, MaxPooling2D, BatchNormalization, Flatten, GlobalAveragePooling2D, Multiply
from keras.callbacks import EarlyStopping, ModelCheckpoint
from keras.optimizers import Adam
from keras import backend as K
from keras.engine.topology import Layer, InputSpec
from keras.utils import Sequence, plot_model
from keras.constraints import unit_norm
from keras import regularizers

 

def SqueezeExcite(tensor, ratio=16):
    nb_channel = K.int_shape(tensor)[-1]  # 獲得tensor的形狀； -1 表示最後一個維度的長度，此處指圖像的通道（5,400,1）
    x = GlobalAveragePooling2D()(tensor) ## 每個通道進行求和平均；實際上只有一個通道；

 
    x = Dense(nb_channel // ratio, activation='relu')(x)

 
    x = Dense(nb_channel, activation='sigmoid')(x)

 
    x = Multiply()([tensor, x])

 
    return x

 

def create_model(width=200):  ## width=400
    K.clear_session()  ## import keras.backend as K
    pool2_list = []
    merge_list = []

    input_size = Input(shape=(5, width, 1))     ## 輸入層 5,400,1  圖像 寬 高 通道    keras.layers.Input

 

 
    conv1_ = Conv2D(128, (5, 10), padding='same',activation='relu')(input_size) ## 128個核； 5*10大小的核；“same”代表保留邊界處的卷積結果，通常會導致輸出shape與輸入shape相同。

 

 
    conv1  = SqueezeExcite(conv1_)
    conv2_ = Conv2D(64, (5, 10), padding='same',activation='relu')(conv1)
    conv2  = SqueezeExcite(conv2_)
    conv3_ = Conv2D(64, (5, 10), padding='same',activation='relu')(conv2)
    conv3  = SqueezeExcite(conv3_)
    conv4_ = Conv2D(128, (5, 10), padding='valid',activation='relu')(conv3)
    conv4  = SqueezeExcite(conv4_)
    pool1  = MaxPooling2D(pool_size=(1, 2))(conv4)
    conv5_ = Conv2D(64, (1, 4), padding='same',activation='relu')(pool1)
    conv5  = SqueezeExcite(conv5_)
    conv6_ = Conv2D(64, (1, 4), padding='same',activation='relu')(conv5)
    conv6  = SqueezeExcite(conv6_)
    conv7_ = Conv2D(128, (1, 4), padding='same',activation='relu')(conv6)
    conv7  = SqueezeExcite(conv7_)
    pool2  = MaxPooling2D(pool_size=(1, 2))(conv7)

    x = Flatten()(pool2)
    dense1 = Dense(256, activation='relu')(x)
    x = Dropout(0.4)(dense1)
    pred_output = Dense(1, activation='sigmoid')(x)
    model = Model(input=[input_size], output=[pred_output])
    model.summary()

    return model

 

 

 

==================================================================================

Init signature:
EarlyStopping(
    monitor='val_loss',
    min_delta=0,
    patience=0,
    verbose=0,
    mode='auto',
    baseline=None,
    restore_best_weights=False,
)
Docstring:     
Stop training when a monitored quantity has stopped improving.

# Arguments
    monitor: quantity to be monitored.
    min_delta: minimum change in the monitored quantity
        to qualify as an improvement, i.e. an absolute
        change of less than min_delta, will count as no
        improvement.
    patience: number of epochs that produced the monitored
        quantity with no improvement after which training will
        be stopped.
        Validation quantities may not be produced for every
        epoch, if the validation frequency
        (`model.fit(validation_freq=5)`) is greater than one.
    verbose: verbosity mode.
    mode: one of {auto, min, max}. In `min` mode,
        training will stop when the quantity
        monitored has stopped decreasing; in `max`
        mode it will stop when the quantity
        monitored has stopped increasing; in `auto`
        mode, the direction is automatically inferred
        from the name of the monitored quantity.
    baseline: Baseline value for the monitored quantity to reach.
        Training will stop if the model doesn't show improvement
        over the baseline.
    restore_best_weights: whether to restore model weights from
        the epoch with the best value of the monitored quantity.
        If False, the model weights obtained at the last step of
        training are used.
File:           //anaconda3/envs/tf115/lib/python3.7/site-packages/keras/callbacks/callbacks.py
Type:           type

==================================================================================

Init signature:
Adam(
    learning_rate=0.001,
    beta_1=0.9,
    beta_2=0.999,
    amsgrad=False,
    **kwargs,
)
Docstring:     
Adam optimizer.

Default parameters follow those provided in the original paper.

# Arguments
    learning_rate: float >= 0. Learning rate.
    beta_1: float, 0 < beta < 1. Generally close to 1.
    beta_2: float, 0 < beta < 1. Generally close to 1.
    amsgrad: boolean. Whether to apply the AMSGrad variant of this
        algorithm from the paper "On the Convergence of Adam and
        Beyond".

# References
    - [Adam - A Method for Stochastic Optimization](
       https://arxiv.org/abs/1412.6980v8)
    - [On the Convergence of Adam and Beyond](
       https://openreview.net/forum?id=ryQu7f-RZ)
File:           /home/software/anaconda3/envs/tf115/lib/python3.7/site-packages/keras/optimizers.py
Type:           type
Subclasses:     

 

==================================================================================

Signature: compile(source, filename, mode, flags=0, dont_inherit=False, optimize=-1)
Docstring:
Compile source into a code object that can be executed by exec() or eval().

The source code may represent a Python module, statement or expression.
The filename will be used for run-time error messages.
The mode must be 'exec' to compile a module, 'single' to compile a
single (interactive) statement, or 'eval' to compile an expression.
The flags argument, if present, controls which future statements influence
the compilation of the code.
The dont_inherit argument, if true, stops the compilation inheriting
the effects of any future statements in effect in the code calling
compile; if absent or false these statements do influence the compilation,
in addition to any features explicitly specified.
Type:      builtin_function_or_method

 

==================================================================================

==================================================================================

 

==================================================================================

# construct the model
model = create_model(width=int(window_size/10))
es = EarlyStopping(monitor='val_loss', patience=2, restore_best_weights=True)
adam = Adam(lr=5e-5, beta_1=0.9, beta_2=0.999, epsilon=1e-08, decay=9e-5)
model.compile(loss='binary_crossentropy', optimizer=adam,
    metrics=['accuracy', auroc, auprc, f1_m, recall_m, precision_m])

if os.path.exists('./saved_models/DNase_hg38.v8.h5'):
    model.load_weights('./saved_models/DNase_hg38.v8.h5')
else:
    #train the model
    history = model.fit(x_train, y_train,
                        batch_size=32,
                        epochs=100,
                        validation_split=0.1,
                        shuffle=True,
                        class_weight=class_weights,
                        callbacks=[es])

    model.save_weights('./saved_models/DNase_hg38.v8.h5')
   

 

==================================================================================

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

==================================================================================