Я пытаюсь отобразить выход слоя кератов GlobalMaxPooling1D обратно в положение входной последовательности, но кажется, что ни одна из сверток на входе выборки не дает такой же результат, как полученный из вывода GlobalMaxPooling1D. Ожидаемый результат и полученный результат показаны ниже.
Основная цель здесь - отобразить важные позиции во входных данных и классифицировать их как положительные способствующие позиции. Аналогичная проблема обсуждается в этой статье, в этой статье используется максимальное объединение с течением времени (https://stats.stackexchange.com/questions/257321/what-is-global-max-pooling-layer-and-what-is-its-advantage-over-maxpooling-layer).
Наша сеть имеет 3 слоя Conv, GlobalMaxPooling1D (GMP) и Dense. После запуска модели я получил веса 20 фильтров, используемых в слое Conv. И выход для слоя GMP получен для тестового образца. Я попытался воссоздать выходные данные слоя GMP, свернув веса фильтра с входными данными, чтобы увидеть, совпадает ли он с выходными данными.
# Packages
import keras
import tensorflow as tf
from keras import backend as k
session_conf = tf.ConfigProto(intra_op_parallelism_threads=1,
inter_op_parallelism_threads=1)
sess = tf.Session(graph = tf.get_default_graph(), config = session_conf)
k.set_session(sess)
import numpy as np
np.random.seed(37)
import pandas as pd
tf.set_random_seed(89)
import random as rn
rn.seed(1254)
import matplotlib.pyplot as plt
from keras.layers import *
from keras.layers import Activation
from keras.layers.core import Dense, Flatten
from keras.optimizers import Adam
from keras.metrics import categorical_crossentropy
from keras.layers.normalization import BatchNormalization
from keras.layers.convolutional import *
import itertools
import math
from keras.models import Sequential, Model
from keras.layers import Input, Flatten, Dense, Dropout, Convolution2D, Conv2D, MaxPooling2D, Lambda, GlobalMaxPooling2D, GlobalAveragePooling2D, BatchNormalization, Activation, AveragePooling2D, Concatenate
from keras.callbacks import EarlyStopping, ModelCheckpoint, ReduceLROnPlateau
from keras.utils import np_utils
from keras.callbacks import CSVLogger
#%matplotlib inline
keras.backend.set_image_data_format('channels_last')
from keras import initializers
# DATA
# 1000 training example data:: each example is 1803 dimensional long
# where dimensional element is encoded using 4 dimensional one hot encoding
# for demonstration lets's use randomly initiated integers
X = np.random.random_integers(0, 1, size=(1000, 1803, 4))
y = np.random.random_integers(0, 1, size=(1000, 1))
print("X.shape : ", X.shape) ## (2237, 95, 95, 1)
print("y.shape : ", y.shape) ## (2237, 1)
train_index = 950
test_index = 950
# HYPERPARAMETERS
ACTIVATION = 'relu'
KERNEL_INITIALIZER = initializers.glorot_uniform(seed=90)
BIAS_INITIALIZER='zeros'
INPUT_SHAPE=X.shape[1:]
OUTPUT_ACTIVATION = 'sigmoid'
OPTIMIZER = Adam
LEARNING_RATE = 0.0001
BETA_1 = 0.9
BETA_2 = 0.999
EPSILON = None
DECAY = 0.0
AMSGRAD = False
LOSS = 'binary_crossentropy'
METRICS = ['accuracy']
BATCH_SIZE = 64
EPOCHS = 20
SHUFFLE = True
VALIDATION_SPLIT = 0.05
FIRST_CONV_LYR = 'conv_lyr_1'
FIRST_CONV_LYR_FILTERS = 20
FIRST_DENSE_LYR = 'dense_lyr_1'
FIRST_DENSE_LYR_UNITS = 50
VALID_PADDING = 'valid'
SAME_PADDING = 'same'
USE_BIAS = True
VERBOSE = 2
DILATION_RATE=1
# MODEL
model_21 = Sequential()
model_21.add(Conv1D(filters=FIRST_CONV_LYR_FILTERS, kernel_size = (2), strides=(1), activation = ACTIVATION,
padding = VALID_PADDING, input_shape =INPUT_SHAPE , kernel_initializer=KERNEL_INITIALIZER,
dilation_rate=DILATION_RATE, bias_initializer=BIAS_INITIALIZER, name =FIRST_CONV_LYR ))
model_21.add(GlobalMaxPooling1D()) # input_shape=(None, 1802, 20)
#model_21.add(Flatten()) ::: GMP does not need flatten()
model_21.add(Dense(1, activation = OUTPUT_ACTIVATION))
model_21.compile(optimizer=OPTIMIZER(lr = LEARNING_RATE, beta_1=BETA_1, beta_2=BETA_2, epsilon=EPSILON, decay=DECAY,
amsgrad=AMSGRAD), loss = LOSS, metrics = METRICS,)
mdl_21 = model_21.fit(X[:train_index], y[:train_index], batch_size = BATCH_SIZE, epochs = EPOCHS, verbose = VERBOSE,
shuffle=SHUFFLE, validation_split=VALIDATION_SPLIT)
accuracy_21 = model_21.evaluate(X[test_index:], y[test_index:])
print(model_21.metrics_names[1], accuracy_21[1]*100)
print("The accuracy of Model 21 is : ", accuracy_21)
# PREDICTION
prediction_21 = model_21.predict(X[test_index:]) # predict() takes <class 'numpy.ndarray'> as input
predicted_classes_21 = model_21.predict_classes(X[test_index:])
print("The predicted classes are : ", predicted_classes_21[:, 0])
print("Actual classes are : ", y[test_index:][:, 0]) # [:, 0]
ВЫВОД:
The output of training above model is
Train on 902 samples, validate on 48 samples
Epoch 1/20
- 3s - loss: 0.7790 - acc: 0.4878 - val_loss: 0.7631 - val_acc: 0.5000
Epoch 2/20
- 3s - loss: 0.7682 - acc: 0.4878 - val_loss: 0.7537 - val_acc: 0.5000
Epoch 3/20
- 3s - loss: 0.7588 - acc: 0.4878 - val_loss: 0.7463 - val_acc: 0.5000
Epoch 4/20
- 3s - loss: 0.7503 - acc: 0.4878 - val_loss: 0.7384 - val_acc: 0.5000
Epoch 5/20
- 3s - loss: 0.7422 - acc: 0.4878 - val_loss: 0.7313 - val_acc: 0.5000
Epoch 6/20
- 3s - loss: 0.7348 - acc: 0.4878 - val_loss: 0.7249 - val_acc: 0.5000
Epoch 7/20
- 3s - loss: 0.7282 - acc: 0.4878 - val_loss: 0.7192 - val_acc: 0.5000
Epoch 8/20
- 3s - loss: 0.7227 - acc: 0.4878 - val_loss: 0.7148 - val_acc: 0.5000
Epoch 9/20
- 3s - loss: 0.7180 - acc: 0.4878 - val_loss: 0.7109 - val_acc: 0.5000
Epoch 10/20
- 3s - loss: 0.7137 - acc: 0.4878 - val_loss: 0.7075 - val_acc: 0.5000
Epoch 11/20
- 3s - loss: 0.7106 - acc: 0.4878 - val_loss: 0.7050 - val_acc: 0.5000
Epoch 12/20
- 3s - loss: 0.7074 - acc: 0.4878 - val_loss: 0.7026 - val_acc: 0.5000
Epoch 13/20
- 3s - loss: 0.7048 - acc: 0.4878 - val_loss: 0.7006 - val_acc: 0.5000
Epoch 14/20
- 3s - loss: 0.7026 - acc: 0.4878 - val_loss: 0.6988 - val_acc: 0.5000
Epoch 15/20
- 3s - loss: 0.7006 - acc: 0.4878 - val_loss: 0.6975 - val_acc: 0.5000
Epoch 16/20
- 3s - loss: 0.6992 - acc: 0.4878 - val_loss: 0.6966 - val_acc: 0.5000
Epoch 17/20
- 3s - loss: 0.6981 - acc: 0.4878 - val_loss: 0.6958 - val_acc: 0.5000
Epoch 18/20
- 3s - loss: 0.6971 - acc: 0.4878 - val_loss: 0.6950 - val_acc: 0.5000
Epoch 19/20
- 3s - loss: 0.6962 - acc: 0.4878 - val_loss: 0.6945 - val_acc: 0.5000
Epoch 20/20
- 3s - loss: 0.6953 - acc: 0.4878 - val_loss: 0.6940 - val_acc: 0.5000
50/50 [==============================] - 0s 439us/step
acc 48.00000047683716
The accuracy of Model 21 is : [0.6956488680839539, 0.48000000476837157]
The predicted values for test dataset is::
The predicted classes are : [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
Actual classes are : [1 0 1 0 1 0 1 0 0 0 1 0 1 1 1 1 0 1 1 1 1 0 1 1 0 1 0 1 1 0 0 1 0 0 0 1 1 0 0 0 0 0 0 0 1 1 0 1 0 0]
Вывод слоя GMP на тестовом образце
from keras.models import Model
model = model_21 # include here your original model
layer_name = 'global_max_pooling1d_1'
intermediate_layer_model = Model(inputs=model.input, outputs=model.get_layer(layer_name).output)
intermediate_output = intermediate_layer_model.predict(X[test_index:test_index+1])
print("intermediate_output shape : \n", intermediate_output.shape)
print("intermediate_output : \n",intermediate_output[0])
ВЫВОД:
intermediate_output shape :
(1, 20)
intermediate_output :
[[0.83934546 0.4251969 0.6270695 1.3905973 0.5667287 0.76842016
0.8666611 0.84354174 0.5817993 0.82427627 0.4142136 0.79649013
0.61913747 1.2524168 1.1239575 0.5584184 0.35370556 1.0718826
0.96888304 0.51134074]]
ВЕСА И СКОРОСТИ извилистого и плотного слоя
# WEIGHTS & BIASES of conv and dense layer
first_layer_weights = model.layers[0].get_weights()[0]
first_layer_biases = model.layers[0].get_weights()[1]
third_layer_weights = model.layers[2].get_weights()[0]
third_layer_biases = model.layers[2].get_weights()[1]
ВОПРОС ЗДЕСЬ
# I tried to recreate the output for all 20 filters used in conv layer
# to map them with output of GMP on test data
# conv_val holds the convoluted value of 20 filters
conv_val = [[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[],[]]
for i in range(len(intermediate_output[0])):
print("Finding the postion of input which resulted into position value : {:d} of GAP ouput".format(i+1))
for j in range(len(X[test_index])-1):
# each sample data is 1803 vocabulary long and each vocabulary is encoded using a 4 dimensional vector representation
# we are using kernel size of 2 in conv1d layer so each filter is 2*4
conv_result = np.sum((X[test_index][j:j+2] * first_layer_weights[:,:,i]) + first_layer_biases[i])
conv_val[i].append(max(0, conv_result))
posn = np.argmax(conv_val[i])
print("position of max element obtained from convolution with text input: posn : ", posn)
print("Actual o/p: vale in resulting convlution on test data ",conv_val[i][posn])
#print("np.max : ", np.max(conv_val[i]))
print("Expected op : The max value obtained from GAP layer on test data ", intermediate_output[0][i])
if ((conv_val[i][posn]) == intermediate_output[0][i]):
print("Convolved at the index : ", j)
print("\n")
else:
print("not found :( ")
print("\n")
OUTPUT
======= Find postion in input which resulted into GMP output position value 1 ======
position of max element obtained from convolution with text input: posn : 612
Actual o/p: vale in resulting convlution on test data 0.9715582579374313
Expected op : The max value obtained from GAP layer on test data 0.83934546
not found :(
======= Find postion in input which resulted into GMP output position value 2 ======
position of max element obtained from convolution with text input: posn : 63
Actual o/p: vale in resulting convlution on test data 0.29808690398931503
Expected op : The max value obtained from GAP layer on test data 0.4251969
not found :(
======= Find postion in input which resulted into GMP output position value 3 ======
position of max element obtained from convolution with text input: posn : 412
Actual o/p: vale in resulting convlution on test data 0.7571525424718857
Expected op : The max value obtained from GAP layer on test data 0.6270695
not found :(
======= Find postion in input which resulted into GMP output position value 4 ======
position of max element obtained from convolution with text input: posn : 292
Actual o/p: vale in resulting convlution on test data 1.5215912163257599
Expected op : The max value obtained from GAP layer on test data 1.3905973
not found :(
======= Find postion in input which resulted into GMP output position value 5 ======
position of max element obtained from convolution with text input: posn : 40
Actual o/p: vale in resulting convlution on test data 0.6967360526323318
Expected op : The max value obtained from GAP layer on test data 0.5667287
not found :(
======= Find postion in input which resulted into GMP output position value 6 ======
position of max element obtained from convolution with text input: posn : 371
Actual o/p: vale in resulting convlution on test data 0.6413831561803818
Expected op : The max value obtained from GAP layer on test data 0.76842016
not found :(
======= Find postion in input which resulted into GMP output position value 7 ======
position of max element obtained from convolution with text input: posn : 166
Actual o/p: vale in resulting convlution on test data 0.9993664026260376
Expected op : The max value obtained from GAP layer on test data 0.8666611
not found :(
======= Find postion in input which resulted into GMP output position value 8 ======
position of max element obtained from convolution with text input: posn : 149
Actual o/p: vale in resulting convlution on test data 0.7193388789892197
Expected op : The max value obtained from GAP layer on test data 0.84354174
not found :(
======= Find postion in input which resulted into GMP output position value 9 ======
position of max element obtained from convolution with text input: posn : 93
Actual o/p: vale in resulting convlution on test data 0.7193349152803421
Expected op : The max value obtained from GAP layer on test data 0.5817993
not found :(
======= Find postion in input which resulted into GMP output position value 10 ======
position of max element obtained from convolution with text input: posn : 30
Actual o/p: vale in resulting convlution on test data 0.6970454901456833
Expected op : The max value obtained from GAP layer on test data 0.82427627
not found :(
======= Find postion in input which resulted into GMP output position value 11 ======
position of max element obtained from convolution with text input: posn : 27
Actual o/p: vale in resulting convlution on test data 0.2878176420927048
Expected op : The max value obtained from GAP layer on test data 0.4142136
not found :(
======= Find postion in input which resulted into GMP output position value 12 ======
position of max element obtained from convolution with text input: posn : 96
Actual o/p: vale in resulting convlution on test data 0.9302686732262373
Expected op : The max value obtained from GAP layer on test data 0.79649013
not found :(
======= Find postion in input which resulted into GMP output position value 13 ======
position of max element obtained from convolution with text input: posn : 2
Actual o/p: vale in resulting convlution on test data 0.7492209821939468
Expected op : The max value obtained from GAP layer on test data 0.61913747
not found :(
======= Find postion in input which resulted into GMP output position value 14 ======
position of max element obtained from convolution with text input: posn : 82
Actual o/p: vale in resulting convlution on test data 1.3909660689532757
Expected op : The max value obtained from GAP layer on test data 1.2524168
not found :(
======= Find postion in input which resulted into GMP output position value 15 ======
position of max element obtained from convolution with text input: posn : 82
Actual o/p: vale in resulting convlution on test data 1.0048598730936646
Expected op : The max value obtained from GAP layer on test data 1.1239575
not found :(
======= Find postion in input which resulted into GMP output position value 16 ======
position of max element obtained from convolution with text input: posn : 986
Actual o/p: vale in resulting convlution on test data 0.43241868913173676
Expected op : The max value obtained from GAP layer on test data 0.5584184
not found :(
======= Find postion in input which resulted into GMP output position value 17 ======
position of max element obtained from convolution with text input: posn : 52
Actual o/p: vale in resulting convlution on test data 0.22693601995706558
Expected op : The max value obtained from GAP layer on test data 0.35370556
not found :(
======= Find postion in input which resulted into GMP output position value 18 ======
position of max element obtained from convolution with text input: posn : 51
Actual o/p: vale in resulting convlution on test data 0.9564715027809143
Expected op : The max value obtained from GAP layer on test data 1.0718826
not found :(
======= Find postion in input which resulted into GMP output position value 19 ======
position of max element obtained from convolution with text input: posn : 106
Actual o/p: vale in resulting convlution on test data 0.8425523824989796
Expected op : The max value obtained from GAP layer on test data 0.96888304
not found :(
======= Find postion in input which resulted into GMP output position value 20 ======
position of max element obtained from convolution with text input: posn : 410
Actual o/p: vale in resulting convlution on test data 0.641106590628624
Expected op : The max value obtained from GAP layer on test data 0.51134074
not found :(
Фактические и ожидаемые значения не совпадают друг с другом. Правильно ли я делаю операцию для нахождения значения GMP, которое принимает максимальное значение полученного выхода ( Макс. За время объединения в Keras ).
После исправления этого я буду отображать выходные данные классификации, полученные в последнем плотном слое, чтобы связать позиции, полученные из входной последовательности для всех последовательностей. Это поможет мне найти позиции последовательности, которые положительно связаны с выходом. Пожалуйста, дайте мне знать, если я смогу использовать какой-то существующий код для поиска таких позиций или генерировать тепловую карту в таких задачах. Извините за длинный пост и заранее спасибо !!