Lernsteuerung
Lernziele
Nach Abschluss dieses Kapitels …
ein einfaches neuronales Netzwerk mit Keras erstellen zur Klassifikation von Hate-Speech.
Überblick
In diesem Kapitel nutzen wir grundlegende Methoden neuronaler Netze, um Hate-Speech vorherzusagen. Dabei findet der Datensatz GermEval Verwendung. Zunächst verwenden wir den schon aufbereiteten Datensatz, das macht es uns einfacher. Dieser aufbereitete Datensatz ist schon “numerisiert”. Der Text der Tweets ist schon in numerische Prädiktoren umgewandelt. Dabei fanden einfache (deutschsprachige) Wordvektoren (wikipedia2vec) Verwendung. In diesem Kapitel arbeiten wir mit ausschließlich mit Python.
Python-Check
reticulate :: py_available ( ) ## [1] FALSE reticulate :: py_config ( ) ## python: /Users/sebastiansaueruser/.virtualenvs/r-tensorflow/bin/python ## libpython: /Users/sebastiansaueruser/.pyenv/versions/3.8.16/lib/libpython3.8.dylib ## pythonhome: /Users/sebastiansaueruser/.virtualenvs/r-tensorflow:/Users/sebastiansaueruser/.virtualenvs/r-tensorflow ## version: 3.8.16 (default, Sep 15 2023, 17:53:02) [Clang 14.0.3 (clang-1403.0.22.14.1)] ## numpy: /Users/sebastiansaueruser/.virtualenvs/r-tensorflow/lib/python3.8/site-packages/numpy ## numpy_version: 1.24.3 ## ## NOTE: Python version was forced by VIRTUAL_ENV
Benötigte Python-Module
import kerasimport pandas as pdimport tensorflow as tffrom tensorflow.keras.models import Sequentialfrom tensorflow.keras.layers import Densefrom sklearn.metrics import accuracy_score
Pipeline mit 1 Hidden Layer
Daten
= pd.read_csv("https://raw.githubusercontent.com/sebastiansauer/Datenwerk2/main/data/germeval/germeval_train_recipe_wordvec_senti.csv" )= d_train_baked.select_dtypes(include= 'number' )= d_train_baked.loc[:, "emo_count" :"wordembed_text_V101" ]= d_train2.values"y" ] = d_train_baked["c1" ].map ({"OTHER" : 0 , "OFFENSE" : 1 })= d_train_baked.loc[:, "y" ].values
Head von y_train:
print (y_train[:6 ])## [0 0 0 0 1 0]
Info zum Objekt:
## <class 'pandas.core.frame.DataFrame'> ## RangeIndex: 5009 entries, 0 to 5008 ## Columns: 119 entries, emo_count to wordembed_text_V101 ## dtypes: float64(119) ## memory usage: 4.5 MB
Head von y_train2:
print (d_train2.head())## emo_count schimpf_count ... wordembed_text_V100 wordembed_text_V101 ## 0 0.574594 -0.450067 ... -0.449265 -0.277801 ## 1 -1.111107 -0.450067 ... 0.974438 0.223422 ## 2 0.186402 -0.450067 ... 0.407285 0.470835 ## 3 0.201551 -0.450067 ... -0.681155 0.351565 ## 4 0.168223 -0.450067 ... -0.674108 0.543312 ## ## [5 rows x 119 columns]
= pd.read_csv("https://raw.githubusercontent.com/sebastiansauer/Datenwerk2/main/data/germeval/germeval_test_recipe_wordvec_senti.csv" )= d_test_baked.select_dtypes(include= 'number' )= d_test_baked.loc[:, "emo_count" :"wordembed_text_V101" ]= d_test2.values"y" ] = d_test_baked["c1" ].map ({"OTHER" : 0 , "OFFENSE" : 1 })= d_test_baked.loc[:, "y" ].values
print (y_test[:5 ])## [0 0 0 0 1]
Modeldefinition
= Sequential()64 , input_dim= X_train.shape[1 ], activation= 'relu' ))1 , activation= 'sigmoid' ))
compile (optimizer= 'adam' , loss= 'binary_crossentropy' , metrics= ['accuracy' ])
Fit
= 10 , batch_size= 64 , validation_data= (X_test, y_test))## Epoch 1/10 ## 1 / 79 [..............................] - ETA: 57 s - loss: 0.9031 - accuracy: 0.4688 32 / 79 [===========> ..................] - ETA: 0 s - loss: 0.7415 - accuracy: 0.5435 66 / 79 [========================> .....] - ETA: 0 s - loss: 0.6537 - accuracy: 0.6264 79 / 79 [============================== ] - 1 s 5 ms / step - loss: 0.6366 - accuracy: 0.6396 - val_loss: 0.5578 - val_accuracy: 0.7166 ## Epoch 2/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.5155 - accuracy: 0.7969 36 / 79 [============> .................] - ETA: 0 s - loss: 0.5192 - accuracy: 0.7374 70 / 79 [=========================> ....] - ETA: 0 s - loss: 0.5059 - accuracy: 0.7473 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.5030 - accuracy: 0.7479 - val_loss: 0.5491 - val_accuracy: 0.7262 ## Epoch 3/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.5058 - accuracy: 0.7344 36 / 79 [============> .................] - ETA: 0 s - loss: 0.4703 - accuracy: 0.7726 71 / 79 [=========================> ....] - ETA: 0 s - loss: 0.4771 - accuracy: 0.7658 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.4733 - accuracy: 0.7678 - val_loss: 0.5540 - val_accuracy: 0.7273 ## Epoch 4/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.4064 - accuracy: 0.7969 36 / 79 [============> .................] - ETA: 0 s - loss: 0.4464 - accuracy: 0.7860 73 / 79 [==========================> ...] - ETA: 0 s - loss: 0.4523 - accuracy: 0.7815 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.4539 - accuracy: 0.7820 - val_loss: 0.5623 - val_accuracy: 0.7240 ## Epoch 5/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.3533 - accuracy: 0.8438 35 / 79 [============> .................] - ETA: 0 s - loss: 0.4413 - accuracy: 0.7915 71 / 79 [=========================> ....] - ETA: 0 s - loss: 0.4400 - accuracy: 0.7931 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.4380 - accuracy: 0.7944 - val_loss: 0.5607 - val_accuracy: 0.7248 ## Epoch 6/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.3845 - accuracy: 0.8281 34 / 79 [===========> ..................] - ETA: 0 s - loss: 0.4141 - accuracy: 0.8051 70 / 79 [=========================> ....] - ETA: 0 s - loss: 0.4201 - accuracy: 0.7998 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.4233 - accuracy: 0.7978 - val_loss: 0.5656 - val_accuracy: 0.7251 ## Epoch 7/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.4457 - accuracy: 0.7812 24 / 79 [========> .....................] - ETA: 0 s - loss: 0.4157 - accuracy: 0.8027 57 / 79 [====================> .........] - ETA: 0 s - loss: 0.4115 - accuracy: 0.8062 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.4098 - accuracy: 0.8065 - val_loss: 0.5689 - val_accuracy: 0.7282 ## Epoch 8/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.3460 - accuracy: 0.8906 32 / 79 [===========> ..................] - ETA: 0 s - loss: 0.3959 - accuracy: 0.8184 67 / 79 [========================> .....] - ETA: 0 s - loss: 0.3945 - accuracy: 0.8190 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.3971 - accuracy: 0.8169 - val_loss: 0.5731 - val_accuracy: 0.7268 ## Epoch 9/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.4085 - accuracy: 0.7969 21 / 79 [======> .......................] - ETA: 0 s - loss: 0.3617 - accuracy: 0.8549 49 / 79 [=================> ............] - ETA: 0 s - loss: 0.3765 - accuracy: 0.8339 76 / 79 [===========================> ..] - ETA: 0 s - loss: 0.3832 - accuracy: 0.8283 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.3849 - accuracy: 0.8271 - val_loss: 0.5836 - val_accuracy: 0.7268 ## Epoch 10/10 ## 1 / 79 [..............................] - ETA: 0 s - loss: 0.2995 - accuracy: 0.8906 26 / 79 [========> .....................] - ETA: 0 s - loss: 0.3683 - accuracy: 0.8335 52 / 79 [==================> ...........] - ETA: 0 s - loss: 0.3717 - accuracy: 0.8287 78 / 79 [============================> .] - ETA: 0 s - loss: 0.3725 - accuracy: 0.8305 79 / 79 [============================== ] - 0 s 3 ms / step - loss: 0.3726 - accuracy: 0.8305 - val_loss: 0.5840 - val_accuracy: 0.7265 ## <keras.src.callbacks.History object at 0x137a43ac0>
Fazit
Schon mit diesem einfachen Netzwerk, das sich schnell berechnen lässt, übertreffen wir auf Anhieb die Modellgüte (Gesamtgenauigkeit) der Shallow-Learners aus früheren Kapiteln.
Pipeline mit 2 Hidden Layers
Wir verwenden die gleichen Daten wie oben.
Wir fügen eine zweite Hidden Layer hinzu. Außerdem verändern wir die Batch-Size.
Modeldefinition
= Sequential()64 , input_dim= X_train.shape[1 ], activation= 'relu' ))= 32 , activation= 'relu' )) # Second hidden layer 1 , activation= 'sigmoid' ))
compile (optimizer= 'adam' , loss= 'binary_crossentropy' , metrics= ['accuracy' ])
Fit
= 10 , batch_size= 8 , validation_data= (X_test, y_test))## Epoch 1/10 ## 1 / 627 [..............................] - ETA: 8 :54 - loss: 0.6587 - accuracy: 0.5000 28 / 627 [> .............................] - ETA: 1 s - loss: 0.6542 - accuracy: 0.6384 56 / 627 [=> ............................] - ETA: 1 s - loss: 0.6259 - accuracy: 0.6562 86 / 627 [===> ..........................] - ETA: 0 s - loss: 0.6204 - accuracy: 0.6570 119 / 627 [====> .........................] - ETA: 0 s - loss: 0.6080 - accuracy: 0.6691 153 / 627 [======> .......................] - ETA: 0 s - loss: 0.6012 - accuracy: 0.6748 189 / 627 [========> .....................] - ETA: 0 s - loss: 0.5924 - accuracy: 0.6779 224 / 627 [=========> ....................] - ETA: 0 s - loss: 0.5813 - accuracy: 0.6864 255 / 627 [===========> ..................] - ETA: 0 s - loss: 0.5744 - accuracy: 0.6926 291 / 627 [============> .................] - ETA: 0 s - loss: 0.5744 - accuracy: 0.6942 327 / 627 [==============> ...............] - ETA: 0 s - loss: 0.5726 - accuracy: 0.6957 332 / 627 [==============> ...............] - ETA: 0 s - loss: 0.5713 - accuracy: 0.6969 366 / 627 [================> .............] - ETA: 0 s - loss: 0.5658 - accuracy: 0.7018 401 / 627 [==================> ...........] - ETA: 0 s - loss: 0.5634 - accuracy: 0.7042 438 / 627 [===================> ..........] - ETA: 0 s - loss: 0.5616 - accuracy: 0.7072 473 / 627 [=====================> ........] - ETA: 0 s - loss: 0.5554 - accuracy: 0.7106 509 / 627 [=======================> ......] - ETA: 0 s - loss: 0.5552 - accuracy: 0.7110 545 / 627 [=========================> ....] - ETA: 0 s - loss: 0.5522 - accuracy: 0.7149 582 / 627 [==========================> ...] - ETA: 0 s - loss: 0.5497 - accuracy: 0.7174 617 / 627 [============================> .] - ETA: 0 s - loss: 0.5499 - accuracy: 0.7180 627 / 627 [============================== ] - 3 s 3 ms / step - loss: 0.5477 - accuracy: 0.7181 - val_loss: 0.5593 - val_accuracy: 0.7214 ## Epoch 2/10 ## 1 / 627 [..............................] - ETA: 0 s - loss: 0.1919 - accuracy: 1.0000 36 / 627 [> .............................] - ETA: 0 s - loss: 0.4186 - accuracy: 0.8194 71 / 627 [==> ...........................] - ETA: 0 s - loss: 0.4415 - accuracy: 0.8046 105 / 627 [====> .........................] - ETA: 0 s - loss: 0.4356 - accuracy: 0.8036 141 / 627 [=====> ........................] - ETA: 0 s - loss: 0.4379 - accuracy: 0.7979 175 / 627 [=======> ......................] - ETA: 0 s - loss: 0.4489 - accuracy: 0.7893 211 / 627 [=========> ....................] - ETA: 0 s - loss: 0.4469 - accuracy: 0.7885 246 / 627 [==========> ...................] - ETA: 0 s - loss: 0.4488 - accuracy: 0.7881 282 / 627 [============> .................] - ETA: 0 s - loss: 0.4509 - accuracy: 0.7877 317 / 627 [==============> ...............] - ETA: 0 s - loss: 0.4441 - accuracy: 0.7898 354 / 627 [===============> ..............] - ETA: 0 s - loss: 0.4508 - accuracy: 0.7871 389 / 627 [=================> ............] - ETA: 0 s - loss: 0.4538 - accuracy: 0.7831 425 / 627 [===================> ..........] - ETA: 0 s - loss: 0.4540 - accuracy: 0.7815 460 / 627 [=====================> ........] - ETA: 0 s - loss: 0.4549 - accuracy: 0.7793 491 / 627 [======================> .......] - ETA: 0 s - loss: 0.4589 - accuracy: 0.7777 525 / 627 [========================> .....] - ETA: 0 s - loss: 0.4598 - accuracy: 0.7769 561 / 627 [=========================> ....] - ETA: 0 s - loss: 0.4598 - accuracy: 0.7790 593 / 627 [===========================> ..] - ETA: 0 s - loss: 0.4589 - accuracy: 0.7789 623 / 627 [============================> .] - ETA: 0 s - loss: 0.4577 - accuracy: 0.7791 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.4568 - accuracy: 0.7796 - val_loss: 0.5706 - val_accuracy: 0.7285 ## Epoch 3/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.4661 - accuracy: 0.5000 36 / 627 [> .............................] - ETA: 0 s - loss: 0.4126 - accuracy: 0.7847 70 / 627 [==> ...........................] - ETA: 0 s - loss: 0.4173 - accuracy: 0.7929 100 / 627 [===> ..........................] - ETA: 0 s - loss: 0.4206 - accuracy: 0.8025 132 / 627 [=====> ........................] - ETA: 0 s - loss: 0.4117 - accuracy: 0.8078 167 / 627 [======> .......................] - ETA: 0 s - loss: 0.4155 - accuracy: 0.8046 202 / 627 [========> .....................] - ETA: 0 s - loss: 0.4129 - accuracy: 0.8069 233 / 627 [==========> ...................] - ETA: 0 s - loss: 0.4146 - accuracy: 0.8063 267 / 627 [===========> ..................] - ETA: 0 s - loss: 0.4173 - accuracy: 0.8034 298 / 627 [=============> ................] - ETA: 0 s - loss: 0.4177 - accuracy: 0.8054 326 / 627 [==============> ...............] - ETA: 0 s - loss: 0.4154 - accuracy: 0.8067 357 / 627 [================> .............] - ETA: 0 s - loss: 0.4143 - accuracy: 0.8074 380 / 627 [=================> ............] - ETA: 0 s - loss: 0.4139 - accuracy: 0.8076 407 / 627 [==================> ...........] - ETA: 0 s - loss: 0.4154 - accuracy: 0.8050 428 / 627 [===================> ..........] - ETA: 0 s - loss: 0.4137 - accuracy: 0.8055 458 / 627 [====================> .........] - ETA: 0 s - loss: 0.4140 - accuracy: 0.8051 489 / 627 [======================> .......] - ETA: 0 s - loss: 0.4164 - accuracy: 0.8037 517 / 627 [=======================> ......] - ETA: 0 s - loss: 0.4159 - accuracy: 0.8044 548 / 627 [=========================> ....] - ETA: 0 s - loss: 0.4174 - accuracy: 0.8043 574 / 627 [==========================> ...] - ETA: 0 s - loss: 0.4170 - accuracy: 0.8051 600 / 627 [===========================> ..] - ETA: 0 s - loss: 0.4162 - accuracy: 0.8067 626 / 627 [============================> .] - ETA: 0 s - loss: 0.4140 - accuracy: 0.8075 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.4140 - accuracy: 0.8075 - val_loss: 0.6027 - val_accuracy: 0.7248 ## Epoch 4/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.3053 - accuracy: 0.7500 32 / 627 [> .............................] - ETA: 0 s - loss: 0.3231 - accuracy: 0.8672 62 / 627 [=> ............................] - ETA: 0 s - loss: 0.3515 - accuracy: 0.8407 93 / 627 [===> ..........................] - ETA: 0 s - loss: 0.3484 - accuracy: 0.8562 120 / 627 [====> .........................] - ETA: 0 s - loss: 0.3372 - accuracy: 0.8583 153 / 627 [======> .......................] - ETA: 0 s - loss: 0.3312 - accuracy: 0.8603 183 / 627 [=======> ......................] - ETA: 0 s - loss: 0.3439 - accuracy: 0.8518 212 / 627 [=========> ....................] - ETA: 0 s - loss: 0.3443 - accuracy: 0.8532 226 / 627 [=========> ....................] - ETA: 0 s - loss: 0.3461 - accuracy: 0.8507 245 / 627 [==========> ...................] - ETA: 0 s - loss: 0.3511 - accuracy: 0.8495 269 / 627 [===========> ..................] - ETA: 0 s - loss: 0.3523 - accuracy: 0.8494 288 / 627 [============> .................] - ETA: 0 s - loss: 0.3567 - accuracy: 0.8455 317 / 627 [==============> ...............] - ETA: 0 s - loss: 0.3589 - accuracy: 0.8442 346 / 627 [===============> ..............] - ETA: 0 s - loss: 0.3659 - accuracy: 0.8396 376 / 627 [================> .............] - ETA: 0 s - loss: 0.3672 - accuracy: 0.8371 397 / 627 [=================> ............] - ETA: 0 s - loss: 0.3673 - accuracy: 0.8369 425 / 627 [===================> ..........] - ETA: 0 s - loss: 0.3697 - accuracy: 0.8350 447 / 627 [====================> .........] - ETA: 0 s - loss: 0.3711 - accuracy: 0.8353 467 / 627 [=====================> ........] - ETA: 0 s - loss: 0.3726 - accuracy: 0.8362 492 / 627 [======================> .......] - ETA: 0 s - loss: 0.3726 - accuracy: 0.8364 511 / 627 [=======================> ......] - ETA: 0 s - loss: 0.3726 - accuracy: 0.8371 537 / 627 [========================> .....] - ETA: 0 s - loss: 0.3756 - accuracy: 0.8350 568 / 627 [==========================> ...] - ETA: 0 s - loss: 0.3755 - accuracy: 0.8343 595 / 627 [===========================> ..] - ETA: 0 s - loss: 0.3756 - accuracy: 0.8334 622 / 627 [============================> .] - ETA: 0 s - loss: 0.3751 - accuracy: 0.8336 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.3753 - accuracy: 0.8333 - val_loss: 0.6474 - val_accuracy: 0.7101 ## Epoch 5/10 ## 1 / 627 [..............................] - ETA: 0 s - loss: 0.4109 - accuracy: 0.7500 37 / 627 [> .............................] - ETA: 0 s - loss: 0.3228 - accuracy: 0.8818 70 / 627 [==> ...........................] - ETA: 0 s - loss: 0.3331 - accuracy: 0.8643 106 / 627 [====> .........................] - ETA: 0 s - loss: 0.3080 - accuracy: 0.8774 144 / 627 [=====> ........................] - ETA: 0 s - loss: 0.3092 - accuracy: 0.8724 180 / 627 [=======> ......................] - ETA: 0 s - loss: 0.3196 - accuracy: 0.8611 216 / 627 [=========> ....................] - ETA: 0 s - loss: 0.3272 - accuracy: 0.8559 247 / 627 [==========> ...................] - ETA: 0 s - loss: 0.3281 - accuracy: 0.8548 274 / 627 [============> .................] - ETA: 0 s - loss: 0.3275 - accuracy: 0.8540 295 / 627 [=============> ................] - ETA: 0 s - loss: 0.3305 - accuracy: 0.8504 314 / 627 [==============> ...............] - ETA: 0 s - loss: 0.3297 - accuracy: 0.8527 335 / 627 [===============> ..............] - ETA: 0 s - loss: 0.3285 - accuracy: 0.8537 365 / 627 [================> .............] - ETA: 0 s - loss: 0.3286 - accuracy: 0.8565 393 / 627 [=================> ............] - ETA: 0 s - loss: 0.3289 - accuracy: 0.8569 423 / 627 [===================> ..........] - ETA: 0 s - loss: 0.3304 - accuracy: 0.8561 453 / 627 [====================> .........] - ETA: 0 s - loss: 0.3323 - accuracy: 0.8551 482 / 627 [======================> .......] - ETA: 0 s - loss: 0.3368 - accuracy: 0.8527 512 / 627 [=======================> ......] - ETA: 0 s - loss: 0.3374 - accuracy: 0.8525 547 / 627 [=========================> ....] - ETA: 0 s - loss: 0.3359 - accuracy: 0.8544 581 / 627 [==========================> ...] - ETA: 0 s - loss: 0.3340 - accuracy: 0.8552 617 / 627 [============================> .] - ETA: 0 s - loss: 0.3361 - accuracy: 0.8539 627 / 627 [============================== ] - 2 s 2 ms / step - loss: 0.3359 - accuracy: 0.8537 - val_loss: 0.6784 - val_accuracy: 0.7129 ## Epoch 6/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.1333 - accuracy: 1.0000 37 / 627 [> .............................] - ETA: 0 s - loss: 0.2734 - accuracy: 0.8784 74 / 627 [==> ...........................] - ETA: 0 s - loss: 0.2664 - accuracy: 0.8953 108 / 627 [====> .........................] - ETA: 0 s - loss: 0.2676 - accuracy: 0.8831 145 / 627 [=====> ........................] - ETA: 0 s - loss: 0.2765 - accuracy: 0.8767 182 / 627 [=======> ......................] - ETA: 0 s - loss: 0.2808 - accuracy: 0.8805 213 / 627 [=========> ....................] - ETA: 0 s - loss: 0.2856 - accuracy: 0.8762 248 / 627 [==========> ...................] - ETA: 0 s - loss: 0.2890 - accuracy: 0.8755 279 / 627 [============> .................] - ETA: 0 s - loss: 0.2914 - accuracy: 0.8768 309 / 627 [=============> ................] - ETA: 0 s - loss: 0.2889 - accuracy: 0.8786 329 / 627 [==============> ...............] - ETA: 0 s - loss: 0.2903 - accuracy: 0.8784 347 / 627 [===============> ..............] - ETA: 0 s - loss: 0.2896 - accuracy: 0.8779 367 / 627 [================> .............] - ETA: 0 s - loss: 0.2911 - accuracy: 0.8764 392 / 627 [=================> ............] - ETA: 0 s - loss: 0.2906 - accuracy: 0.8753 418 / 627 [===================> ..........] - ETA: 0 s - loss: 0.2951 - accuracy: 0.8747 447 / 627 [====================> .........] - ETA: 0 s - loss: 0.2958 - accuracy: 0.8753 470 / 627 [=====================> ........] - ETA: 0 s - loss: 0.2948 - accuracy: 0.8771 501 / 627 [======================> .......] - ETA: 0 s - loss: 0.2992 - accuracy: 0.8743 533 / 627 [========================> .....] - ETA: 0 s - loss: 0.2989 - accuracy: 0.8748 570 / 627 [==========================> ...] - ETA: 0 s - loss: 0.2995 - accuracy: 0.8750 606 / 627 [===========================> ..] - ETA: 0 s - loss: 0.2992 - accuracy: 0.8744 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.2980 - accuracy: 0.8746 - val_loss: 0.6883 - val_accuracy: 0.7189 ## Epoch 7/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.2289 - accuracy: 1.0000 30 / 627 [> .............................] - ETA: 1 s - loss: 0.1825 - accuracy: 0.9583 57 / 627 [=> ............................] - ETA: 1 s - loss: 0.1969 - accuracy: 0.9364 76 / 627 [==> ...........................] - ETA: 1 s - loss: 0.1903 - accuracy: 0.9391 101 / 627 [===> ..........................] - ETA: 1 s - loss: 0.2084 - accuracy: 0.9295 133 / 627 [=====> ........................] - ETA: 0 s - loss: 0.2183 - accuracy: 0.9295 158 / 627 [======> .......................] - ETA: 0 s - loss: 0.2289 - accuracy: 0.9217 186 / 627 [=======> ......................] - ETA: 0 s - loss: 0.2223 - accuracy: 0.9267 213 / 627 [=========> ....................] - ETA: 0 s - loss: 0.2283 - accuracy: 0.9255 239 / 627 [==========> ...................] - ETA: 0 s - loss: 0.2371 - accuracy: 0.9179 270 / 627 [===========> ..................] - ETA: 0 s - loss: 0.2388 - accuracy: 0.9171 301 / 627 [=============> ................] - ETA: 0 s - loss: 0.2454 - accuracy: 0.9128 333 / 627 [==============> ...............] - ETA: 0 s - loss: 0.2507 - accuracy: 0.9088 365 / 627 [================> .............] - ETA: 0 s - loss: 0.2482 - accuracy: 0.9106 396 / 627 [=================> ............] - ETA: 0 s - loss: 0.2515 - accuracy: 0.9056 426 / 627 [===================> ..........] - ETA: 0 s - loss: 0.2526 - accuracy: 0.9040 459 / 627 [====================> .........] - ETA: 0 s - loss: 0.2537 - accuracy: 0.9022 490 / 627 [======================> .......] - ETA: 0 s - loss: 0.2545 - accuracy: 0.9003 522 / 627 [=======================> ......] - ETA: 0 s - loss: 0.2552 - accuracy: 0.8997 553 / 627 [=========================> ....] - ETA: 0 s - loss: 0.2553 - accuracy: 0.8996 586 / 627 [===========================> ..] - ETA: 0 s - loss: 0.2591 - accuracy: 0.8974 620 / 627 [============================> .] - ETA: 0 s - loss: 0.2612 - accuracy: 0.8964 627 / 627 [============================== ] - 2 s 2 ms / step - loss: 0.2609 - accuracy: 0.8962 - val_loss: 0.7289 - val_accuracy: 0.7152 ## Epoch 8/10 ## 1 / 627 [..............................] - ETA: 0 s - loss: 0.2016 - accuracy: 0.8750 39 / 627 [> .............................] - ETA: 0 s - loss: 0.1916 - accuracy: 0.9519 75 / 627 [==> ...........................] - ETA: 0 s - loss: 0.2067 - accuracy: 0.9317 109 / 627 [====> .........................] - ETA: 0 s - loss: 0.2097 - accuracy: 0.9255 125 / 627 [====> .........................] - ETA: 0 s - loss: 0.2040 - accuracy: 0.9290 142 / 627 [=====> ........................] - ETA: 0 s - loss: 0.2028 - accuracy: 0.9278 165 / 627 [======> .......................] - ETA: 0 s - loss: 0.2021 - accuracy: 0.9258 190 / 627 [========> .....................] - ETA: 0 s - loss: 0.2058 - accuracy: 0.9257 211 / 627 [=========> ....................] - ETA: 0 s - loss: 0.2065 - accuracy: 0.9259 234 / 627 [==========> ...................] - ETA: 0 s - loss: 0.2115 - accuracy: 0.9252 259 / 627 [===========> ..................] - ETA: 0 s - loss: 0.2183 - accuracy: 0.9213 283 / 627 [============> .................] - ETA: 0 s - loss: 0.2159 - accuracy: 0.9231 308 / 627 [=============> ................] - ETA: 0 s - loss: 0.2170 - accuracy: 0.9225 331 / 627 [==============> ...............] - ETA: 0 s - loss: 0.2183 - accuracy: 0.9226 358 / 627 [================> .............] - ETA: 0 s - loss: 0.2184 - accuracy: 0.9214 388 / 627 [=================> ............] - ETA: 0 s - loss: 0.2218 - accuracy: 0.9188 421 / 627 [===================> ..........] - ETA: 0 s - loss: 0.2238 - accuracy: 0.9175 455 / 627 [====================> .........] - ETA: 0 s - loss: 0.2256 - accuracy: 0.9170 485 / 627 [======================> .......] - ETA: 0 s - loss: 0.2263 - accuracy: 0.9160 512 / 627 [=======================> ......] - ETA: 0 s - loss: 0.2263 - accuracy: 0.9155 533 / 627 [========================> .....] - ETA: 0 s - loss: 0.2239 - accuracy: 0.9172 557 / 627 [=========================> ....] - ETA: 0 s - loss: 0.2243 - accuracy: 0.9165 581 / 627 [==========================> ...] - ETA: 0 s - loss: 0.2242 - accuracy: 0.9163 610 / 627 [============================> .] - ETA: 0 s - loss: 0.2253 - accuracy: 0.9152 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.2258 - accuracy: 0.9152 - val_loss: 0.8327 - val_accuracy: 0.6937 ## Epoch 9/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.1488 - accuracy: 1.0000 34 / 627 [> .............................] - ETA: 0 s - loss: 0.1648 - accuracy: 0.9706 59 / 627 [=> ............................] - ETA: 1 s - loss: 0.1746 - accuracy: 0.9513 90 / 627 [===> ..........................] - ETA: 0 s - loss: 0.1658 - accuracy: 0.9528 110 / 627 [====> .........................] - ETA: 0 s - loss: 0.1710 - accuracy: 0.9477 135 / 627 [=====> ........................] - ETA: 0 s - loss: 0.1700 - accuracy: 0.9463 170 / 627 [=======> ......................] - ETA: 0 s - loss: 0.1739 - accuracy: 0.9426 204 / 627 [========> .....................] - ETA: 0 s - loss: 0.1777 - accuracy: 0.9400 240 / 627 [==========> ...................] - ETA: 0 s - loss: 0.1774 - accuracy: 0.9406 271 / 627 [===========> ..................] - ETA: 0 s - loss: 0.1780 - accuracy: 0.9414 304 / 627 [=============> ................] - ETA: 0 s - loss: 0.1796 - accuracy: 0.9404 335 / 627 [===============> ..............] - ETA: 0 s - loss: 0.1770 - accuracy: 0.9410 367 / 627 [================> .............] - ETA: 0 s - loss: 0.1786 - accuracy: 0.9390 402 / 627 [==================> ...........] - ETA: 0 s - loss: 0.1780 - accuracy: 0.9384 434 / 627 [===================> ..........] - ETA: 0 s - loss: 0.1828 - accuracy: 0.9335 469 / 627 [=====================> ........] - ETA: 0 s - loss: 0.1851 - accuracy: 0.9315 500 / 627 [======================> .......] - ETA: 0 s - loss: 0.1865 - accuracy: 0.9310 535 / 627 [========================> .....] - ETA: 0 s - loss: 0.1868 - accuracy: 0.9301 568 / 627 [==========================> ...] - ETA: 0 s - loss: 0.1894 - accuracy: 0.9294 604 / 627 [===========================> ..] - ETA: 0 s - loss: 0.1912 - accuracy: 0.9280 627 / 627 [============================== ] - 2 s 3 ms / step - loss: 0.1926 - accuracy: 0.9275 - val_loss: 0.8951 - val_accuracy: 0.7143 ## Epoch 10/10 ## 1 / 627 [..............................] - ETA: 1 s - loss: 0.4191 - accuracy: 0.7500 35 / 627 [> .............................] - ETA: 0 s - loss: 0.1559 - accuracy: 0.9536 63 / 627 [==> ...........................] - ETA: 0 s - loss: 0.1439 - accuracy: 0.9643 99 / 627 [===> ..........................] - ETA: 0 s - loss: 0.1418 - accuracy: 0.9646 129 / 627 [=====> ........................] - ETA: 0 s - loss: 0.1377 - accuracy: 0.9632 164 / 627 [======> .......................] - ETA: 0 s - loss: 0.1360 - accuracy: 0.9604 198 / 627 [========> .....................] - ETA: 0 s - loss: 0.1369 - accuracy: 0.9621 231 / 627 [==========> ...................] - ETA: 0 s - loss: 0.1409 - accuracy: 0.9605 266 / 627 [===========> ..................] - ETA: 0 s - loss: 0.1423 - accuracy: 0.9586 303 / 627 [=============> ................] - ETA: 0 s - loss: 0.1435 - accuracy: 0.9571 340 / 627 [===============> ..............] - ETA: 0 s - loss: 0.1466 - accuracy: 0.9555 377 / 627 [=================> ............] - ETA: 0 s - loss: 0.1508 - accuracy: 0.9532 408 / 627 [==================> ...........] - ETA: 0 s - loss: 0.1513 - accuracy: 0.9522 444 / 627 [====================> .........] - ETA: 0 s - loss: 0.1544 - accuracy: 0.9507 482 / 627 [======================> .......] - ETA: 0 s - loss: 0.1568 - accuracy: 0.9502 516 / 627 [=======================> ......] - ETA: 0 s - loss: 0.1575 - accuracy: 0.9494 554 / 627 [=========================> ....] - ETA: 0 s - loss: 0.1601 - accuracy: 0.9483 590 / 627 [===========================> ..] - ETA: 0 s - loss: 0.1594 - accuracy: 0.9481 622 / 627 [============================> .] - ETA: 0 s - loss: 0.1589 - accuracy: 0.9488 627 / 627 [============================== ] - 1 s 2 ms / step - loss: 0.1596 - accuracy: 0.9483 - val_loss: 1.0364 - val_accuracy: 0.7044 ## <keras.src.callbacks.History object at 0x13828df40>
Modellgüte
= (model.predict(X_test) > 0.5 ).astype("int32" )## 1 / 111 [..............................] - ETA: 10 s 50 / 111 [============> .................] - ETA: 0 s 107 / 111 [===========================> ..] - ETA: 0 s 111 / 111 [============================== ] - 0 s 951 us / step= accuracy_score(y_test, y_pred)print (f"Test Accuracy: { accuracy} " )## Test Accuracy: 0.7044167610419027
Fazit
Die Modellgüte der 2. Pipeline ist etwas geringer als in der ersten. Die zweite Hidden-Layer muss also nicht zur Modellgüte positiv beitragen. Ähnliches gilt für die Batch-Size; wobei eigentlich kleine Batch-Sizes für diesen eher kleinen Datensatz sinnvoll sein sollten …
Pipeline mit (englischen) Word Embedding
Diese Pipeline orientiert sich an diesem Beispiel von Tensorflow .
Daten
import pandas as pd= "https://github.com/sebastiansauer/pradadata/raw/master/data-raw/germeval_train.csv" = pd.read_csv(train_file_path)= "https://github.com/sebastiansauer/pradadata/raw/master/data-raw/germeval_test.csv" = pd.read_csv(test_file_path)
Prädiktor-Dataframes als Arrays:
= d_train["text" ].values= d_test["text" ].values
Module
tensorflow-hub ist übrigens NICHT mehr nötig. Das Paket ist jetzt Teil von tensorflow.
import osimport numpy as npimport tensorflow as tfimport tensorflow_hub as hub
GPU
Testen, ob eine GPU verfügbar ist:
'GPU' ) ## []
print ("TF Version: " , tf.__version__)## TF Version: 2.13.1 print ("Eager mode: " , tf.executing_eagerly())## Eager mode: True print ("Hub version: " , hub.__version__)## Hub version: 0.14.0 print ("GPU is" , "available" if tf.config.list_physical_devices("GPU" ) else "NOT AVAILABLE" )## GPU is NOT AVAILABLE
Tja, leider nein.
Wort-Einbettungen
= "https://tfhub.dev/google/nnlm-en-dim50/2" = hub.KerasLayer(embedding, input_shape= [], = tf.string, trainable= True )
Modell
= tf.keras.Sequential()16 , activation= 'relu' ))1 ))## Model: "sequential_2" ## _________________________________________________________________ ## Layer (type) Output Shape Param # ## ================================================================= ## keras_layer (KerasLayer) (None, 50) 48190600 ## ## dense_5 (Dense) (None, 16) 816 ## ## dense_6 (Dense) (None, 1) 17 ## ## ================================================================= ## Total params: 48191433 (183.84 MB) ## Trainable params: 48191433 (183.84 MB) ## Non-trainable params: 0 (0.00 Byte) ## _________________________________________________________________
compile (optimizer= 'adam' ,= tf.keras.losses.BinaryCrossentropy(from_logits= True ),= ['accuracy' ])
Trainieren
= 10 , = 8 , = (X_test, y_test),= 1 )
Epoch 1/10
627/627 [==============================] - 490s 781ms/step - loss: 0.6232 - accuracy: 0.6638 - val_loss: 0.6093 - val_accuracy: 0.6628
Epoch 2/10
627/627 [==============================] - 477s 760ms/step - loss: 0.4541 - accuracy: 0.7686 - val_loss: 0.6536 - val_accuracy: 0.6761
Epoch 3/10
627/627 [==============================] - 482s 769ms/step - loss: 0.2762 - accuracy: 0.8794 - val_loss: 0.8118 - val_accuracy: 0.6526
Epoch 4/10
627/627 [==============================] - 521s 831ms/step - loss: 0.1671 - accuracy: 0.9367 - val_loss: 1.0416 - val_accuracy: 0.6467
Epoch 5/10
627/627 [==============================] - 456s 727ms/step - loss: 0.0936 - accuracy: 0.9689 - val_loss: 1.2981 - val_accuracy: 0.6486
Epoch 6/10
627/627 [==============================] - 455s 726ms/step - loss: 0.0478 - accuracy: 0.9872 - val_loss: 1.5631 - val_accuracy: 0.6297
Epoch 7/10
627/627 [==============================] - 456s 727ms/step - loss: 0.0240 - accuracy: 0.9954 - val_loss: 1.8281 - val_accuracy: 0.6285
Epoch 8/10
627/627 [==============================] - 455s 726ms/step - loss: 0.0101 - accuracy: 0.9982 - val_loss: 2.0636 - val_accuracy: 0.6334
Epoch 9/10
627/627 [==============================] - 459s 732ms/step - loss: 0.0067 - accuracy: 0.9986 - val_loss: 2.2470 - val_accuracy: 0.6291
Epoch 10/10
627/627 [==============================] - 455s 727ms/step - loss: 0.0046 - accuracy: 0.9992 - val_loss: 2.3786 - val_accuracy: 0.6277
<keras.src.callbacks.History object at 0x148309730>
Modellgüte
= (model.predict(X_test) > 0.5 ).astype("int32" )= accuracy_score(y_test, y_pred)print (f"Test Accuracy: { accuracy} " )
111/111 [==============================] - 17s 151ms/step
Test Accuracy: 0.6276896942242356
Fazit
Naja, dafür dass es englische Wortvektoren waren, gar nicht so schlecht 🤣