Example: Drug rating regressor tutorial
We are going to use the Drug Review dataset. (Felix Gräßer, Surya Kallumadi, Hagen Malberg, and Sebastian Zaunseder. 2018. Aspect-Based Sentiment Analysis of Drug Reviews Applying Cross-Domain and Cross-Data Learning. In Proceedings of the 2018 International Conference on Digital Health (DH ‘18). ACM, New York, NY, USA, 121-125. DOI: Web Link)
!curl -O https://archive.ics.uci.edu/ml/machine-learning-databases/00461/drugLib_raw.zip
!unzip drugLib_raw.zip
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
100 1106k 100 1106k 0 0 173k 0 0:00:06 0:00:06 --:--:-- 224k
Archive: drugLib_raw.zip
inflating: drugLibTest_raw.tsv
inflating: drugLibTrain_raw.tsv
import pandas as pd
train = pd.read_csv("drugLibTrain_raw.tsv", sep='\t')
test = pd.read_csv("drugLibTest_raw.tsv", sep='\t')
print(train)
Unnamed: 0 urlDrugName rating effectiveness \
0 2202 enalapril 4 Highly Effective
1 3117 ortho-tri-cyclen 1 Highly Effective
2 1146 ponstel 10 Highly Effective
3 3947 prilosec 3 Marginally Effective
4 1951 lyrica 2 Marginally Effective
... ... ... ... ...
3102 1039 vyvanse 10 Highly Effective
3103 3281 zoloft 1 Ineffective
3104 1664 climara 2 Marginally Effective
3105 2621 trileptal 8 Considerably Effective
3106 2748 micardis 4 Moderately Effective
sideEffects condition \
0 Mild Side Effects management of congestive heart failure
1 Severe Side Effects birth prevention
2 No Side Effects menstrual cramps
3 Mild Side Effects acid reflux
4 Severe Side Effects fibromyalgia
... ... ...
3102 Mild Side Effects adhd
3103 Extremely Severe Side Effects depression
3104 Moderate Side Effects total hysterctomy
3105 Mild Side Effects epilepsy
3106 Moderate Side Effects high blood pressure
benefitsReview \
0 slowed the progression of left ventricular dys...
1 Although this type of birth control has more c...
2 I was used to having cramps so badly that they...
3 The acid reflux went away for a few months aft...
4 I think that the Lyrica was starting to help w...
... ...
3102 Increased focus, attention, productivity. Bett...
3103 Emotions were somewhat blunted. Less moodiness.
3104 ---
3105 Controlled complex partial seizures.
3106 The drug Micardis did seem to alleviate my hig...
sideEffectsReview \
0 cough, hypotension , proteinuria, impotence , ...
1 Heavy Cycle, Cramps, Hot Flashes, Fatigue, Lon...
2 Heavier bleeding and clotting than normal.
3 Constipation, dry mouth and some mild dizzines...
4 I felt extremely drugged and dopey. Could not...
... ...
3102 Restless legs at night, insomnia, headache (so...
3103 Weight gain, extreme tiredness during the day,...
3104 Constant issues with the patch not staying on....
3105 Dizziness, fatigue, nausea
3106 I find when I am taking Micardis that I tend t...
commentsReview
0 monitor blood pressure , weight and asses for ...
1 I Hate This Birth Control, I Would Not Suggest...
2 I took 2 pills at the onset of my menstrual cr...
3 I was given Prilosec prescription at a dose of...
4 See above
... ...
3102 I took adderall once as a child, and it made m...
3103 I was on Zoloft for about 2 years total. I am ...
3104 ---
3105 Started at 2 doses of 300 mg a day and worked ...
3106 I take Micardis in pill form once daily.
[3107 rows x 9 columns]
Multi-text model building
The rating will be predicted using benefitsReview, commentsReview, sideEffectsReview and urlDrugName
from gianlp.models import CharEmbeddingSequence, RNNDigest, KerasWrapper, PreTrainedWordEmbeddingSequence
WARNING:nlp_builder:The NLP builder disables all tensorflow-related logging
Word embedding for reviews
We are going to use a word embedding for the benefitsReview, commentsReview and sideEffectsReview. We first download uncased 50 dimensions glove word embedding, then we translate it to word2vec format.
!curl -O http://downloads.cs.stanford.edu/nlp/data/glove.6B.zip
!unzip glove.6B
!python -m gensim.scripts.glove2word2vec --input glove.6B.50d.txt --output glove.6B.50d.w2vformat.txt
% Total % Received % Xferd Average Speed Time Time Time Current
Dload Upload Total Spent Left Speed
0 822M 0 14216 0 0 29616 0 8:05:12 --:--:-- 8:05:12 29555
We create our word embedding
help(PreTrainedWordEmbeddingSequence.__init__)
Help on function __init__ in module gianlp.models.text_representations.pre_trained_word_embedding:
__init__(self, word2vec_src: Union[str, gensim.models.keyedvectors.Word2VecKeyedVectors], tokenizer: Callable[[str], List[str]], trainable: bool = False, sequence_maxlen: int = 20, random_state: int = 42)
:param word2vec_src: path to word2vec format .txt file or gensim KeyedVectors
:param tokenizer: a tokenizer function that transforms each string into a list of string tokens
the tokens transformed should match the keywords in the pretrained word embeddings
the function must support serialization through pickle
:param trainable: if the vectors are trainable with keras backprop
:param sequence_maxlen: The maximum allowed sequence length
:param random_state: the random seed used for random processes
We will see the 80% percentile as a guideline for defining our sequence_maxlen.
train['commentsReview'].fillna("").map(lambda x: len(x.split(" "))).quantile(0.8)
76.0
With the tokenizer created, we can now create the PreTrainedWordEmbeddingSequence.
def split_tokenizer(text):
return text.lower().split(" ")
word_emb = PreTrainedWordEmbeddingSequence("glove.6B.50d.w2vformat.txt",
tokenizer=split_tokenizer,
sequence_maxlen=90)
word_emb.outputs_shape
(90, 50), float32
Review digest
We build a different RNN for each review
benefits_digest = RNNDigest(word_emb, units_per_layer=20, rnn_type='gru', stacked_layers=1)
comments_digest = RNNDigest(word_emb, units_per_layer=20, rnn_type='gru', stacked_layers=1)
side_eff_digest = RNNDigest(word_emb, units_per_layer=20, rnn_type='gru', stacked_layers=1)
benefits_digest.outputs_shape, comments_digest.outputs_shape, side_eff_digest.outputs_shape
WARNING:nlp_builder:If the model and wrapper inputs mismatch it will only be noticed when building, before that output shape is an estimate and does not assert inputs.
WARNING:nlp_builder:If the model and wrapper inputs mismatch it will only be noticed when building, before that output shape is an estimate and does not assert inputs.
WARNING:nlp_builder:If the model and wrapper inputs mismatch it will only be noticed when building, before that output shape is an estimate and does not assert inputs.
((20,), float32, (20,), float32, (20,), float32)
Char embedding for drug name
Since we know nothing about drug names I hope char embedding will help.
char_emb = CharEmbeddingSequence(embedding_dimension=16, sequence_maxlen=30)
char_emb.outputs_shape
(30, 16), float32
Convolutional digest for drug name
We convolute the drug name to, hopefully, extract useful features.
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Dense, Conv1D, GlobalMaxPooling1D, Concatenate, Input
text_input = Input((30, 16))
conv_1char = Conv1D(5, kernel_size=1, padding="same")(text_input)
conv_2char = Conv1D(5, kernel_size=2, padding="same")(text_input)
conv_3char = Conv1D(5, kernel_size=3, padding="same")(text_input)
conv_4char = Conv1D(5, kernel_size=4, padding="same")(text_input)
state_seq = Concatenate()([conv_1char, conv_2char, conv_3char, conv_4char])
result = GlobalMaxPooling1D()(state_seq)
drug_encoder = Model(inputs=text_input, outputs=result)
drug_encoder.summary()
Model: "model"
__________________________________________________________________________________________________
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_4 (InputLayer) [(None, 30, 16)] 0 []
conv1d (Conv1D) (None, 30, 5) 85 ['input_4[0][0]']
conv1d_1 (Conv1D) (None, 30, 5) 165 ['input_4[0][0]']
conv1d_2 (Conv1D) (None, 30, 5) 245 ['input_4[0][0]']
conv1d_3 (Conv1D) (None, 30, 5) 325 ['input_4[0][0]']
concatenate (Concatenate) (None, 30, 20) 0 ['conv1d[0][0]',
'conv1d_1[0][0]',
'conv1d_2[0][0]',
'conv1d_3[0][0]']
global_max_pooling1d (GlobalMa (None, 20) 0 ['concatenate[0][0]']
xPooling1D)
==================================================================================================
Total params: 820
Trainable params: 820
Non-trainable params: 0
__________________________________________________________________________________________________
drug_encoder = KerasWrapper(char_emb, drug_encoder)
print(drug_encoder)
WARNING:nlp_builder:If the model and wrapper inputs mismatch it will only be noticed when building, before that output shape is an estimate and does not assert inputs.
Model | Inputs shape | Output shape |Trainable| Total | Connected to
| | | weights | weights |
==============================================================================================================
7fcd9cdd9a60 CharEmbe| (30,), int32 | (30, 16), float32 | ? | ? |
7fcd201327c0 KerasWra| (30, 16), float32 | (20,), float32 | ? | ? |7fcd9cdd9a60 CharEmb
==============================================================================================================
| | | ? | ? |
Final model building
We have 20 dimensions for each review and 20 for the drug name, we have to build a regressor for that inputs
drug_name = Input((20,))
review1 = Input((20,))
review2 = Input((20,))
review3 = Input((20,))
full_state = Concatenate()([drug_name, review1, review2, review3])
dense1 = Dense(50, activation='tanh')(full_state)
dense2 = Dense(50, activation='relu')(dense1)
dense3 = Dense(50, activation='relu')(dense2)
out = Dense(1, activation='relu')(dense3)
regressor = Model(inputs = [drug_name, review1, review2, review3], outputs = out)
train.columns
Index(['Unnamed: 0', 'urlDrugName', 'rating', 'effectiveness', 'sideEffects',
'condition', 'benefitsReview', 'sideEffectsReview', 'commentsReview'],
dtype='object')
We use multi-text for the inputs indicating we will be fitting different texts
regressor = KerasWrapper([('urlDrugName',[drug_encoder]),
('benefitsReview', [benefits_digest]),
('sideEffectsReview', [comments_digest]),
('commentsReview', [side_eff_digest])], regressor)
Model build
regressor.build(train['commentsReview'].fillna("").tolist()+
train['sideEffectsReview'].fillna("").tolist()+
train['benefitsReview'].fillna("").tolist()+
train['urlDrugName'].fillna("").tolist())
print(regressor)
Model | Inputs shape | Output shape |Trainable| Total | Connected to
| | | weights | weights |
==============================================================================================================
7fcda9e18c10 PreTrain| (90,), int32 | (90, 50), float32 | 0 | 20000100|
7fcd20356460 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fcda9e18c10 PreTrai
7fcd205f1430 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fcda9e18c10 PreTrai
7fcd900b0700 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fcda9e18c10 PreTrai
7fcd9cdd9a60 CharEmbe| (30,), int32 | (30, 16), float32 | 1456 | 1456 |
7fcd201327c0 KerasWra| (30, 16), float32 | (20,), float32 | 2276 | 2276 |7fcd9cdd9a60 CharEmb
7fcd205f5940 KerasWra| (20,), float32 | (1,), float32 | 24437 | 20024537|"urlDrugName": 7fcd2
| (20,), float32 | | | |"benefitsReview": 7f
| (20,), float32 | | | |"sideEffectsReview":
| (20,), float32 | | | |"commentsReview": 7f
==============================================================================================================
| | | 24437 | 20024537|
regressor.compile(optimizer="adam", loss="mse", metrics=["mae"])
from tensorflow.keras.callbacks import EarlyStopping
early_stopping = EarlyStopping(patience=20, monitor='val_loss', restore_best_weights=True)
hst = regressor.fit(train, train['rating'].values,
batch_size=256, epochs=200, validation_split=0.1,
callbacks=[early_stopping])
Epoch 1/200
10/10 [==============================] - 15s 481ms/step - loss: 44.8526 - mae: 6.0467 - val_loss: 26.3905 - val_mae: 4.5891
Epoch 2/200
10/10 [==============================] - 2s 212ms/step - loss: 16.2550 - mae: 3.5820 - val_loss: 9.6836 - val_mae: 2.6122
Epoch 3/200
10/10 [==============================] - 2s 202ms/step - loss: 9.3684 - mae: 2.2984 - val_loss: 11.4970 - val_mae: 2.4750
Epoch 4/200
10/10 [==============================] - 2s 208ms/step - loss: 9.0430 - mae: 2.3599 - val_loss: 9.3396 - val_mae: 2.6122
Epoch 5/200
10/10 [==============================] - 2s 200ms/step - loss: 8.5684 - mae: 2.5149 - val_loss: 9.6518 - val_mae: 2.6384
Epoch 6/200
10/10 [==============================] - 2s 200ms/step - loss: 8.5556 - mae: 2.4241 - val_loss: 10.2140 - val_mae: 2.6156
Epoch 7/200
10/10 [==============================] - 2s 201ms/step - loss: 8.4898 - mae: 2.3640 - val_loss: 9.7481 - val_mae: 2.5665
Epoch 8/200
10/10 [==============================] - 2s 198ms/step - loss: 8.0966 - mae: 2.3293 - val_loss: 9.5756 - val_mae: 2.5541
Epoch 9/200
10/10 [==============================] - 2s 199ms/step - loss: 8.4019 - mae: 2.3864 - val_loss: 9.5679 - val_mae: 2.5755
Epoch 10/200
10/10 [==============================] - 2s 200ms/step - loss: 8.3462 - mae: 2.3844 - val_loss: 9.5266 - val_mae: 2.5618
Epoch 11/200
10/10 [==============================] - 2s 202ms/step - loss: 8.3780 - mae: 2.3694 - val_loss: 9.9598 - val_mae: 2.6336
Epoch 12/200
10/10 [==============================] - 2s 192ms/step - loss: 8.2200 - mae: 2.3633 - val_loss: 9.3986 - val_mae: 2.5457
Epoch 13/200
10/10 [==============================] - 2s 206ms/step - loss: 8.1503 - mae: 2.3485 - val_loss: 8.8225 - val_mae: 2.4505
Epoch 14/200
10/10 [==============================] - 2s 207ms/step - loss: 8.1233 - mae: 2.3438 - val_loss: 9.4460 - val_mae: 2.5943
Epoch 15/200
10/10 [==============================] - 2s 207ms/step - loss: 8.0737 - mae: 2.3433 - val_loss: 9.3705 - val_mae: 2.5378
Epoch 16/200
10/10 [==============================] - 2s 204ms/step - loss: 8.0723 - mae: 2.3258 - val_loss: 9.4884 - val_mae: 2.6086
Epoch 17/200
10/10 [==============================] - 2s 206ms/step - loss: 7.9773 - mae: 2.3296 - val_loss: 9.4471 - val_mae: 2.5593
Epoch 18/200
10/10 [==============================] - 2s 199ms/step - loss: 7.5464 - mae: 2.2537 - val_loss: 8.9789 - val_mae: 2.5062
Epoch 19/200
10/10 [==============================] - 2s 208ms/step - loss: 7.4574 - mae: 2.2392 - val_loss: 8.3895 - val_mae: 2.4031
Epoch 20/200
10/10 [==============================] - 2s 199ms/step - loss: 7.0000 - mae: 2.1540 - val_loss: 8.7761 - val_mae: 2.3631
Epoch 21/200
10/10 [==============================] - 2s 201ms/step - loss: 6.9204 - mae: 2.1558 - val_loss: 8.6919 - val_mae: 2.3652
Epoch 22/200
10/10 [==============================] - 2s 203ms/step - loss: 6.3791 - mae: 2.0272 - val_loss: 8.7601 - val_mae: 2.3497
Epoch 23/200
10/10 [==============================] - 2s 202ms/step - loss: 6.0366 - mae: 1.9443 - val_loss: 8.1706 - val_mae: 2.2950
Epoch 24/200
10/10 [==============================] - 2s 202ms/step - loss: 5.8741 - mae: 1.9132 - val_loss: 8.1911 - val_mae: 2.3053
Epoch 25/200
10/10 [==============================] - 2s 207ms/step - loss: 5.4787 - mae: 1.8285 - val_loss: 7.4014 - val_mae: 2.1681
Epoch 26/200
10/10 [==============================] - 2s 210ms/step - loss: 5.2076 - mae: 1.7880 - val_loss: 7.2368 - val_mae: 2.0988
Epoch 27/200
10/10 [==============================] - 2s 214ms/step - loss: 4.8116 - mae: 1.7061 - val_loss: 6.7278 - val_mae: 2.0793
Epoch 28/200
10/10 [==============================] - 2s 209ms/step - loss: 4.7443 - mae: 1.6859 - val_loss: 7.3081 - val_mae: 2.1378
Epoch 29/200
10/10 [==============================] - 2s 207ms/step - loss: 4.6793 - mae: 1.6791 - val_loss: 6.6134 - val_mae: 2.0334
Epoch 30/200
10/10 [==============================] - 2s 203ms/step - loss: 4.3276 - mae: 1.6048 - val_loss: 6.6267 - val_mae: 2.0093
Epoch 31/200
10/10 [==============================] - 2s 209ms/step - loss: 4.3379 - mae: 1.6148 - val_loss: 6.1516 - val_mae: 1.9605
Epoch 32/200
10/10 [==============================] - 2s 200ms/step - loss: 4.0363 - mae: 1.5443 - val_loss: 6.4251 - val_mae: 1.9917
Epoch 33/200
10/10 [==============================] - 2s 207ms/step - loss: 3.9923 - mae: 1.5460 - val_loss: 6.3441 - val_mae: 1.9595
Epoch 34/200
10/10 [==============================] - 2s 204ms/step - loss: 3.8146 - mae: 1.4986 - val_loss: 6.4183 - val_mae: 1.9715
Epoch 35/200
10/10 [==============================] - 2s 205ms/step - loss: 3.7252 - mae: 1.4755 - val_loss: 6.3719 - val_mae: 1.9541
Epoch 36/200
10/10 [==============================] - 2s 211ms/step - loss: 3.8593 - mae: 1.5039 - val_loss: 5.9912 - val_mae: 1.8993
Epoch 37/200
10/10 [==============================] - 2s 203ms/step - loss: 3.4515 - mae: 1.4152 - val_loss: 6.1809 - val_mae: 1.8888
Epoch 38/200
10/10 [==============================] - 2s 204ms/step - loss: 3.5217 - mae: 1.4361 - val_loss: 6.2674 - val_mae: 1.9094
Epoch 39/200
10/10 [==============================] - 2s 201ms/step - loss: 3.6698 - mae: 1.4673 - val_loss: 6.7534 - val_mae: 1.9594
Epoch 40/200
10/10 [==============================] - 2s 208ms/step - loss: 3.4128 - mae: 1.3989 - val_loss: 5.5831 - val_mae: 1.8298
Epoch 41/200
10/10 [==============================] - 2s 206ms/step - loss: 3.3113 - mae: 1.3933 - val_loss: 6.0933 - val_mae: 1.8533
Epoch 42/200
10/10 [==============================] - 2s 203ms/step - loss: 3.2604 - mae: 1.3622 - val_loss: 6.1341 - val_mae: 1.8971
Epoch 43/200
10/10 [==============================] - 2s 202ms/step - loss: 3.0881 - mae: 1.3521 - val_loss: 6.1046 - val_mae: 1.8776
Epoch 44/200
10/10 [==============================] - 2s 206ms/step - loss: 3.0806 - mae: 1.3335 - val_loss: 5.5088 - val_mae: 1.7957
Epoch 45/200
10/10 [==============================] - 2s 198ms/step - loss: 2.9925 - mae: 1.3121 - val_loss: 6.2356 - val_mae: 1.8951
Epoch 46/200
10/10 [==============================] - 2s 211ms/step - loss: 2.9967 - mae: 1.3252 - val_loss: 6.5368 - val_mae: 1.9281
Epoch 47/200
10/10 [==============================] - 2s 211ms/step - loss: 2.8649 - mae: 1.2917 - val_loss: 6.2750 - val_mae: 1.8579
Epoch 48/200
10/10 [==============================] - 2s 200ms/step - loss: 2.9601 - mae: 1.3048 - val_loss: 5.7847 - val_mae: 1.8140
Epoch 49/200
10/10 [==============================] - 2s 202ms/step - loss: 2.8342 - mae: 1.2702 - val_loss: 5.8610 - val_mae: 1.8357
Epoch 50/200
10/10 [==============================] - 2s 200ms/step - loss: 2.8497 - mae: 1.2969 - val_loss: 5.8004 - val_mae: 1.8159
Epoch 51/200
10/10 [==============================] - 2s 201ms/step - loss: 2.6426 - mae: 1.2303 - val_loss: 6.0414 - val_mae: 1.8613
Epoch 52/200
10/10 [==============================] - 2s 205ms/step - loss: 2.6518 - mae: 1.2339 - val_loss: 6.0281 - val_mae: 1.8042
Epoch 53/200
10/10 [==============================] - 2s 200ms/step - loss: 2.5120 - mae: 1.2305 - val_loss: 6.5550 - val_mae: 1.9098
Epoch 54/200
10/10 [==============================] - 2s 201ms/step - loss: 2.5509 - mae: 1.2092 - val_loss: 6.2900 - val_mae: 1.8720
Epoch 55/200
10/10 [==============================] - 2s 204ms/step - loss: 2.4135 - mae: 1.1927 - val_loss: 5.6932 - val_mae: 1.7901
Epoch 56/200
10/10 [==============================] - 2s 200ms/step - loss: 2.3593 - mae: 1.1683 - val_loss: 5.9240 - val_mae: 1.7881
Epoch 57/200
10/10 [==============================] - 2s 204ms/step - loss: 2.3442 - mae: 1.1779 - val_loss: 6.2036 - val_mae: 1.8304
Epoch 58/200
10/10 [==============================] - 2s 204ms/step - loss: 2.2430 - mae: 1.1506 - val_loss: 6.2946 - val_mae: 1.8458
Epoch 59/200
10/10 [==============================] - 2s 203ms/step - loss: 2.2977 - mae: 1.1396 - val_loss: 5.9096 - val_mae: 1.8211
Epoch 60/200
10/10 [==============================] - 2s 205ms/step - loss: 2.1639 - mae: 1.1418 - val_loss: 5.5687 - val_mae: 1.7562
Epoch 61/200
10/10 [==============================] - 2s 207ms/step - loss: 2.0508 - mae: 1.1006 - val_loss: 5.4898 - val_mae: 1.7547
Epoch 62/200
10/10 [==============================] - 2s 209ms/step - loss: 2.1567 - mae: 1.1206 - val_loss: 5.9577 - val_mae: 1.8105
Epoch 63/200
10/10 [==============================] - 2s 203ms/step - loss: 2.0232 - mae: 1.0940 - val_loss: 6.1129 - val_mae: 1.8515
Epoch 64/200
10/10 [==============================] - 2s 207ms/step - loss: 2.0656 - mae: 1.0940 - val_loss: 6.0248 - val_mae: 1.8243
Epoch 65/200
10/10 [==============================] - 2s 202ms/step - loss: 1.8792 - mae: 1.0665 - val_loss: 5.8070 - val_mae: 1.7824
Epoch 66/200
10/10 [==============================] - 2s 202ms/step - loss: 1.9632 - mae: 1.0711 - val_loss: 5.9124 - val_mae: 1.7691
Epoch 67/200
10/10 [==============================] - 2s 194ms/step - loss: 2.0155 - mae: 1.0864 - val_loss: 6.1221 - val_mae: 1.8192
Epoch 68/200
10/10 [==============================] - 2s 201ms/step - loss: 1.8581 - mae: 1.0461 - val_loss: 6.1425 - val_mae: 1.8245
Epoch 69/200
10/10 [==============================] - 2s 204ms/step - loss: 1.8283 - mae: 1.0437 - val_loss: 6.1746 - val_mae: 1.8088
Epoch 70/200
10/10 [==============================] - 2s 204ms/step - loss: 1.7469 - mae: 1.0271 - val_loss: 6.1359 - val_mae: 1.8005
Epoch 71/200
10/10 [==============================] - 2s 199ms/step - loss: 1.7665 - mae: 1.0200 - val_loss: 6.1712 - val_mae: 1.7875
Epoch 72/200
10/10 [==============================] - 2s 204ms/step - loss: 1.7146 - mae: 1.0164 - val_loss: 6.4598 - val_mae: 1.8229
Epoch 73/200
10/10 [==============================] - 2s 202ms/step - loss: 1.7306 - mae: 1.0091 - val_loss: 6.0780 - val_mae: 1.7763
Epoch 74/200
10/10 [==============================] - 2s 199ms/step - loss: 1.6304 - mae: 0.9804 - val_loss: 5.8898 - val_mae: 1.7796
Epoch 75/200
10/10 [==============================] - 2s 201ms/step - loss: 1.5902 - mae: 0.9769 - val_loss: 5.8832 - val_mae: 1.7269
Epoch 76/200
10/10 [==============================] - 2s 202ms/step - loss: 1.5511 - mae: 0.9606 - val_loss: 6.5852 - val_mae: 1.8586
Epoch 77/200
10/10 [==============================] - 2s 208ms/step - loss: 1.5500 - mae: 0.9463 - val_loss: 6.2416 - val_mae: 1.8236
Epoch 78/200
10/10 [==============================] - 2s 199ms/step - loss: 1.5069 - mae: 0.9439 - val_loss: 6.6130 - val_mae: 1.9194
Epoch 79/200
10/10 [==============================] - 2s 207ms/step - loss: 1.4580 - mae: 0.9301 - val_loss: 6.3010 - val_mae: 1.8422
Epoch 80/200
10/10 [==============================] - 2s 202ms/step - loss: 1.4190 - mae: 0.9184 - val_loss: 5.8961 - val_mae: 1.7574
Epoch 81/200
10/10 [==============================] - 2s 207ms/step - loss: 1.4412 - mae: 0.9182 - val_loss: 6.0843 - val_mae: 1.7733
Evaluation
preds = regressor.predict(test)
preds
array([[7.9903536],
[9.336563 ],
[1.3078178],
...,
[8.757404 ],
[9.284457 ],
[8.320618 ]], dtype=float32)
preds.min(), preds.max()
(0.0, 10.139272)
print("MAE: ", abs(test['rating'].values-preds.flatten()).mean())
MAE: 1.8476456113691841
Serialization and deserialization
from gianlp.models import BaseModel
data = regressor.serialize()
regressor2 = BaseModel.deserialize(data)
regressor2.predict(test)
array([[7.9903536],
[9.336563 ],
[1.3078178],
...,
[8.757404 ],
[9.284457 ],
[8.320618 ]], dtype=float32)
print(regressor2)
Model | Inputs shape | Output shape |Trainable| Total | Connected to
| | | weights | weights |
==============================================================================================================
7fcbfb9dbdc0 PreTrain| (90,), int32 | (90, 50), float32 | 0 | 20000100|
7fcc095f1130 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fcbfb9dbdc0 PreTrai
7fccada16100 PreTrain| (90,), int32 | (90, 50), float32 | 0 | 20000100|
7fccae14c8e0 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fccada16100 PreTrai
7fccd117e280 PreTrain| (90,), int32 | (90, 50), float32 | 0 | 20000100|
7fcd2065d160 KerasWra| (90, 50), float32 | (20,), float32 | 4320 | 20004420|7fccd117e280 PreTrai
7fccd3688280 CharEmbe| (30,), int32 | (30, 16), float32 | 1456 | 1456 |
7fccd3688f40 KerasWra| (30, 16), float32 | (20,), float32 | 2276 | 2276 |7fccd3688280 CharEmb
7fcd2065d250 KerasWra| (20,), float32 | (1,), float32 | 24437 | 60024737|"urlDrugName": 7fccd
| (20,), float32 | | | |"benefitsReview": 7f
| (20,), float32 | | | |"sideEffectsReview":
| (20,), float32 | | | |"commentsReview": 7f
==============================================================================================================
| | | 24437 | 60024737|