keras val_loss：0.0000e+00 - val_accuracy：0.0000e+00

Question

In my network the validation metrics if fixed on 0.0000e+00 from the epoch.

I've looked around that few people had the same problem but I'm not be able to fix it following same advices.

Rows are shuffled and label is already transformaned into float32. These are suggestions I've found on similar questions. Can you tell me what i'm wrong?

# data here https://www.kaggle.com/crowdflower/twitter-airline-sentiment
dtf_data = pd.read_csv(str_path + "Tweets.csv")

def clean_data(str_text): 
    lst_tokens = str_text.split()
    lst_tokens = [w for w in lst_tokens if not any(True for e in ["@", "#"] if e in w)]
    table = str.maketrans('', '', string.punctuation)
    lst_tokens = [c.translate(table) for c in lst_tokens]
    lst_tokens = [str_word for str_word in lst_tokens if str_word.isalpha()] # solo parole
    lst_stop_words = set(stopwords.words('english'))
    lst_tokens = [str_word for str_word in lst_tokens if not str_word in lst_stop_words and len(str_word) > 1]
    return " ".join(lst_tokens)

dtf_data["text_cleaned"] = dtf_data["text"].apply(lambda x: clean_data(x))
dtf_data["y"] = (dtf_data["airline_sentiment"] == "positive").astype(int)

X_train, X_test, y_train, y_test = train_test_split(dtf_data["text_cleaned"], dtf_data["y"], test_size=.25, random_state=0)

y_train = np.array(y_train, dtype = 'float32')
y_test = np.array(y_test, dtype = 'float32')

def build_corpus(dtf_in, str_col):
    corpus = []
    for sentence in dtf_in[str_col].iteritems():
        word_list = sentence[1].split()
        corpus.append(word_list)

    return corpus

corpus = build_corpus(dtf_data, 'text_cleaned')

model2vec = word2vec.Word2Vec(corpus, size=50)

tokenizer = Tokenizer()
tokenizer.fit_on_texts(corpus)

corpus_train, corpus_test = train_test_split(corpus, test_size=.25, random_state=0)
sequences_train = tokenizer.texts_to_sequences(corpus_train)
sequences_test = tokenizer.texts_to_sequences(corpus_test)

word_index = tokenizer.word_index
print('Found %s unique tokens' % len(word_index))

data_train = pad_sequences(sequences_train, maxlen=20)
data_test = pad_sequences(sequences_test, maxlen=20)
# pred = np.array(data.sentiment.values)
print('Shape of data tensor:', data_train.shape)
print('Shape of label tensor:', data_test.shape)

nb_words = min(200000, len(word_index))+1

embedding_matrix = np.zeros((nb_words, 50))
for word, i in word_index.items():
    if word in model2vec.wv.vocab:
        embedding_matrix[i] = model2vec.wv[word]

data_train, y_train = shuffle(data_train, y_train, random_state=0)
data_test, y_test = shuffle(data_test, y_test, random_state=0)

model_input = Input(shape=(20,))

model = Embedding(nb_words, 50, weights=[embedding_matrix],input_length=20,
        trainable=False)(model_input)
model = Dropout(0.25)(model)

model = LSTM(20, dropout=0.35,recurrent_dropout=0.35)(model)
model = BatchNormalization()(model)

model = Dropout(0.25)(model)
model = Dense(32)(model)
model = Activation('relu')(model)
model = BatchNormalization()(model)

model = Dense(1)(model)
out = Activation('sigmoid')(model)

opt = SGD(lr = 0.1, momentum = 0.9, nesterov=True)

model = Model(inputs=model_input, outputs=out)

model.compile(loss='binary_crossentropy',
              optimizer=opt,
              metrics=['accuracy'])

model.fit(data_train, y_train,
          batch_size=128,
          epochs=5,
          verbose=1,
          validation_data=[data_test, y_test])

Epoch 1/5
68/68 [==============================] - 1s 22ms/step - loss: 0.5084 - accuracy: 0.7975 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 2/5
68/68 [==============================] - 1s 19ms/step - loss: 0.5109 - accuracy: 0.7976 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 3/5
68/68 [==============================] - 1s 19ms/step - loss: 0.5095 - accuracy: 0.7976 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 4/5
68/68 [==============================] - 1s 19ms/step - loss: 0.5084 - accuracy: 0.7975 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00
Epoch 5/5
68/68 [==============================] - 1s 20ms/step - loss: 0.5083 - accuracy: 0.7973 - val_loss: 0.0000e+00 - val_accuracy: 0.0000e+00

In addition I found that the metrics in validation is actually different

accuracy_score(y_test, [c[0]>.3 for c in model.predict(data_test)])
0.7661122661122661

I really do not understand why this issue occured. Any idea? Thank you

Answer 1

Fixed. I've just downgraded my tensorflow to 2.0. pip install tensorflow==2.0