I am trying to find out the reason behind why my RNN network won't go beyond 50% for binary classification. My input data is of the shape:

- TensorShape([9585, 25, 2])

My labels are a single dimension vector with the values 1.0 and 0.0:

- <tf.Tensor: shape=(9585,), dtype=float32, numpy=array([1., 0., 1., ..., 0., 0., 1.], dtype=float32)>

I have created the classification class as below:

batch_size = 4 # hyperparameter
max_seqlen = 25 # the second dimension (time) in the data, first being the number of datapoints
features = 2 # 2 features per timestamp

class Model(tf.keras.Model):
    def __init__(self, max_seqlen, **kwargs):
        super(Model, self).__init__(**kwargs)
        self.bilstm = tf.keras.layers.Bidirectional(
            tf.keras.layers.LSTM(128, return_sequences=False, input_shape=(max_seqlen, features))
        self.dense = tf.keras.layers.Dense(50, activation="relu") # nn with relu non-linearity
        self.out = tf.keras.layers.Dense(1, activation="sigmoid") # for final binary prediction
    def call(self, x):
        x = self.bilstm(x)
        x = self.dense(x)
        x = self.out(x)
        return x
model = Model(max_seqlen)
model.build(input_shape=(batch_size, max_seqlen, features))

I am preparing the dataset and running the training and validation as follows:

dataset = tf.data.Dataset.from_tensor_slices((X, y))
dataset = dataset.shuffle(10000)
n = len(y)
test_size = n // 8
val_size = (n - test_size) // 10

test_dataset = dataset.take(test_size)
val_dataset = dataset.skip(test_size).take(val_size)
train_dataset = dataset.skip(test_size + val_size)
train_dataset = train_dataset.batch(batch_size)
val_dataset = val_dataset.batch(batch_size)
test_dataset = test_dataset.batch(batch_size)


# train
data_dir = "./data"
logs_dir = os.path.join("./logs")
best_model_file = os.path.join(data_dir, "best_model.h5")

checkpoint = tf.keras.callbacks.ModelCheckpoint(best_model_file,
tensorboard = tf.keras.callbacks.TensorBoard(log_dir=logs_dir)
num_epochs = 10
history = model.fit(train_dataset, epochs=num_epochs,
    callbacks=[checkpoint, tensorboard])

During the epochs, the accuracy does not improve beyond 50%. Is there something wrong with what I am doing? I also tried normalizing my dataset.


2 Answers 2


To split your dataset try this :

train_ds = dataset.take(train_size)    
val_ds = dataset.skip(train_size).take(val_size)
test_ds = dataset.skip(train_size).skip(val_size)
  • $\begingroup$ Don't see the point. I am making test dataset, skipping and taking validation dataset and then skipping both test and validation datasets. $\endgroup$ Nov 9, 2023 at 21:48
  • $\begingroup$ Using your way of splitting datasets, found some samples appearing in both test and val, or other way around. Anyway it is great if you found solution. $\endgroup$
    – aRedDish
    Nov 10, 2023 at 21:46

Found the issue. The y were not correct to the corresponding X.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.