How to solve ValueError in LSTM input?

Question

I have the training and testing numpy arrays in the following shapes

TrainX = (1234, 50, 50) Type: <class 'numpy.ndarray'> # 1234 arrays of 50 by 50 floats
TrainY = (1234, 2) Type: <class 'numpy.ndarray'>
# TrainY was one column of binary class 0 or 1. Converted it through to_categorical()

TestX = (123, 50, 50) Type: <class 'numpy.ndarray'>
TestY = (123, 2) Type: <class 'numpy.ndarray'>

I use the following code for the LSTM,

from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import LSTM, Dense, Dropout

model = Sequential()
model.add(LSTM(50, input_shape=(TrainX.shape[1], TrainX.shape[2])))
model.add(Dense(50))
model.add(Dropout(0.3))
model.add(Dense(2, activation="softmax"))

model.compile("adam", "categorical_crossentropy", metrics=["accuracy"])

model.fit(
    TrainX,
    TrainY,
    batch_size=24,
    epochs=48,
    validation_data=[np.asarray(TestX).all(), np.asarray(TestY).all()],
    class_weight=classweights,#calculated class weights
    verbose=2,
)

1. First I used TrainX and TrainY but got an error ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

2. Based on this Keras LSTM input ValueError: Shapes are incompatible, I used np.asarray(TrainX). But got the same error.

3. So I added np.asarray(TrainX).all() one time and np.asarray(TrainX).any() another time. But got a different value error: ValueError: Failed to find data adapter that can handle input: <class 'numpy.bool_'>, <class 'numpy.bool_'>

4. Finally, I tried pandas.DataFrame(TrainX) for input. But it showed the following error, ValueError: Must pass 2-d input. shape=(1234, 50, 50)

trace back of the errors 1 and 2:

Traceback (most recent call last):

  File "C:\Users\...\main.py", line 73, in <module>
    lmodel.lstm(TrainX, TestX, TrainY, TestY)
  File "C:\Users\...\llmodel.py", line 65, in lstm
    model.fit(
  File "C:\Users\...\AppData\Roaming\Python\Python310\site-packages\tensorflow\python\keras\engine\training.py", line 1137, in fit
    data_handler = data_adapter.get_data_handler(
  File "C:\Users\...\AppData\Roaming\Python\Python310\site-packages\tensorflow\python\keras\engine\data_adapter.py", line 1394, in get_data_handler
    return DataHandler(*args, **kwargs)
  File "C:\Users\...\AppData\Roaming\Python\Python310\site-packages\tensorflow\python\keras\engine\data_adapter.py", line 1169, in __init__
    self._configure_dataset_and_inferred_steps(strategy, x, steps_per_epoch,
  File "C:\Users\...\AppData\Roaming\Python\Python310\site-packages\tensorflow\python\keras\engine\data_adapter.py", line 1177, in _configure_dataset_and_inferred_steps
    if class_weight:
ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

How can this issue be solved?

Answer 1

The log suggests that the problem is in your class_weight argument. According to the docs, this argument should be

Optional dictionary mapping class indices (integers) to a weight (float) value, used for weighting the loss function (during training only). This can be useful to tell the model to "pay more attention" to samples from an under-represented class.

which should be something like

class_weights = {0: 0.75, 1: 1.25}

if you want to assign a weight of 0.75 and 1.25 to class 0 and 1, respectively.

From the log, I suspect that instead of passing such a dictionary, you passed a numpy array of shape (n_samples, n_classes) as class_weight, which is inappropriate.