We used the class_weight parameter to assign different weights for samples of different classes in keras.Model.fit().
Our data (input) and labels (output) have an extra time-step dimension besides the batch dimension.
The labels are indices of classes (i.e. not one-hot encoded), used along with the sparse CE loss function.
We've written a minimal reproduction script (as presented below) to simplify the situation.

It looks like that class_weight is only designed for outputs of shape (batch_dim, n_classes).

Possible workarounds:

1. convert class_weight to sample_weight
2. collapse time-step axis into batch axis using tf.reshape() in a Lambda layer, which would mess the code up

I've noticed there are relevant issues but they are left there and closed as the authors did not reply in time.

System information

• Have I written custom code: Yes
• OS Platform and Distribution: Linux Ubuntu 16.04
• TensorFlow installed from (source or binary): Anaconda, binary
• TensorFlow version (use command below): 2.2.0
• Python version: 3.8.5
• CUDA/cuDNN version: cudnn 7.6.5 cuda10.1_0.conda (irrelevant)
• GPU model and memory: Tesla K40c, 11441MiB (irrelevant)

Describe the current behavior
Crashed on model.fit with the following exception (traceback appended at the end as it is too long):

InvalidArgumentError: indices[0] = 3 is not in [0, 2)
[[{{node GatherV2}}]] [Op:IteratorGetNext]

The training progress bar did not appear.

Describe the expected behavior
Complete the training, though the minimal reproduction script has no actual trainable parameter.

Standalone code to reproduce the issue

import numpy
import tensorflow as tf
from tensorflow import keras

def test_class_weight_error():
    # the model simply return the input time_step*n_class=20x2 data as-is
    model = keras.Sequential([keras.layers.Reshape((20, 2), input_shape=(20, 2))])
    # run_eagerly improves the readability of the traceback a bit
    model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', run_eagerly=True)
    model.summary()
    # X (inputs, as well as y_pred): samples*time_step*n_classes=15x20x2
    xs = tf.reshape(tf.one_hot(tf.ones(300, dtype=tf.int32), 2), [-1, 20, 2]).numpy()
    # Y (labels i.e. y_true): samples*time_step=15x20, class labels of 0 or 1
    ys = np.ones([15, 20], dtype=np.int32)
    # without the line below (a.k.a. all labels being 1) there's no exception
    ys[:,:3] = 0
    # here's the crash
    model.fit(xs, ys, batch_size=3, class_weight={0:1.,1:1.})
    
test_class_weight_error()

Other info / logs

Traceback:
traceback.txt

Model Structure (in minimal repro script):

Model: "sequential_64"
Layer (type) Output Shape Param
reshape_64 (Reshape) (None, 20, 2) 0
Total params: 0
Trainable params: 0
Non-trainable params: 0

---