Validation accuracy jumps badly between 0 and 1

Question

Issue:

My validation accuracy jumps between 0% and 100%.
This seems fairly implausible to me because the predictions are between two classes only and I am validating on full valSet (41802 records) each epoch. I doubt that a model which just started learning gets 100% accuracy on one epoch and then 0% accuracy the next epoch.

Code samples:

I am validating on the full valSet per epoch. This is achieved by using validation_steps = {steps for full validation file} and dataset = dataset.repeat(numEpochsTrainSet)

all my datasets are produced with the following code

def modelDataset(sourcepath, badgesize, repeat = False, repeatCount = 15):
    #get all files
    list = os.listdir(sourcepath)
    pathfiles = [sourcepath+x for x in list]

    #get metrics
    rows_per_file = count_lines(sourcepath+"0.data")
    number_of_files = len(list)
    total_rows = (rows_per_file * number_of_files)
    # get number of steps per Epoch
    epochs = number_of_files
    if  epochs == 1 and repeat:
        epochs = repeatCount
    elif epochs == 1:
        badgesize = rows_per_file
    steps_per_epoch = int(rows_per_file / badgesize) # 2000 badges per epoch   
    # model interleaved dataset
    dataset = (tf.data.Dataset.from_tensor_slices(pathfiles).interleave(lambda x:
        tf.data.TextLineDataset(x).map(parse_csv, num_parallel_calls=4),
        cycle_length=4, block_length=counter+1))
    if epochs != 1 or epochs != repeatCount:
        dataset = dataset.shuffle(buffer_size=badgesize)
    if repeat:
        dataset = dataset.repeat(repeatCount)
        #dataset = dataset.cache().repeat(repeatCount)
    dataset = dataset.batch(badgesize)
    dataset = dataset.prefetch(2)  # prefetch one batch
    return dataset, steps_per_epoch, epochs, badgesize

# generate dataset
valSet, valSteps, maxValEpochs, valBadgeSize = modelDataset(valPath, 1, True, maxTrainEpochs) 
# check if dataset is modeled correctly
print(f"number valSteps {valSteps}, maxepochs {maxValEpochs}, badgesize {valBadgeSize}") 
for elem in valSet:
  print(elem) 

The output looks fine to me:

finally, the model is fit like the following:

model.fit(trainSet, 
    epochs = epochs, 
    steps_per_epoch = trainSteps, 
    shuffle = True, 
    validation_data = valSet, 
    validation_steps = valSteps, # how many batches do you want to validate on, NOT how many records
    class_weight=classDistr,
    verbose = verbose, 
    callbacks = [tensorboard,reduce_lr])#,cp_callback])

Issue assumptions

I would expect such validation results in one of the following scenarios:

• I am only validating on one record per epoch
• The validation Data is inconsistent with the train data
• The validation data contains labels of all the same class

I already tried to exclude those assumption with the following steps:

• I checked the source data csv file which looks fine, it has 41802 records
• In between those val_accuracies there are some which suggest against a validation on only one record:
• output of print(valset[0]) looks fine
• I found out that val-batchsize = train-batchsize so I adjusted te code but it did not make any difference