How to configure dataset pipelines with Tensorflow make_csv_dataset for Keras Model

问题

I have a structured dataset(csv features files) of around 200 GB. I'm using make_csv_dataset to make the input pipelines. Here is my code

def pack_features_vector(features, labels):
    """Pack the features into a single array."""
    features = tf.stack(list(features.values()), axis=1)
    return features, labels
def main():    
    defaults=[float()]*len(selected_columns)
    data_set=tf.data.experimental.make_csv_dataset(
        file_pattern = "./../path-to-dataset/Train_DS/*/*.csv",
        column_names=all_columns,    # all_columns=["col1,col2,..."]
        select_columns=selected_columns,   # selected_columns= a subset of all_columns
        column_defaults=defaults,
        label_name="Target",
        batch_size=1000, 
        num_epochs=20,
        num_parallel_reads=50,
    #    shuffle_buffer_size=10000,
        ignore_errors=True)

    data_set = data_set.map(pack_features_vector)

    N_VALIDATION = int(1e3)
    N_TRAIN= int(1e4)
    BUFFER_SIZE = int(1e4)
    BATCH_SIZE = 1000
    STEPS_PER_EPOCH = N_TRAIN//BATCH_SIZE

    validate_ds = data_set.take(N_VALIDATION).cache().repeat()
    train_ds = data_set.skip(N_VALIDATION).take(N_TRAIN).cache().repeat()

    # validate_ds = validate_ds.batch(BATCH_SIZE)
    # train_ds = train_ds.batch(BATCH_SIZE)

    model = tf.keras.Sequential([
    layers.Flatten(),
    layers.Dense(256, activation='elu'),
    layers.Dense(256, activation='elu'),
    layers.Dense(128, activation='elu'),  
    layers.Dense(64, activation='elu'), 
    layers.Dense(32, activation='elu'), 
    layers.Dense(1,activation='sigmoid') 
    ])
    model.compile(optimizer='adam',
                loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
                metrics=['accuracy'])    
    model.fit(train_ds,
            validation_data=validate_ds,
            validation_steps=1,
            steps_per_epoch= 1,
            epochs=20,
            verbose=1
            )
if __name__ == "__main__":
    main()

print('Training completed!')

Now, when I execute this code , it's completed within few minutes (I think not going through the whole training data) with the following warnings:

W tensorflow/core/kernels/data/cache_dataset_ops.cc:798] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.

As per this warning and as training is completed in few minutes meaning that... input pipeline is not configured correctly... Can anyone please guide me, how to correct this problem.

GPU of my system is NVIDIA Quadro RTX 6000 (compute capability 7.5).

A solution based on some other function like experimental.CsvDataset would work as well.

Edit

That warning gone by changing the code to avoid any cache as

    validate_ds = data_set.take(N_VALIDATION).repeat()
    train_ds = data_set.skip(N_VALIDATION).take(N_TRAIN).repeat()

But now the problem is I'm getting zero accuracy, even on the training data. Which I think is a problem of input pipelines. Here is the output.

Edit2

After some efforts, I managed to resolve the known issues by using a bit lower level but similar API, CsvDataset. But now, I'm getting the accuracy=1.00 which I think is not OK. At first epoch, it's .95 and then for next 19 epochs, it's 1.00. Here is my final code.

def preprocess(*fields):
    features=tf.stack(fields[:-1])
    # convert Target column values to int to make it work for binary classification
    labels=tf.stack([int(x) for x in fields[-1:]])
    return features,labels  # x, y


def main():
    # selected_columns=["col1,col2,..."]
    selected_indices=[]
    for selected_column in selected_columns:
        index=all_columns.index(selected_column)
        selected_indices.append(index)
        
    print("All_columns length"+str(len(all_columns)))
    print("selected_columns length"+str(len(selected_columns)))
    print("selected_indices length"+str(len(selected_indices)))
    print(selected_indices)
    defaults=[float()]*(len(selected_columns))
    #defaults.append(int())
    print("defaults"+str(defaults))
    print("defaults length"+str(len(defaults)))
    FEATURES = len(selected_columns) - 1
    training_csvs =  sorted(str(p) for p in pathlib.Path('.').glob("path-to-data/Train_DS/*/*.csv"))
    testing_csvs =  sorted(str(p) for p in pathlib.Path('.').glob("path-to-data/Test_DS/*/*.csv"))

    training_csvs
    testing_csvs

    training_dataset=tf.data.experimental.CsvDataset(        
        training_csvs,
        record_defaults=defaults, 
        compression_type=None, 
        buffer_size=None,
        header=True, 
        field_delim=',',
        # use_quote_delim=True,
        # na_value="",
        select_cols=selected_indices
        )
    
    print(type(training_dataset))
    for features in training_dataset.take(1):
        print("Training samples before mapping")
        print(features)
    
    validate_ds = training_dataset.map(preprocess).take(10).batch(100).repeat()
    train_ds = training_dataset.map(preprocess).skip(10).take(90).batch(100).repeat()
    validate_ds
    train_ds
    for features,labels in train_ds.take(1):
        print("Training samples")
        print(features)
        print(labels)
    
    testing_dataset=tf.data.experimental.CsvDataset(        
        testing_csvs,
        record_defaults=defaults, 
        compression_type=None, 
        buffer_size=None,
        header=True, 
        field_delim=',',
        use_quote_delim=True,
        na_value="",
        select_cols=selected_indices
        )
    
    print(type(testing_dataset))
    test_ds = testing_dataset.map(preprocess).batch(100).repeat()
    test_ds
    for features,labels in test_ds.take(1):
        print("Testing samples")
        print(features)
        print(labels)
    
    model = tf.keras.Sequential([        
        layers.Dense(256,activation='elu'),  
        layers.Dense(128,activation='elu'),  
        layers.Dense(64,activation='elu'),  
        layers.Dense(1,activation='sigmoid') 
        ])
    history = model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(from_logits=False),
                            metrics=['accuracy'])
    
    model.fit(train_ds,
        validation_data=validate_ds,
        validation_steps=20,
        steps_per_epoch= 20,
        epochs=20,
        verbose=1
        )
    
    loss, accuracy = model.evaluate(test_ds)
    print("Test Accuracy", accuracy)


if __name__ == "__main__":
    main()

print('Training completed!')

I tried to feed just the few useless features to the model, but still, it's giving accuracy=1.00 or 100 %. Which is going wrong now? Overfitting etc?

回答1:

In the snippets, you wrote

model.fit(train_ds,
          validation_data=validate_ds,
          validation_steps=1,
          steps_per_epoch= 1,
          epochs=20,
          verbose=1)

Is the steps_per_epoch= 1 a typo? If not, that would mean you only use one batch per training, which explains the fast training and the low accuracy. validation_steps=1 is also an issue

来源：https://stackoverflow.com/questions/64725275/how-to-configure-dataset-pipelines-with-tensorflow-make-csv-dataset-for-keras-mo