Shapes (None, 1) and (None, 3) are incompatible, multi-class classification

Question

So I have multi-class classification. I want to compile my model:

feature_layer = DenseFeatures(feature_columns) # A layer that produces a dense Tensor
model = Sequential([

feature_layer,
  Dense(32, activation='relu'),
  Dense(3, activation='softmax')
])

So I use categorical_crossentropy loss:

model.compile(optimizer='adam',
              loss='categorical_crossentropy',
              metrics=['accuracy'])

model.fit(train_ds,
          validation_data=val_ds,
          epochs=10)

But, of course, it gives me an error. I know about to_categorical method, but it doesn't take BatchDataset as a parameter, which train_ds and val_ds are.

Please direct me

UPDATE: I tried to do something like this:

def df_to_dataset(df, shuffle=True, batch_size=32): 
  df = df.copy()
  labels = df.pop('class')
  ds = tf.data.Dataset.from_tensor_slices((dict(df), labels))
  if shuffle:
    ds = ds.shuffle(buffer_size=len(df))
    ds = ds.batch(batch_size).map(lambda x, y: (x, tf.one_hot(y, depth=3)))
  return ds

batch_size = 32
train_ds = df_to_dataset(train, batch_size=batch_size)  # error
val_ds = df_to_dataset(val, shuffle=False, batch_size=batch_size)
test_ds = df_to_dataset(test, shuffle=False, batch_size=batch_size)

And it gives me:

Value passed to parameter 'indices' has DataType string not in list of allowed values: uint8, int32, int64

My class column has string values (it tells if the object is star, galaxy or quazar, others are int/float), but I popped it:

def df_to_dataset(df, shuffle=True, batch_size=32): 
  df = df.copy()
  labels = df.pop('class')
  ds = tf.data.Dataset.from_tensor_slices((dict(df), labels))
  if shuffle:
    ds = ds.shuffle(buffer_size=len(df))
  return ds
df_to_dataset(df)

labels:

<ShuffleDataset shapes: ({objid: (), ra: (), dec: (), u: (), g: (), r: (), i: (), z: (), run: (), rerun: (), camcol: (), field: (), specobjid: (), redshift: (), plate: (), mjd: (), fiberid: ()}, ()), types: ({objid: tf.float64, ra: tf.float64, dec: tf.float64, u: tf.float64, g: tf.float64, r: tf.float64, i: tf.float64, z: tf.float64, run: tf.int64, rerun: tf.int64, camcol: tf.int64, field: tf.int64, specobjid: tf.float64, redshift: tf.float64, plate: tf.int64, mjd: tf.int64, fiberid: tf.int64}, tf.string)>

Answer 1

You can either convert your labels to one-hot encoded labels and use the categorical_crossentropy loss function:

one_hot_encoded_train_ds = train_ds.map(lambda x, y: (x, tf.one_hot(y, depth=3)))
one_hot_encoded_val_ds = val_ds.map(lambda x, y: (x, tf.one_hot(y, depth=3)))

Or change your loss function to sparse_categorical_crossentropy and leave your labels the way they are as integers:

model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

Update: Convert your string labels to integers:

def df_to_dataset(df, shuffle=True, batch_size=32): 
  df = df.copy()
  labels = df.pop('class')
  dicts = {'STAR': 0, 'GALAXY': 1, 'QSO': 2}
  converted_labels = np.array([dicts[l] for l in labels.to_list()])
  ds = tf.data.Dataset.from_tensor_slices((dict(df), converted_labels))
  if shuffle:
    ds = ds.shuffle(buffer_size=len(df))
  return ds
df_to_dataset(df)