Python no puede aplicar fit_generator al modelo de keras con entrada múltiple

Question

Tengo el siguiente modelo: este es LSTM + CNN con 3 entradas.

Y construí esta función de generador para entrenar el modelo usando fit_generator (basado en esto: https://stanford.edu/~shervine/blog/keras-how-to-generate-data-on-the-fly):

class MultiInputDataGenerator(keras.utils.Sequence):
    'Generates data for Keras'

    def __init__(self, list_IDs, labels, shuffle=True):
        'Initialization'
        self.batch_size = 8
        self.labels = labels
        self.list_IDs = list_IDs
        self.n_classes = 5
        self.shuffle = shuffle
        self.on_epoch_end()

def __len__(self):
    'Denotes the number of batches per epoch'
    return int(np.floor(len(self.list_IDs) / self.batch_size))

def __getitem__(self, index):
    'Generate one batch of data'
    # Generate indexes of the batch
    indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]

    # Find list of IDs
    list_IDs_temp = [self.list_IDs[k] for k in indexes]

    # Generate data
    X, y = self.__data_generation(list_IDs_temp)

    return X, y

def on_epoch_end(self):
    'Updates indexes after each epoch'
    self.indexes = np.arange(len(self.list_IDs))
    if self.shuffle == True:
        np.random.shuffle(self.indexes)

def __data_generation(self, list_IDs_temp):
    'Generates data containing batch_size samples' # X : (n_samples, *dim, n_channels)
    # Initialization
    
    X = np.empty((self.batch_size, 1, 3), dtype=object)
    y = np.empty((self.batch_size), dtype=object)

    # Generate data
    for i, ID in enumerate(list_IDs_temp):
        X_id = []
        x_features = df.iloc[id][et_cols].values #ET_COLS are 14 columns so I get 1X14 here
        x_text = df.iloc[id].text_col #x_text is 1X768
        x_vid = df.iloc[id].frame_col #x_vid is (3,244,244)
        
        X_id.append(x_features)
        X_id.append(x_text) 
        X_id.append(x_vid)
           
        X[i,] = X_id
        y[i] = self.labels[ID]

    y_mat = tf.convert_to_tensor(pd.get_dummies(y))
    return X, y_mat


training_generator = MultiModelDataGenerator(generator_partition['train'], generator_labels)
validation_generator = MultiModelDataGenerator(generator_partition['val'], generator_labels)
net = build_LSTMCNN_net()
net.compile(keras.optimizers.Adam(0.001),'categorical_crossentropy',metrics=['acc'])
net.fit_generator(generator=training_generator,
                    validation_data=validation_generator,)
                    use_multiprocessing=True)#,    workers=6)

Y me sale el error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-38-669153f703e6> in <module>()
      
      net.fit_generator(generator=training_generator,
--->                      validation_data=validation_generator,)
                          #use_multiprocessing=True)#,    workers=6)

/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/constant_op.py in convert_to_eager_tensor(value, ctx, dtype)
     96       dtype = dtypes.as_dtype(dtype).as_datatype_enum
     97   ctx.ensure_initialized()
---> 98   return ops.EagerTensor(value, ctx.device_name, dtype)
     99 
    100 

ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type numpy.ndarray).

También probé varias variaciones, como agregar:

x_features = np.asarray(x_features).astype(object)
x_text = np.asarray(x_text).astype(object)
x_vid = np.asarray(x_text).astype(object)

O en X[i,] = [X_id]lugar de X[i,] = X_idPero ninguno funcionó. ¿Alguna idea de cómo solucionar el problema?

Editado: al agregar:

astype(np.float32) 

y
tf.convert_to_tensor(X)

Recibo el error: ValueError Traceback (última llamada más reciente) en ()

      net.fit_generator(generator=training_generator,
--->                      validation_data=validation_generator,
                          use_multiprocessing=True,    workers=6)


/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/constant_op.py in convert_to_eager_tensor(value, ctx, dtype)
            dtype = dtypes.as_dtype(dtype).as_datatype_enum
        ctx.ensure_initialized()
--->    return ops.EagerTensor(value, ctx.device_name, dtype)
  

Answer 1

Antes de resolver el problema, resumamos primero el conjunto de datos con el que está trabajando. Basándome en tu descripción, creé un ejemplo DataFrameque podría parecerse al tuyo

import pandas as pd

dataset_size = 500
train_idx,val_idx = train_test_split(range(dataset_size),test_size=0.2,) 

# create an example DataFrame that I assume will be resemble yours 
example_df = pd.DataFrame({'vids':np.random.randint(0,10000,dataset_size)})
# create feature columns 
for ind in range(14): example_df['feature_%i' % ind] = np.random.rand(dataset_size)
# each cell contains a list 
example_df['text'] = np.random.randint(dataset_size)
example_df['text'] = example_df['text'].astype('object')
for ind in range(dataset_size):example_df.at[ind,'text'] = np.random.rand(768).tolist()
# create the label column
example_df['label'] = np.random.randint(low=0,high=5,size=dataset_size)

# extract information from the dataframe, and create data generators 
all_vids = example_df['vids'].values
feature_columns = ['feature_%i' % ind for ind in range(14)]
all_features = example_df[feature_columns].values
all_text = example_df['text'].values
all_labels = example_df['label'].values

Como puede ver, la columna textes una columna de listas, en la que cada lista contiene 768 elementos. La columna labelscontiene las etiquetas de los ejemplos, no importa si usa codificación one-hot u otros tipos de codificación, siempre que su forma coincida con la forma de la capa de salida del modelo de red neuronal general. La columna vidses una columna de seeds para generar imágenes aleatorias sobre la marcha.

---

Resolver el problema (basado en el conjunto de datos anterior)

Puede utilizar esta sintaxis return {'feature':features,'text':text,'vid':vid},ypara el método __getitem__, en lugar de apilar las tres matrices de entrada.

Para explicar esto, primero construyamos un modelo de juguete que se parezca al tuyo

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input,Dense,Flatten,Add


def features_part(x):
    y = Dense(14)(x)
    y = Dense(10,activation='linear')(y)
    return y

def text_part(x):
    y = Dense(768)(x)
    y = Dense(10,activation='linear')(y)
    return y

def vid_part(x):
    y = Flatten()(x)
    y = Dense(10,activation='linear')(y)
    return y

input_features = Input(shape=(14,),name='feature')
input_text = Input(shape=(768,),name='text')
input_vid = Input(shape=(3,244,244,),name='vid')

feature_block = features_part(input_features)
text_block = text_part(input_text)
vid_block = vid_part(input_vid)
added = Add()([feature_block,text_block,vid_block])
# you have five classes at the end of the day 
pred = Dense(1)(added)
# build model
model = Model(inputs=[input_features,input_text,input_vid],outputs=pred)
model.compile(loss='mae',optimizer='adam',metrics=['mae'])

Lo más importante de este modelo es que especifiqué los nombres de las tres capas de entrada

input_features = Input(shape=(14,),name='feature')
input_text = Input(shape=(768,),name='text')
input_vid = Input(shape=(3,244,244,),name='vid')

Para este modelo, puede construir un generador como

# provide a seed for generating a random image 
def fn2img(seed):
    np.random.seed(seed)
    # fake an image with three channels 
    return np.random.randint(low=0,high=255,size=(3,244,244))


class MultiInputDataGenerator(keras.utils.Sequence):

    def __init__(self, 
                 all_inds,labels, 
                 features,text,vid, 
                 shuffle=True):
        self.batch_size = 8
        self.labels = labels
        self.all_inds = all_inds
        self.shuffle = shuffle
        self.on_epoch_end()
        
        self.features = features
        self.text = text
        self.vid = vid

    def __len__(self): 
        return int(np.floor(len(self.all_inds) / self.batch_size))


    def __getitem__(self,index):
        indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]
        batch_indices = [self.all_inds[k] for k in indexes]
        features,text,vid,y = self.__data_generation(batch_indices)

        return {'feature':features,'text':text,'vid':vid},y

    def on_epoch_end(self):
        self.indexes = np.arange(len(self.all_inds))
        if self.shuffle == True:
            np.random.shuffle(self.indexes)

    def __data_generation(self,batch_indices):
        # Generate data
        features = self.features[batch_indices,:]
        # note that you need to stack the slice in order to reshape it to (num_samples,768)
        text = np.stack(self.text[batch_indices])
        # since batch_size is not a super large number, you can stack here
        vid = np.stack([fn2img(seed) for seed in self.vid[batch_indices]])
        y = self.labels[batch_indices]

        return features,text,vid,y

como puede ver, el __getitem__método devuelve un diccionario {'feature':features,'text':text,'vid':vid},y. Las claves del diccionario coinciden con los nombres de las tres capas de entrada. Además, las imágenes aleatorias se generan sobre la marcha.

Para asegurarse de que todo funcione, puede ejecutar el siguiente script,

import numpy as np
import pandas as pd
from tensorflow import keras 
from sklearn.model_selection import train_test_split

from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input,Dense,Flatten,Add


# provide a seed for generating a random image
def fn2img(seed):
    np.random.seed(seed)
    # fake an image with three channels
    return np.random.randint(low=0,high=255,size=(3,244,244))


class MultiInputDataGenerator(keras.utils.Sequence):

    def __init__(self,
                 all_inds,labels,
                 features,text,vid,
                 shuffle=True):
        self.batch_size = 8
        self.labels = labels
        self.all_inds = all_inds
        self.shuffle = shuffle
        self.on_epoch_end()
        
        self.features = features
        self.text = text
        self.vid = vid

    def __len__(self):
        return int(np.floor(len(self.all_inds) / self.batch_size))


    def __getitem__(self,index):
        indexes = self.indexes[index*self.batch_size:(index+1)*self.batch_size]
        batch_indices = [self.all_inds[k] for k in indexes]
        features,text,vid,y = self.__data_generation(batch_indices)

        return {'feature':features,'text':text,'vid':vid},y

    def on_epoch_end(self):
        self.indexes = np.arange(len(self.all_inds))
        if self.shuffle == True:
            np.random.shuffle(self.indexes)

    def __data_generation(self,batch_indices):
        # Generate data
        features = self.features[batch_indices,:]
        # note that you need to stack the slice in order to reshape it to (num_samples,768)
        text = np.stack(self.text[batch_indices])
        # since batch_size is not a super large number, you can stack here
        vid = np.stack([fn2img(seed) for seed in self.vid[batch_indices]])
        y = self.labels[batch_indices]

        return features,text,vid,y


# fake a dataset
dataset_size = 500
train_idx,val_idx = train_test_split(range(dataset_size),test_size=0.2,)

# create an example DataFrame that I assume will be resemble yours
example_df = pd.DataFrame({'vids':np.random.randint(0,10000,dataset_size)})
# create feature columns
for ind in range(14): example_df['feature_%i' % ind] = np.random.rand(dataset_size)
# each cell contains a list
example_df['text'] = np.random.randint(dataset_size)
example_df['text'] = example_df['text'].astype('object')
for ind in range(dataset_size):example_df.at[ind,'text'] = np.random.rand(768).tolist()
# create the label column
example_df['label'] = np.random.randint(low=0,high=5,size=dataset_size)

# extract information from the dataframe, and create data generators
all_vids = example_df['vids'].values
feature_columns = ['feature_%i' % ind for ind in range(14)]
all_features = example_df[feature_columns].values
all_text = example_df['text'].values
all_labels = example_df['label'].values

training_generator = MultiInputDataGenerator(train_idx,all_labels,all_features,all_text,all_vids)

# create model
def features_part(x):
    y = Dense(14)(x)
    y = Dense(10,activation='linear')(y)
    return y

def text_part(x):
    y = Dense(768)(x)
    y = Dense(10,activation='linear')(y)
    return y

def vid_part(x):
    y = Flatten()(x)
    y = Dense(10,activation='linear')(y)
    return y

input_features = Input(shape=(14,),name='feature')
input_text = Input(shape=(768,),name='text')
input_vid = Input(shape=(3,244,244,),name='vid')

feature_block = features_part(input_features)
text_block = text_part(input_text)
vid_block = vid_part(input_vid)
added = Add()([feature_block,text_block,vid_block])
# you have five classes at the end of the day 
pred = Dense(1)(added)
# build model
model = Model(inputs=[input_features,input_text,input_vid],outputs=pred)
model.compile(loss='mae',optimizer='adam',metrics=['mae'])

model.fit_generator(generator=training_generator,epochs=10)

print(model.history.history)