AttributeError: 'tuple' object has no attribute 'rank' when calling fit on a Keras model with custom generator
Question:
I want to build a Neural Network with two inputs: for image data and for numeric data. So I wrote custom data generator for that. The train and validation dataframes contain 11 columns:
image_name — path to the image;- 9 numeric features;
target — class for the item (last column).
The code for custom generator (based on this answer):
target_size = (224, 224)
batch_size = 1
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True)
val_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_dataframe(
train,
x_col='image_name',
y_col=train.columns[1:],
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='raw')
validation_generator = val_datagen.flow_from_dataframe(
validation,
x_col='image_name',
y_col=validation.columns[1:],
target_size=target_size,
shuffle=False,
batch_size=batch_size,
class_mode='raw')
def train_generator_func():
count = 0
while True:
if count == len(train.index):
train_generator.reset()
break
count += 1
data = train_generator.next()
imgs = []
cols = []
targets = []
for k in range(batch_size):
imgs.append(data[0][k])
cols.append(data[1][k][:-1])
targets.append(data[1][k][-1])
yield [imgs, cols], targets
def validation_generator_func():
count = 0
while True:
if count == len(validation.index):
validation_generator.reset()
break
count += 1
data = validation_generator.next()
imgs = []
cols = []
targets = []
for k in range(batch_size):
imgs.append(data[0][k])
cols.append(data[1][k][:-1])
targets.append(data[1][k][-1])
yield [imgs, cols], targets
Model building:
def mlp_model(dim):
model = Sequential()
model.add(Dense(8, input_dim=dim, activation="relu"))
model.add(Dense(4, activation="relu"))
return model
def vgg16_model():
model = VGG16(weights='imagenet', include_top=False, input_shape=target_size+(3,))
x=Flatten()(model.output)
output=Dense(1,activation='sigmoid')(x) # because we have to predict the AUC
model=Model(model.input,output)
return model
def concatenated_model(cnn, mlp):
combinedInput = concatenate([cnn.output, mlp.output])
x = Dense(4, activation="relu")(combinedInput)
x = Dense(1, activation="sigmoid")(x)
model = Model(inputs=[cnn.input, mlp.input], outputs=x)
return model
def focal_loss(alpha=0.25,gamma=2.0):
def focal_crossentropy(y_true, y_pred):
bce = K.binary_crossentropy(y_true, y_pred)
y_pred = K.clip(y_pred, K.epsilon(), 1.- K.epsilon())
p_t = (y_true*y_pred) + ((1-y_true)*(1-y_pred))
alpha_factor = 1
modulating_factor = 1
alpha_factor = y_true*alpha + ((1-alpha)*(1-y_true))
modulating_factor = K.pow((1-p_t), gamma)
# compute the final loss and return
return K.mean(alpha_factor*modulating_factor*bce, axis=-1)
return focal_crossentropy
cnn = vgg16_model()
mlp = mlp_model(9)
model = concatenated_model(cnn, mlp)
opt = Adam(lr=1e-5)
model.compile(loss=focal_loss(), metrics=[tf.keras.metrics.AUC()],optimizer=opt)
nb_epochs = 2
nb_train_steps = train.shape[0]//batch_size
nb_val_steps = validation.shape[0]//batch_size
model.fit(
train_generator_func(),
steps_per_epoch=nb_train_steps,
epochs=nb_epochs,
validation_data=validation_generator_func(),
validation_steps=nb_val_steps)
And fitting doesn’t work with error message:
AttributeError Traceback (most recent call last)
<ipython-input-53-253849fd34d6> in <module>
9 epochs=nb_epochs,
10 validation_data=validation_generator_func(),
---> 11 validation_steps=nb_val_steps)
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginetraining.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginetraining.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing)
1061 use_multiprocessing=use_multiprocessing,
1062 model=self,
-> 1063 steps_per_execution=self._steps_per_execution)
1064
1065 # Container that configures and calls `tf.keras.Callback`s.
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in __init__(self, x, y, sample_weight, batch_size, steps_per_epoch, initial_epoch, epochs, shuffle, class_weight, max_queue_size, workers, use_multiprocessing, model, steps_per_execution)
1108 use_multiprocessing=use_multiprocessing,
1109 distribution_strategy=ds_context.get_strategy(),
-> 1110 model=model)
1111
1112 strategy = ds_context.get_strategy()
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in __init__(self, x, y, sample_weights, workers, use_multiprocessing, max_queue_size, model, **kwargs)
796 return tensor_shape.TensorShape([None for _ in shape.as_list()])
797
--> 798 output_shapes = nest.map_structure(_get_dynamic_shape, peek)
799 output_types = nest.map_structure(lambda t: t.dtype, peek)
800
d:pyenvkeras-gpulibsite-packagestensorflowpythonutilnest.py in map_structure(func, *structure, **kwargs)
633
634 return pack_sequence_as(
--> 635 structure[0], [func(*x) for x in entries],
636 expand_composites=expand_composites)
637
d:pyenvkeras-gpulibsite-packagestensorflowpythonutilnest.py in <listcomp>(.0)
633
634 return pack_sequence_as(
--> 635 structure[0], [func(*x) for x in entries],
636 expand_composites=expand_composites)
637
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in _get_dynamic_shape(t)
792 shape = t.shape
793 # Unknown number of dimensions, `as_list` cannot be called.
--> 794 if shape.rank is None:
795 return shape
796 return tensor_shape.TensorShape([None for _ in shape.as_list()])
AttributeError: 'tuple' object has no attribute 'rank'
So I tried to look at Keras sources but without any success.
If I use modified train_generator and validation_generator (y_col='target' instead of y_col=train.columns[1:]) everything works fine.
Answers:
You need to convert all the individual objects returned by both the training and validation generators to Numpy arrays:
yield [np.array(imgs), np.array(cols)], np.array(targets)
Alternatively, a simpler and much more efficient solution is to not iterate over the data batch at all; instead, we can take advantage of the fact that these objects are already Numpy arrays when returned by ImageDataGenerator, so we can write:
imgs = data[0]
cols = data[1][:,:-1]
targets = data[1][:,-1:]
yield [imgs, cols], targets
A different solution worked for me, just posting it here.
I ran into the problem working with two very similar dataframes in one notebook, where for one of them the error occurred.
I noticed the dtypes were slightly different int64 vs Int64, where the target column coded as Int64 gave the error.
For me the following worked:
dataframe[target_col] = dataframe[target_col].astype(int)
I want to build a Neural Network with two inputs: for image data and for numeric data. So I wrote custom data generator for that. The train and validation dataframes contain 11 columns:
image_name— path to the image;- 9 numeric features;
target— class for the item (last column).
The code for custom generator (based on this answer):
target_size = (224, 224)
batch_size = 1
train_datagen = ImageDataGenerator(
rescale=1./255,
rotation_range=20,
width_shift_range=0.2,
height_shift_range=0.2,
horizontal_flip=True)
val_datagen = ImageDataGenerator(rescale=1./255)
train_generator = train_datagen.flow_from_dataframe(
train,
x_col='image_name',
y_col=train.columns[1:],
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='raw')
validation_generator = val_datagen.flow_from_dataframe(
validation,
x_col='image_name',
y_col=validation.columns[1:],
target_size=target_size,
shuffle=False,
batch_size=batch_size,
class_mode='raw')
def train_generator_func():
count = 0
while True:
if count == len(train.index):
train_generator.reset()
break
count += 1
data = train_generator.next()
imgs = []
cols = []
targets = []
for k in range(batch_size):
imgs.append(data[0][k])
cols.append(data[1][k][:-1])
targets.append(data[1][k][-1])
yield [imgs, cols], targets
def validation_generator_func():
count = 0
while True:
if count == len(validation.index):
validation_generator.reset()
break
count += 1
data = validation_generator.next()
imgs = []
cols = []
targets = []
for k in range(batch_size):
imgs.append(data[0][k])
cols.append(data[1][k][:-1])
targets.append(data[1][k][-1])
yield [imgs, cols], targets
Model building:
def mlp_model(dim):
model = Sequential()
model.add(Dense(8, input_dim=dim, activation="relu"))
model.add(Dense(4, activation="relu"))
return model
def vgg16_model():
model = VGG16(weights='imagenet', include_top=False, input_shape=target_size+(3,))
x=Flatten()(model.output)
output=Dense(1,activation='sigmoid')(x) # because we have to predict the AUC
model=Model(model.input,output)
return model
def concatenated_model(cnn, mlp):
combinedInput = concatenate([cnn.output, mlp.output])
x = Dense(4, activation="relu")(combinedInput)
x = Dense(1, activation="sigmoid")(x)
model = Model(inputs=[cnn.input, mlp.input], outputs=x)
return model
def focal_loss(alpha=0.25,gamma=2.0):
def focal_crossentropy(y_true, y_pred):
bce = K.binary_crossentropy(y_true, y_pred)
y_pred = K.clip(y_pred, K.epsilon(), 1.- K.epsilon())
p_t = (y_true*y_pred) + ((1-y_true)*(1-y_pred))
alpha_factor = 1
modulating_factor = 1
alpha_factor = y_true*alpha + ((1-alpha)*(1-y_true))
modulating_factor = K.pow((1-p_t), gamma)
# compute the final loss and return
return K.mean(alpha_factor*modulating_factor*bce, axis=-1)
return focal_crossentropy
cnn = vgg16_model()
mlp = mlp_model(9)
model = concatenated_model(cnn, mlp)
opt = Adam(lr=1e-5)
model.compile(loss=focal_loss(), metrics=[tf.keras.metrics.AUC()],optimizer=opt)
nb_epochs = 2
nb_train_steps = train.shape[0]//batch_size
nb_val_steps = validation.shape[0]//batch_size
model.fit(
train_generator_func(),
steps_per_epoch=nb_train_steps,
epochs=nb_epochs,
validation_data=validation_generator_func(),
validation_steps=nb_val_steps)
And fitting doesn’t work with error message:
AttributeError Traceback (most recent call last)
<ipython-input-53-253849fd34d6> in <module>
9 epochs=nb_epochs,
10 validation_data=validation_generator_func(),
---> 11 validation_steps=nb_val_steps)
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginetraining.py in _method_wrapper(self, *args, **kwargs)
106 def _method_wrapper(self, *args, **kwargs):
107 if not self._in_multi_worker_mode(): # pylint: disable=protected-access
--> 108 return method(self, *args, **kwargs)
109
110 # Running inside `run_distribute_coordinator` already.
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginetraining.py in fit(self, x, y, batch_size, epochs, verbose, callbacks, validation_split, validation_data, shuffle, class_weight, sample_weight, initial_epoch, steps_per_epoch, validation_steps, validation_batch_size, validation_freq, max_queue_size, workers, use_multiprocessing)
1061 use_multiprocessing=use_multiprocessing,
1062 model=self,
-> 1063 steps_per_execution=self._steps_per_execution)
1064
1065 # Container that configures and calls `tf.keras.Callback`s.
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in __init__(self, x, y, sample_weight, batch_size, steps_per_epoch, initial_epoch, epochs, shuffle, class_weight, max_queue_size, workers, use_multiprocessing, model, steps_per_execution)
1108 use_multiprocessing=use_multiprocessing,
1109 distribution_strategy=ds_context.get_strategy(),
-> 1110 model=model)
1111
1112 strategy = ds_context.get_strategy()
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in __init__(self, x, y, sample_weights, workers, use_multiprocessing, max_queue_size, model, **kwargs)
796 return tensor_shape.TensorShape([None for _ in shape.as_list()])
797
--> 798 output_shapes = nest.map_structure(_get_dynamic_shape, peek)
799 output_types = nest.map_structure(lambda t: t.dtype, peek)
800
d:pyenvkeras-gpulibsite-packagestensorflowpythonutilnest.py in map_structure(func, *structure, **kwargs)
633
634 return pack_sequence_as(
--> 635 structure[0], [func(*x) for x in entries],
636 expand_composites=expand_composites)
637
d:pyenvkeras-gpulibsite-packagestensorflowpythonutilnest.py in <listcomp>(.0)
633
634 return pack_sequence_as(
--> 635 structure[0], [func(*x) for x in entries],
636 expand_composites=expand_composites)
637
d:pyenvkeras-gpulibsite-packagestensorflowpythonkerasenginedata_adapter.py in _get_dynamic_shape(t)
792 shape = t.shape
793 # Unknown number of dimensions, `as_list` cannot be called.
--> 794 if shape.rank is None:
795 return shape
796 return tensor_shape.TensorShape([None for _ in shape.as_list()])
AttributeError: 'tuple' object has no attribute 'rank'
So I tried to look at Keras sources but without any success.
If I use modified train_generator and validation_generator (y_col='target' instead of y_col=train.columns[1:]) everything works fine.
You need to convert all the individual objects returned by both the training and validation generators to Numpy arrays:
yield [np.array(imgs), np.array(cols)], np.array(targets)
Alternatively, a simpler and much more efficient solution is to not iterate over the data batch at all; instead, we can take advantage of the fact that these objects are already Numpy arrays when returned by ImageDataGenerator, so we can write:
imgs = data[0]
cols = data[1][:,:-1]
targets = data[1][:,-1:]
yield [imgs, cols], targets
A different solution worked for me, just posting it here.
I ran into the problem working with two very similar dataframes in one notebook, where for one of them the error occurred.
I noticed the dtypes were slightly different int64 vs Int64, where the target column coded as Int64 gave the error.
For me the following worked:
dataframe[target_col] = dataframe[target_col].astype(int)