Numpy Array of tensorflow.keras.preprocessing.text.Tokenizer.texts_to_sequences is giving weird output, list([2]) instead of [[2]]

Question

Numpy Array of tensorflow.keras.preprocessing.text.Tokenizer.texts_to_sequences is giving weird output for Training Labels as shown below:

(training_label_list[0:10]) = [list([1]) list([1]) list([1]) list([1]) list([1]) list([1]) list([1]) list([1]) list([1]) list([1])]

but is printing Normal Array for the Validation Labels,

(validation_label_list[0:10]) = [[16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]]

In other words, type(training_label_list[0]) = <class 'list'> but

type(validation_label_list[0]) =  <class 'numpy.ndarray'>

Consequently, while Training the Model using Keras Model.fit, it is resulting in the below Error,

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

This is the Link of the Google Colab, to reproduce the error easily.

Complete Code to reproduce the Error is given below:

!pip install tensorflow==2.1

# For Preprocessing the Text => To Tokenize the Text
from tensorflow.keras.preprocessing.text import Tokenizer
# If the Two Articles are of different length, pad_sequences will make the length equal
from tensorflow.keras.preprocessing.sequence import pad_sequences

# Package for performing Numerical Operations
import numpy as np

Unique_Labels_List = ['India', 'USA', 'Australia', 'Germany', 'Bhutan', 'Nepal', 'New Zealand', 'Israel', 'Canada', 'France', 'Ireland', 'Poland', 'Egypt', 'Greece', 'China', 'Spain', 'Mexico']


Train_Labels = Unique_Labels_List[0:14]
#print('Train Labels = {}'.format(Train_Labels))

Val_Labels =  Unique_Labels_List[14:]
#print('Val_Labels = {}'.format(Val_Labels))

No_Of_Train_Items = [248, 200, 200, 218, 248, 248, 249, 247, 220, 200, 200, 211, 224, 209]
No_Val_Items = [212, 200, 219]

T_L = []
for Each_Label, Item in zip(Train_Labels, No_Of_Train_Items):
    T_L.append([Each_Label] * Item)

T_L = [item for sublist in T_L for item in sublist]

V_L = []
for Each_Label, Item in zip(Val_Labels, No_Val_Items):
    V_L.append([Each_Label] * Item)

V_L = [item for sublist in V_L for item in sublist]


len(T_L)

len(V_L)

label_tokenizer = Tokenizer()

label_tokenizer.fit_on_texts(Unique_Labels_List)

# Since it should be a Numpy Array, we should Convert the Sequences to Numpy Array, for both Training and 
# Test Labels

training_label_list = np.array(label_tokenizer.texts_to_sequences(T_L))

validation_label_list = np.array(label_tokenizer.texts_to_sequences(V_L))

print('(training_label_list[0:10]) = {}'.format((training_label_list[0:10])))
print('(validation_label_list[0:10]) = {}'.format((validation_label_list[0:10])))

print('type(training_label_list[0]) = ', type(training_label_seq[0]))
print('type(validation_label_seq[0]) = ', type(validation_label_seq[0]))

I will be Grateful if someone can suggest me how can I get both Training Labels and Validation Labels in same Format, as I have spent so much time on it.

Answer 1

Replacing np.array with np.hstack as mentioned in this Stack Overflow Answer has fixed that problem for me.

Now, the Correct Output is

(training_label_seq[0:10]) = [1 1 1 1 1 1 1 1 1 1]
(validation_label_seq[0:10]) = [16 16 16 16 16 16 16 16 16 16]
type(training_label_list[0]) =  <class 'numpy.int64'>
type(validation_label_seq[0]) =  <class 'numpy.int64'>

Link of the working code is in this Google Colab.

Mentioned below is the working code (just in case if the above link doesn't work):

!pip install tensorflow==2.1

# For Preprocessing the Text => To Tokenize the Text
from tensorflow.keras.preprocessing.text import Tokenizer
# If the Two Articles are of different length, pad_sequences will make the length equal
from tensorflow.keras.preprocessing.sequence import pad_sequences

# Package for performing Numerical Operations
import numpy as np

Unique_Labels_List = ['India', 'USA', 'Australia', 'Germany', 'Bhutan', 'Nepal', 'New Zealand', 'Israel', 'Canada', 'France', 'Ireland', 'Poland', 'Egypt', 'Greece', 'China', 'Spain', 'Mexico']


Train_Labels = Unique_Labels_List[0:14]
#print('Train Labels = {}'.format(Train_Labels))

Val_Labels =  Unique_Labels_List[14:]
#print('Val_Labels = {}'.format(Val_Labels))

No_Of_Train_Items = [248, 200, 200, 218, 248, 248, 249, 247, 220, 200, 200, 211, 224, 209]
No_Val_Items = [212, 200, 219]

T_L = []
for Each_Label, Item in zip(Train_Labels, No_Of_Train_Items):
    T_L.append([Each_Label] * Item)

T_L = [item for sublist in T_L for item in sublist]

V_L = []
for Each_Label, Item in zip(Val_Labels, No_Val_Items):
    V_L.append([Each_Label] * Item)

V_L = [item for sublist in V_L for item in sublist]


len(T_L)

len(V_L)

label_tokenizer = Tokenizer()

label_tokenizer.fit_on_texts(Unique_Labels_List)

# Since it should be a Numpy Array, we should Convert the Sequences to Numpy Array, for both Training and 
# Test Labels

training_label_list = np.hstack(label_tokenizer.texts_to_sequences(T_L))

validation_label_list = np.hstack(label_tokenizer.texts_to_sequences(V_L))

print('(training_label_list[0:10]) = {}'.format((training_label_list[0:10])))
print('(validation_label_list[0:10]) = {}'.format((validation_label_list[0:10])))

print('type(training_label_list[0]) = ', type(training_label_seq[0]))
print('type(validation_label_seq[0]) = ', type(validation_label_seq[0]))

Answer 2

Your problem is that, while your are converting your training data to a numpy array, that specific numpy array consists of list elements, hence the error

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

The error is subtler than it appears; some have reported that they had to switch back from 2.1.0 to 2.0.0. What is the difference between Numpy's array() and asarray() functions?

I would personally try this:

1. Use training_label_list = np.asarray(label_tokenizer.texts_to_sequences(T_L)), instead of np.array. Tensorflow - ValueError: Failed to convert a NumPy array to a Tensor (Unsupported object type float)
2. According to this:

List of lists into numpy array

you will have to force the casting(although weird yet this should work):

x=[[1,2],[1,2,3],[1]]
y=numpy.array([numpy.array(xi) for xi in x])
type(y)
>>><type 'numpy.ndarray'>
type(y[0])
>>><type 'numpy.ndarray'>

While trying to help you on this issue, I discovered an interesting fact about numpy casting:

CASE 1:

   my_list = [[1,2],[2],[3]]
   my_numpy_array = np.array(my_list)
   print(type(my_numpy_array))
   print(type(my_numpy_array[0]))
   <class 'numpy.ndarray'>
   <class 'list'>

CASE 2:

    my_list = [[1],[2],[3]]
    my_numpy_array = np.array(my_list)
    print(type(my_numpy_array))
    print(type(my_numpy_array[0]))
    <class 'numpy.ndarray'>
    <class 'numpy.ndarray'>

Short conclusion: If the sublists lengths differ, apparently they are left as lists and not converted to numpy arrays.

I tested on your code, now it works:

training_label_seq = np.asarray(label_tokenizer.texts_to_sequences(T_L))

training_label_seq = np.array([np.array(training_element) for training_element in training_label_seq])

validation_label_seq = np.asarray(label_tokenizer.texts_to_sequences(V_L))



print('(training_label_seq[0:10]) = {}'.format((training_label_seq[0:10])))
print('(validation_label_seq[0:10]) = {}'.format((validation_label_seq[0:10])))

print('type(training_label_list[0]) = ', type(training_label_seq[0]))
print('type(validation_label_seq[0]) = ', type(validation_label_seq[0]))



(training_label_seq[0:10]) = [array([1]) array([1]) array([1]) array([1]) array([1]) array([1])
 array([1]) array([1]) array([1]) array([1])]
(validation_label_seq[0:10]) = [[16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]
 [16]]
type(training_label_list[0]) =  <class 'numpy.ndarray'>
type(validation_label_seq[0]) =  <class 'numpy.ndarray'>