Batch generators in tf2

2023-03-11

...And how to deal with "AttributeError: tensor object has no attribute numpy" 

Below is my example inspired by https://github.com/mostafaalishahi/eICU_Benchmark_updated

Notes:

• tf.print() and run_eagerly=True

• yield function. which returns an object defining the generator to the caller  (see https://docs.python.org/3/reference/expressions.html?highlight=yield%20return#yield-expressions)

# Batch generator for binary classification problems that may be heavily imbalanced

def batch_generator( config, X, Y, rng=np.random.RandomState(0) ):

            c1 = np.where(Y.squeeze() ==1 )[0]

            c2 = np.where(Y.squeeze() ==0 )[0]

            print( 'size of class1:', len(c1),'size of class2:', len(c2) )                 

            all_index = list(range(X.shape[0]))

            while len(all_index) > (config.batch_size*0.2):

                idx1 = rng.choice( c1, int(batch_size/2) )

                idx2 = rng.choice( c2, int(batch_size/2) )

                idx = np.hstack( (idx1, idx2))

                x_batch = X[idx]

                y_batch = Y[idx]                

                idx = list(set(all_index) - set(idx))

                data_selection = list(zip(x_batch, y_batch))

                # randomize the subset

                random.shuffle(data_selection)

                x_batch, y_batch = zip(*data_selection)

                x_batch = np.array(x_batch)

                y_batch = np.array(y_batch)

            yield x_batch, y_batch 

def my_custom_metric( y_true, y_pred):   

        y = y_true[0].numpy(); print( '???\n\n', y )   # print as numpy

        tf.print(y_pred[0], output_stream=sys.stdout)  # print as tensor

        tf.print(y_pred[1], output_stream=sys.stdout)   

        return -1   # TODO: return something meaningful

model.compile( loss = my_custom_loss( n_intervals ),  

               optimizer = optim,  

               metrics=[ my_custom_metric ], 

               run_eagerly=True )  # to enable tf.printing