Getting NaNs while training

TensorFlow General Discussion
, ,
1

import tensorflow as tf
from keras import layers, models
import numpy as np
from keras.preprocessing import image

img_height = 128
img_width = 128
img_channels = 3

shape = (img_height, img_width, img_channels)
classes = 2

def create_model(input_shape, num_classes):
model = models.Sequential()

model.add(layers.Conv2D(32, (3, 3), activation='relu', input_shape=input_shape))
model.add(layers.MaxPooling2D((2, 2)))

model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))

model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))

model.add(layers.Flatten())

model.add(layers.Dense(128, activation='relu'))
model.add(layers.Dense(num_classes, activation='softmax'))

return model

model = create_model(shape, classes)
model.compile(optimizer=‘adam’,
loss=‘sparse_categorical_crossentropy’,
metrics=[‘accuracy’])
model.summary()

def parse_tfrecord_fn(example):
feature_description = {
‘image/encoded’: tf.io.FixedLenFeature(, tf.string),
‘image/filename’: tf.io.FixedLenFeature(, tf.string),
‘image/format’: tf.io.FixedLenFeature(, tf.string),
‘image/height’: tf.io.FixedLenFeature(, tf.int64),
‘image/width’: tf.io.FixedLenFeature(, tf.int64),
‘image/object/bbox/xmin’: tf.io.FixedLenFeature(, tf.float32),
‘image/object/bbox/xmax’: tf.io.FixedLenFeature(, tf.float32),
‘image/object/bbox/ymin’: tf.io.FixedLenFeature(, tf.float32),
‘image/object/bbox/ymax’: tf.io.FixedLenFeature(, tf.float32),
‘image/object/class/label’: tf.io.FixedLenFeature(, tf.int64),
‘image/object/class/text’: tf.io.FixedLenFeature(, tf.string),
}

example = tf.io.parse_single_example(example, feature_description)

filename = example['image/filename']
label = example['image/object/class/label']
print(f"Processing {filename}, label: {label}")

image = tf.image.decode_jpeg(example['image/encoded'], channels=3)
image = tf.image.resize(image, (img_height, img_width))
image = tf.image.convert_image_dtype(image, tf.float32)

return image, label

def create_dataset(tfrecord_path, batch_size, buffer_size):
dataset = tf.data.TFRecordDataset(tfrecord_path)
dataset = dataset.map(parse_tfrecord_fn)
dataset = dataset.shuffle(buffer_size)
dataset = dataset.batch(batch_size)
return dataset

train_dataset = create_dataset(‘default.tfrecord’, batch_size=12, buffer_size=1000)
test_dataset = create_dataset(‘default.tfrecord’, batch_size=12, buffer_size=1000)
model.fit(train_dataset, epochs=100, validation_data=test_dataset)
model.save(“model.keras”)

def load_label_map(label_map_path):
label_map = {}
with open(label_map_path, ‘r’) as file:
lines = file.readlines()
for line in lines:
if ‘id’ in line:
id_line = line.strip().split(‘:’)
label_id = int(id_line[1].strip())
elif ‘name’ in line:
name_line = line.strip().split(‘:’)
label_name = name_line[1].strip().replace(“'”, “”)
label_map[label_id] = label_name
return label_map

label_map_path = ‘label_map.pbtxt’
label_map = load_label_map(label_map_path)

img_path = ‘bmw.png’
img = image.load_img(img_path, target_size=(img_height, img_width))
img_array = image.img_to_array(img)
img_array = np.expand_dims(img_array, axis=0)
img_array /= 255.0

predictions = model.predict(img_array)

for i, prob in enumerate(predictions[0]):
label_id = i + 1
class_name = label_map.get(label_id, f’Unknown Class {label_id}')
print(f"{class_name}: {prob:.4f}")

When i have “classes = 2” i get accuracy, loss, etc equals to NaN. When i put a 3 there, it seems to be working, but of course i have a third class, that i dont really need.

3

Hi @raaw, Even though if you have the 3rd class in the dataset that don’t need while training, the model will also train on that class and you need to mention numbers of classes=3.

If you want to train the model on only 2 classes which you need, then you need to remove the data related to the 3 classes from the training dataset and need to train on those 2 classes only.

while training the model with 2 classes it will come under binary classification, so for the last layer you need to use

tf.keras.layers.Dense(1,activation='sigmoid')

and while compiling the model the loss should be binary_crossentropy. Thank You.

4

But I do not have third class. The task is object detection, so image could include more than one object and even both classes. I I’m kinda a newbie, but I’m not sure if it’s still binary classification. Correct me if I’m wrong

6

Hi @raaw, If you are training the model with only 2 classes it will comes under binary classification. Thank You.

8

So - as a sanity check - what are the max/min values of the label value across the training and test sets? Should be {0,1} for the sparse_categorical_crossentropy to work. The label_id = i + 1 in your class_name outputting makes me suspicious that there may be an issue…

You also mentioned that there might be multiple instances of objects in the image : You are only training for 1 target label for each image, correct?