Tensorflow 训练准确率和损失与评估相同数据集不同

Question

I try to train a Tensorflow model with two classes. My Trainingsdata is balanced (~11k images for both classes). I am using Tranferlearning and try to continue on the InceptionV3 Model with the following code:

BUFFER_SIZE = 1000
BATCH_SIZE = 32

def get_label(file_path, class_names):
  # convert the path to a list of path components
  parts = tf.strings.split(file_path, os.path.sep)
  # The second to last is the class-directory
  return parts[-2] == class_names

def parse_image(filename):
    parts = tf.strings.split(filename, "\\")
    label = get_label(filename, CLASS_NAMES)

    image = tf.io.read_file(filename)
    image = tf.image.decode_png(image, channels=3)
    image = tf.image.convert_image_dtype(image, tf.float32)
    image = tf.image.resize(image, [299,299])/255.0
    return image, label

datasetFilePath = "Path\To\BalancedData"
IMAGESIZE = 299
AUTOTUNE = tf.data.experimental.AUTOTUNE
datasetPath = pathlib.Path(datasetFilePath)
list_ds = tf.data.Dataset.list_files(str(datasetPath/"*/*"))

CLASS_NAMES = np.array([item.name for item in datasetPath.glob('*')])

# labeled_ds = list_ds.map(process_path)

images_ds = list_ds.map(parse_image, num_parallel_calls=AUTOTUNE)
images_ds = images_ds.shuffle(BATCH_SIZE)

dataset = images_ds.batch(BATCH_SIZE, drop_remainder=True)

for image_batch, label_batch in dataset.take(1):
    pass

def build_and_compile_model():
    base_model =tf.keras.applications.InceptionV3(include_top=False, weights = "imagenet", input_shape=(IMAGESIZE,IMAGESIZE,3))
    feature_batch = base_model(image_batch)
    print(feature_batch.shape)

    base_model.trainable = False
    x = base_model.output
    x = tf.keras.layers.GlobalAveragePooling2D(name="avg_pool")(x)
    x = tf.keras.layers.Dense(256, activation="relu")(x)
    predictions = tf.keras.layers.Dense(2, activation="softmax")(x)
    model = tf.keras.models.Model(inputs=base_model.input, outputs=predictions)

    base_learning_rate = 0.00001
    model.compile(optimizer=tf.keras.optimizers.Adam(lr=base_learning_rate),
                 loss="categorical_crossentropy",
                 metrics=["accuracy"])
    return model


multi_worker_model = build_and_compile_model()
tensorboard = tf.keras.callbacks.TensorBoard(logdir, histogram_freq=1)
history = multi_worker_model.fit(dataset, epochs=2, callbacks=[tensorboard])

=>Epoch 1/2
=>711/711 [==============================] - 176s 247ms/step - loss: 0.3309 - accuracy: 0.8945
=>Epoch 2/2
=>711/711 [==============================] - 166s 233ms/step - loss: 0.1410 - accuracy: 0.9632

predictions = multi_worker_model.evaluate(dataset)
=>711/711 [==============================] - 150s 212ms/step - loss: 0.7538 - accuracy: 0.4999

So all in all my model just predicts everything to be of class 1 and i dont understand why that is and mostly why it claims to have a training accuracy of 96%

It might be a problem of the Data, but then again, i would expect a training accuracy of also ~50% as both training and evaluation happen on the same dataset.

Any help would be much appreciated and if you need more information let me know!

Answer 1

Ok, after a few more hours of reseach i might have the Problem. InceptionV3 uses

def preprocess_input(x):
    x /= 255.
    x -= 0.5
    x *= 2.
    return x

as preprocess function. And i am only doing the x/=255 part wich is not enough. The difference of the accuracy might be because in the .fit function automatically does the preprocess, the .evaluate on the other hand cant do that.

I am not 100% sure that this is the correct answer, but its my best guess. And it also works with setting the base_model.trainable = False to base_model.trainable = True . I think this allows the model to train with my own preprocessing and so the evaluation works also, because it gets my preprocessing.

I hope this might help someone with a similar problem. But i am still glad for any input from people who might give a better insight into the topic.