TensorFlow 模型的准确率为零

Question

I am currently training a model using the Cars196 dataset from Stanford. However, with the dataset correctly imported and recognized by TensorFlow, my accuracy is still 0. I used a similar approach to train the model on other datasets and it works. Did I do anything wrong?

Here is my code

import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import csv
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.python.keras.models import Sequential
from tensorflow.python.keras.layers import Flatten,Dense

car_dir = './src/'
test_dir = './src/cars_test/'
train_dir = './src/cars_train/'
train_labels_file = './src/labels-train.csv'
test_labels_file = './src/labels-test.csv'

IMG_SIZE = (150,150)
def read_labels(label_file:str):

    pathAndClass = list()
    
    with open(label_file) as csv_file:
        reader = csv.reader(csv_file)
        next(reader) # skip first row
        for row in reader:
            pathAndClass.append([row[5].lower(), row[4]])
        
        return pd.DataFrame(pathAndClass,columns=['path', 'class'])

pathAndClass = read_labels(train_labels_file)
n_classes = np.size(np.unique(pathAndClass['class']))
pathAndClass['path'] = pathAndClass['path'].astype(str)
pathAndClass['class'] = pathAndClass['class'].astype(str)
data_gen = ImageDataGenerator(rescale = 1.0/255.0, validation_split=0.25)
BATCH_SIZE = 32
index_list = []
for i in range(0, n_classes):
    index_list.append(str(i))
train_flow = data_gen.flow_from_dataframe(
    dataframe=pathAndClass,
    x_col='path',
    y_col='class',
    directory=train_dir,
    subset="training",
    seed=42,
    target_size=IMG_SIZE,
    batch_size=BATCH_SIZE, 
    shuffle=True, 
    classes=index_list,
    class_mode='categorical')

valid_flow = data_gen.flow_from_dataframe(
    dataframe=pathAndClass,
    x_col='path',
    y_col='class',
    directory=train_dir,
    subset="validation",
    seed=42,
    target_size=IMG_SIZE,
    batch_size=BATCH_SIZE, 
    shuffle=True, 
    classes=index_list,
    class_mode='categorical')
model_nn = Sequential()
model_nn.add(Flatten(input_shape=(150,150, 3)))
model_nn.add(Dense(300, activation="relu"))
model_nn.add(Dense(n_classes, activation="softmax"))
model_nn.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
print(model_nn.summary())
training = model_nn.fit(
        train_flow,
        steps_per_epoch=train_flow.n//train_flow.batch_size,
        epochs=10,
        validation_data=valid_flow,
        validation_steps=valid_flow.n//valid_flow.batch_size)
print(model_nn.evaluate(train_flow))
plt.plot(training.history['accuracy'])
plt.plot(training.history['val_accuracy'])
plt.plot(training.history['loss'])
plt.plot(training.history['val_loss'])
plt.title('Model accuracy/loss')
plt.ylabel('accuracy/loss')
plt.xlabel('epoch')
plt.legend(['accuracy', 'val_accuracy', 'loss', 'val_loss'])
plt.show()

The output I got

Found 6078 validated image filenames belonging to 196 classes.
Found 2026 validated image filenames belonging to 196 classes.
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
flatten_1 (Flatten)          (None, 67500)             0         
_________________________________________________________________
dense_2 (Dense)              (None, 300)               20250300  
_________________________________________________________________
dense_3 (Dense)              (None, 196)               58996     
=================================================================
Total params: 20,309,296
Trainable params: 20,309,296
Non-trainable params: 0
_________________________________________________________________
None

Epoch 1/10
189/189 [==============================] - 68s 361ms/step - loss: 9.6809 - accuracy: 0.0036 - val_loss: 5.2785 - val_accuracy: 0.0030
Epoch 2/10
189/189 [==============================] - 58s 307ms/step - loss: 5.2770 - accuracy: 0.0055 - val_loss: 5.2785 - val_accuracy: 0.0089
Epoch 3/10
189/189 [==============================] - 58s 307ms/step - loss: 5.2743 - accuracy: 0.0083 - val_loss: 5.2793 - val_accuracy: 0.0104
Epoch 4/10
189/189 [==============================] - 58s 306ms/step - loss: 5.2728 - accuracy: 0.0089 - val_loss: 5.2800 - val_accuracy: 0.0089
Epoch 5/10
189/189 [==============================] - 58s 307ms/step - loss: 5.2710 - accuracy: 0.0084 - val_loss: 5.2806 - val_accuracy: 0.0089
Epoch 6/10
189/189 [==============================] - 57s 305ms/step - loss: 5.2698 - accuracy: 0.0086 - val_loss: 5.2815 - val_accuracy: 0.0089
Epoch 7/10
189/189 [==============================] - 58s 307ms/step - loss: 5.2695 - accuracy: 0.0083 - val_loss: 5.2822 - val_accuracy: 0.0089
Epoch 8/10
189/189 [==============================] - 58s 310ms/step - loss: 5.2681 - accuracy: 0.0086 - val_loss: 5.2834 - val_accuracy: 0.0089
Epoch 9/10
189/189 [==============================] - 58s 306ms/step - loss: 5.2679 - accuracy: 0.0083 - val_loss: 5.2840 - val_accuracy: 0.0089
Epoch 10/10
189/189 [==============================] - 58s 308ms/step - loss: 5.2669 - accuracy: 0.0083 - val_loss: 5.2848 - val_accuracy: 0.0089
   1578/Unknown - 339s 215ms/step - loss: 5.2657 - accuracy: 0.0085

Update 1

I have increased the training sample by decreasing the batch size to 8 . I tried to train the model again. However, the accuracy is still nearly 0.

Epoch 1/10
759/759 [==============================] - 112s 147ms/step - loss: 7.6876 - accuracy: 0.0051 - val_loss: 5.2779 - val_accuracy: 0.0089
Epoch 2/10
759/759 [==============================] - 112s 148ms/step - loss: 5.2728 - accuracy: 0.0086 - val_loss: 5.2792 - val_accuracy: 0.0089
Epoch 3/10
759/759 [==============================] - 112s 148ms/step - loss: 5.2695 - accuracy: 0.0087 - val_loss: 5.2808 - val_accuracy: 0.0089
Epoch 4/10
759/759 [==============================] - 109s 143ms/step - loss: 5.2671 - accuracy: 0.0087 - val_loss: 5.2828 - val_accuracy: 0.0089
Epoch 5/10
759/759 [==============================] - 111s 146ms/step - loss: 5.2661 - accuracy: 0.0086 - val_loss: 5.2844 - val_accuracy: 0.0089
Epoch 6/10
759/759 [==============================] - 114s 151ms/step - loss: 5.2648 - accuracy: 0.0089 - val_loss: 5.2862 - val_accuracy: 0.0089
Epoch 7/10
759/759 [==============================] - 118s 156ms/step - loss: 5.2646 - accuracy: 0.0086 - val_loss: 5.2881 - val_accuracy: 0.0089
Epoch 8/10
759/759 [==============================] - 117s 155ms/step - loss: 5.2639 - accuracy: 0.0087 - val_loss: 5.2891 - val_accuracy: 0.0089
Epoch 9/10
759/759 [==============================] - 115s 151ms/step - loss: 5.2635 - accuracy: 0.0087 - val_loss: 5.2903 - val_accuracy: 0.0089
Epoch 10/10
759/759 [==============================] - 112s 147ms/step - loss: 5.2634 - accuracy: 0.0086 - val_loss: 5.2915 - val_accuracy: 0.0089
   2390/Unknown - 141s 59ms/step - loss: 5.2611 - accuracy: 0.0088

Indeed the last dataset I used had less classes but more samples. Maybe there is another model that fits my dataset, any suggestions?

Answer 1

For computer vision problems, you want to look at Convolutional Neural Networks. If you're unfamiliar with them, they learn to identify features in images. Examples could be edges and textures in early layers, and then wheels, windows, doors, etc in later layers.

For this problem, I would suggest using an existing, pretrained network such as MobileNet V2 or InceptionNetV3 as a backbone, and then building your own classifier on top. This tutorial on the Tensorflow website will get you started https://www.tensorflow.org/tutorials/images/transfer_learning#create_the_base_model_from_the_pre-trained_convnets

Here's an excerpt from this tutorial:

base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
                                               include_top=False,
                                               weights='imagenet')

Then adding your model code from above, you could try:

model = tf.keras.Sequential([
  base_model,
  Flatten(input_shape=(150,150, 3)),  
  Dense(300, activation="relu")),  
  Dense(n_classes, activation='softmax')
])

model_nn.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])

This is the model I've used on similar datasets and got reasonable accuracy with it:

base_model = tf.keras.applications.MobileNetV2(input_shape=IMG_SHAPE,
                                               include_top=False,
                                               weights='imagenet')
model = tf.keras.Sequential([
  base_model,
  GlobalAveragePooling2D(),,  
  Dense(n_classes, activation='softmax')
])

In you current model, you are not trying to extract any features from the images. A single hidden layer with 300 neurons is nowhere near enough to be able to learn the features in images and give meaningful results.

You also need to check your input image size. MobileNet V2 works well with 224x224 colour images.

As per other comments, you will need to use the full dataset, you are not going to get any meaningful results with a few hundred images.

Answer 2

I would suggest that the ~6000 Training Samples for almost 200 classes is simply way to little for the model to work well.
The model did ~2000 wheight updates (200 in each epoch), which is way to few for it to learn to distinguish between ~200 classifications.

Maybe you had less classes and more training data in the other training sets?