TensorFlow训练Cars196数据集模型准确率接近0如何解决

我目前正在使用斯坦福提供的Cars196数据集训练模型，数据集已正确导入且可被TensorFlow正常识别，但模型准确率始终接近0。我曾使用相似的训练流程在其他数据集上完成训练且效果正常，不清楚当前配置是否存在错误。

我的实现代码如下：

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) # 跳过首行表头
        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()

运行得到的输出如下：

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

更新1 #

我尝试通过将batch size下调至8增加每轮训练的迭代次数，重新训练后准确率仍然接近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

我此前使用的数据集类别数更少、单类样本量更大，想了解是否有更适配该数据集的模型方案，恳请给出相关建议。

内容的提问来源于stack exchange，提问作者phantom0202