Question

以这种方式使用Keras训练预训练模型：

baseModel = keras.applications.resnet50.ResNet50(include_top=False, weights='imagenet')
t = baseModel.output
t = MaxPooling2D()(t)
t = Dense(1000, activation='relu', kernel_regularizer=regularizers.l2(0.01))(t)
predictions = Dense(NUMCLASSES, activation='softmax')(t)
model = Model(inputs=baseModel.input, outputs=predictions)

for layer in baseModel.layers:
    layer.trainable = False

model.compile(loss=losses.categorical_crossentropy, optimizer=keras.optimizers.Adam())

# loading the data
files = np.array(list(train_gt.keys()))
np.random.shuffle(files)
pics = [resize(io.imread(join(trainImgDir, f)), INPUTSHAPE, mode='reflect') for f in files]
pics = np.array(pics)
classes = np.array([train_gt[f] for f in files])
classes = to_categorical(classes, NUMCLASSES)

train = pics[: int(pics.shape[0] * ((SPLITDATA - 1) / SPLITDATA))]
classesTr = classes[: int(classes.shape[0] * ((SPLITDATA - 1) / SPLITDATA))]

# training
fIn = open("Error", 'w')

batchSize = 64
for ep in range(1000):
    # train data
    trLosses = np.array([], dtype='Float64')
    for s in range(train.shape[0] // batchSize + (train.shape[0] % batchSize != 0)):
        batch = train[s * batchSize : (s + 1) * batchSize]
        batchClasses = classesTr[s * batchSize : (s + 1) * batchSize]
        trLosses = np.append(trLosses, model.train_on_batch(batch, batchClasses))

我有一个错误：

  File "/home/mark/miniconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1636, in train_on_batch
check_batch_axis=True)
File "/home/mark/miniconda3/lib/python3.6/site-packages/keras/engine/training.py", line 1315, in _standardize_user_data
    exception_prefix='target')
  File "/home/mark/miniconda3/lib/python3.6/site-packages/keras/engine/training.py", line 127, in _standardize_input_data
    str(array.shape))
ValueError: Error when checking target: expected dense_2 to have 4 dimensions, but got array with shape (64, 50)

我尝试过其他损失，但这并没有帮助。 batchClasses有形状（batchSize，NUMCLASSES）=（64,50），我希望在Dense的输出中有这个形状。

Answer 1

MaxPooling2D()不会删除宽度和高度尺寸，因此t = MaxPooling2D()(t)的输出将是形状(batch_size, w, h, 2048)的张量。这就是为什么以下Dense图层会为您提供4D张量的原因。

此外，由于您没有向MaxPooling2D()提供任何参数，并使用默认参数pool_size=(2, 2)，因此w和h都可能较大比1。

所以你基本上有两个选择，取决于你认为哪个更适合你的问题：

在Flatten()之后添加MaxPooling2D()：我不确定这是否是您想要的，因为如果w和h很大，展平会导致相当大的矢量。
删除t = MaxPooling2D()(t)并使用：
1. ResNet50(..., pooling='max')（推荐）或
2. t = GlobalMaxPooling2D()(t)

ValueError：检查目标时出错：预期dense_2有4个维度，但得到的数组有形状（64,50）（Keras）

1 个答案: