我想构建一个神经网络，它将查看图像并为我提供 64 个浮点数。
起初我有 pandas 数据框。目标帧看起来像这样（我每张图像有 64 个数字）：
a1 a2 a3 a4 a5 a6 a7 a8 b1 b2 ... h8
257 0.08 0.0 1.0 1.0 1.0 1.0 0.5 0.58 0.17 0.0 ... 0.58
258 0.08 0.0 1.0 1.0 1.0 1.0 0.5 0.58 0.17 0.0 ... 0.58
259 0.08 0.0 1.0 1.0 1.0 1.0 0.5 0.58 0.17 0.0 ... 0.58
260 0.08 0.0 1.0 1.0 1.0 1.0 0.5 0.58 0.17 0.0 ... 0.58
261 0.08 0.0 1.0 1.0 1.0 1.0 0.5 0.58 0.17 0.0 ... 0.58
图像框架如下所示：
257 chessred/images/2/G002_IMG000.jpg
258 chessred/images/2/G002_IMG001.jpg
259 chessred/images/2/G002_IMG002.jpg
260 chessred/images/2/G002_IMG003.jpg
261 chessred/images/2/G002_IMG004.jpg
然后我制作我的 tf.dataset
train_dataset = tf.data.Dataset.from_tensor_slices((train_images,train_target))
并映射图像：
def read(image_path, position):
#Load images
file = tf.io.read_file(image_path)
#Convert image into numpy array
file = tf.image.decode_image(file, channels=3, dtype=tf.float64, expand_animations = False)
#Resize images
file = tf.image.resize(file, (224, 224))
return file, position
train_dataset = train_dataset.map(read).shuffle(shuffle_num).batch(batch_size)
然后我构建我的神经网络
pre_trained = tf.keras.applications.MobileNetV2(
include_top=False,
weights='imagenet',
input_shape=(image_size, image_size, 3))
pre_trained.trainable = False
#Add pooling layer and one dense layer
pre_trained_output = pre_trained.output
global_average_pooling2d = tf.keras.layers.GlobalAveragePooling2D()(pre_trained_output)
intermediate = tf.keras.layers.Dense(128, activation='sigmoid')(global_average_pooling2d)
#Building 64 output layers
a1 = tf.keras.layers.Dense(1, name='a1', activation='sigmoid')(intermediate)
.
h8 = tf.keras.layers.Dense(1, name='h8', activation='sigmoid')(intermediate)
我编译了
output_layers = [a1, a2, a3, a4, a5, a6, a7, a8, b1, b2, b3, b4, b5, b6, b7, b8,
c1, c2, c3, c4, c5, c6, c7, c8, d1, d2, d3, d4, d5, d6, d7, d8,
e1, e2, e3, e4, e5, e6, e7, e8, f1, f2, f3, f4, f5, f6, f7, f8,
g1, g2, g3, g4, g5, g6, g7, g8, h1, h2, h3, h4, h5, h6, h7, h8]
model = tf.keras.Model(
inputs=pre_trained.inputs,
outputs=output_layers)
model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001),
loss={"a1":"mse", "a2":"mse", "a3":"mse", "a4":"mse", "a5":"mse", "a6":"mse","a7":"mse", "a8":"mse",
"b1":"mse", "b2":"mse", "b3":"mse", "b4":"mse", "b5":"mse", "b6":"mse","b7":"mse", "b8":"mse",
"c1":"mse", "c2":"mse", "c3":"mse", "c4":"mse", "c5":"mse", "c6":"mse","c7":"mse", "c8":"mse",
"d1":"mse", "d2":"mse", "d3":"mse", "d4":"mse", "d5":"mse", "d6":"mse","d7":"mse", "d8":"mse",
"e1":"mse", "e2":"mse", "e3":"mse", "e4":"mse", "e5":"mse", "e6":"mse","e7":"mse", "e8":"mse",
"f1":"mse", "f2":"mse", "f3":"mse", "f4":"mse", "f5":"mse", "f6":"mse","f7":"mse", "f8":"mse",
"g1":"mse", "g2":"mse", "g3":"mse", "g4":"mse", "g5":"mse", "g6":"mse","g7":"mse", "g8":"mse",
"h1":"mse", "h2":"mse", "h3":"mse", "h4":"mse", "h5":"mse", "h6":"mse","h7":"mse", "h8":"mse"},
metrics={"a1":"accuracy", "a2":"accuracy", "a3":"accuracy", "a4":"accuracy", "a5":"accuracy", "a6":"accuracy","a7":"accuracy", "a8":"accuracy",
"b1":"accuracy", "b2":"accuracy", "b3":"accuracy", "b4":"accuracy", "b5":"accuracy", "b6":"accuracy","b7":"accuracy", "b8":"accuracy",
"c1":"accuracy", "c2":"accuracy", "c3":"accuracy", "c4":"accuracy", "c5":"accuracy", "c6":"accuracy","c7":"accuracy", "c8":"accuracy",
"d1":"accuracy", "d2":"accuracy", "d3":"accuracy", "d4":"accuracy", "d5":"accuracy", "d6":"accuracy","d7":"accuracy", "d8":"accuracy",
"e1":"accuracy", "e2":"accuracy", "e3":"accuracy", "e4":"accuracy", "e5":"accuracy", "e6":"accuracy","e7":"accuracy", "e8":"accuracy",
"f1":"accuracy", "f2":"accuracy", "f3":"accuracy", "f4":"accuracy", "f5":"accuracy", "f6":"accuracy","f7":"accuracy", "f8":"accuracy",
"g1":"accuracy", "g2":"accuracy", "g3":"accuracy", "g4":"accuracy", "g5":"accuracy", "g6":"accuracy","g7":"accuracy", "g8":"accuracy",
"h1":"accuracy", "h2":"accuracy", "h3":"accuracy", "h4":"accuracy", "h5":"accuracy", "h6":"accuracy","h7":"accuracy", "h8":"accuracy"})
最后我训练它：
training = model.fit(train_dataset,
epochs=epochs_num)
当我这样做时，我收到此错误：
Traceback (most recent call last):
File "/ChessRec/ChessRec.py", line 154, in <module>
training = model.fit(train_dataset,
File "/.local/lib/python3.10/site-packages/keras/src/utils/traceback_utils.py", line 122, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/.local/lib/python3.10/site-packages/keras/src/trainers/compile_utils.py", line 561, in build
for name, yt, yp in zip(output_names, y_true, y_pred):
tensorflow.python.framework.errors_impl.OperatorNotAllowedInGraphError: Iterating over a symbolic `tf.Tensor` is not allowed. You can attempt the following resolutions to the problem: If you are running in Graph mode, use Eager execution mode or decorate this function with @tf.function. If you are using AutoGraph, you can try decorating this function with @tf.function. If that does not work, then you may be using an unsupported feature or your source code may not be visible to AutoGraph. See https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/autograph/g3doc/reference/limitations.md#access-to-source-code for more information.
train_dataset 的形状似乎是正确的。当我这样做时
train_dataset.take(1)<_TakeDataset element_spec=(TensorSpec(shape=(None, 224, 224, 3), dtype=tf.float32, name=None), TensorSpec(shape=(None, 64), dtype=tf.float64, name=None))>编辑：
batch_size = 50
epochs_num = 2
shuffle_num = 6000
image_size = 224`enter code here`
编辑2：
我尝试将数据集制作成numpy数组
train_dataset = train_dataset.unbatch()
images = np.asarray(list(train_dataset.map(lambda x, y: x)))
labels = np.asarray(list(train_dataset.map(lambda x, y: y)))
编译之前，然后训练
training = model.fit(x=images, y=labels,
epochs=epochs_num)
但问题仍然存在。
编辑3：
使用 numpy 数组并在
'a1':labels[:,0]'a2':labels[:,1]fit()所以我认为问题是你有一个具有多个输出的模型。 输出层 = [a1,....h8]
由于您使用多个输出，Tensorflow 使用字典将标签与输出进行匹配。每个输出层都应该有一个与您的输入标签匹配的唯一名称。
labels_dict = {col: train_targets[col].values for col in train_targets.columns}
train_dataset = tf.data.Dataset.from_tensor_slices((train_images, labels_dict))
train_dataset = train_dataset.map(read).shuffle(shuffle_num).batch(batch_size)
并删除你的线：
train_dataset = train_dataset.unbatch()
作为额外的内容，我建议也使用以下方法对图像数据进行标准化：
file = tf.image.resize(file, (224, 224))
file = tf.cast(file, tf.float32) / 255.0 # add this line
Tensorflow 在处理小数据时表现更好。