python - Keras model.fit ValueError: 形状 (None, 43) 和 (None, 1, 1, 43) 不兼容

Question

编辑：我在 ResNet50 和 Dense 层之间添加了一个 Flatten() 层，错误消失了

一般来说，我对使用 Keras 和机器学习非常陌生。我的目标是使用经过 ImageNet 权重预训练的 Keras ResNet50 模型对 GTSRB 中不同类型的标志进行分类。我正在使用 Google CoLab，加载和预处理所有图像/标签的代码是由我的教授编写的。我可以用模型编译和预测，我遇到的问题是训练。我正在使用 model.fit() 函数来尝试训练模型。

这些图像是 32x32 图像，所以我将 input_tensor 设置为 (32, 32, 3) 的形状。有 43 个标签可供分类，我使用的是单热编码，所以我在 ResNet50 之后添加了一个大小为 43 的层。我不确定 ValueError 到底出现在哪里，但我很确定这与标签有关。我真的不知道形状的问题是什么，我一直在查看文档和其他 Stack Overflow 答案，但我找不到任何适用于此的东西（我的理解有限）。

以下是错误：

---------------------------------------------------------------------------

ValueError                                Traceback (most recent call last)

<ipython-input-21-2f55b5d19997> in <module>()
      6 
      7 resnet.compile(optimizer='adam', loss=keras.losses.CategoricalCrossentropy(from_logits=True), metrics=['acc'])
----> 8 resnet.fit(x=imgs_train, y=labels_train, batch_size=BATCH_SIZE, epochs=NUM_EPOCH, validation_data=(imgs_val, labels_val))
      9 
     10 predictions = resnet.predict(x=imgs_test)

9 frames

/usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    992           except Exception as e:  # pylint:disable=broad-except
    993             if hasattr(e, "ag_error_metadata"):
--> 994               raise e.ag_error_metadata.to_exception(e)
    995             else:
    996               raise

ValueError: in user code:

    /usr/local/lib/python3.7/dist-packages/keras/engine/training.py:853 train_function  *
        return step_function(self, iterator)
    /usr/local/lib/python3.7/dist-packages/keras/engine/training.py:842 step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/distribute/distribute_lib.py:1286 run
        return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/distribute/distribute_lib.py:2849 call_for_each_replica
        return self._call_for_each_replica(fn, args, kwargs)
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/distribute/distribute_lib.py:3632 _call_for_each_replica
        return fn(*args, **kwargs)
    /usr/local/lib/python3.7/dist-packages/keras/engine/training.py:835 run_step  **
        outputs = model.train_step(data)
    /usr/local/lib/python3.7/dist-packages/keras/engine/training.py:789 train_step
        y, y_pred, sample_weight, regularization_losses=self.losses)
    /usr/local/lib/python3.7/dist-packages/keras/engine/compile_utils.py:201 __call__
        loss_value = loss_obj(y_t, y_p, sample_weight=sw)
    /usr/local/lib/python3.7/dist-packages/keras/losses.py:141 __call__
        losses = call_fn(y_true, y_pred)
    /usr/local/lib/python3.7/dist-packages/keras/losses.py:245 call  **
        return ag_fn(y_true, y_pred, **self._fn_kwargs)
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/util/dispatch.py:206 wrapper
        return target(*args, **kwargs)
    /usr/local/lib/python3.7/dist-packages/keras/losses.py:1666 categorical_crossentropy
        y_true, y_pred, from_logits=from_logits, axis=axis)
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/util/dispatch.py:206 wrapper
        return target(*args, **kwargs)
    /usr/local/lib/python3.7/dist-packages/keras/backend.py:4839 categorical_crossentropy
        target.shape.assert_is_compatible_with(output.shape)
    /usr/local/lib/python3.7/dist-packages/tensorflow/python/framework/tensor_shape.py:1161 assert_is_compatible_with
        raise ValueError("Shapes %s and %s are incompatible" % (self, other))

    ValueError: Shapes (None, 43) and (None, 1, 1, 43) are incompatible

这是我的代码：

drive.mount('/content/drive')

# Import libraries
import numpy as np
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import Input
from tensorflow.keras.utils import to_categorical
from tensorflow.keras.applications.resnet50 import ResNet50, decode_predictions
import matplotlib.pyplot as plt
import pickle
import pandas as pd
import cv2
from itertools import chain
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

### Load data

# Path to project directory
PROJECT_ROOT_DIR = 'drive/My Drive/Colab Notebooks/Assignment1/'

# Path to directory containing the dataset
DATA_DIR = PROJECT_ROOT_DIR + 'Data/'

# Function for loading the dataset
# Code from advml-traffic-sign (https://github.com/inspire-group/advml-traffic-sign)
def load_dataset_GTSRB(n_channel=3, train_file_name=None):
    """
    Load GTSRB data as a (datasize) x (channels) x (height) x (width) numpy
    matrix. Each pixel is rescaled to the range [0,1].
    """

    def load_pickled_data(file, columns):
        """
        Loads pickled training and test data.

        Parameters
        ----------
        file    : string
                          Name of the pickle file.
        columns : list of strings
                          List of columns in pickled data we're interested in.

        Returns
        -------
        A tuple of datasets for given columns.
        """

        with open(file, mode='rb') as f:
            dataset = pickle.load(f)
        return tuple(map(lambda c: dataset[c], columns))

    def preprocess(x, n_channel):
        """
        Preprocess dataset: turn images into grayscale if specified, normalize
        input space to [0,1], reshape array to appropriate shape for NN model
        """

        if n_channel == 3:
            # Scale features to be in [0, 1]
            x = (x / 255.).astype(np.float32)
        else:
            # Convert to grayscale, e.g. single Y channel
            x = 0.299 * x[:, :, :, 0] + 0.587 * x[:, :, :, 1] + \
                0.114 * x[:, :, :, 2]
            # Scale features to be in [0, 1]
            x = (x / 255.).astype(np.float32)
            x = x[:, :, :, np.newaxis]
        return x

    # Load pickle dataset
    if train_file_name is None:
        x_train, y_train = load_pickled_data(
            DATA_DIR + 'train.p', ['features', 'labels'])
    else:
        x_train, y_train = load_pickled_data(
            DATA_DIR + train_file_name, ['features', 'labels'])
    x_val, y_val = load_pickled_data(
        DATA_DIR + 'valid.p', ['features', 'labels'])
    x_test, y_test = load_pickled_data(
        DATA_DIR + 'test.p', ['features', 'labels'])

    # Preprocess loaded data
    x_train = preprocess(x_train, n_channel)
    x_val = preprocess(x_val, n_channel)
    x_test = preprocess(x_test, n_channel)
    return x_train, y_train, x_val, y_val, x_test, y_test

# Set constants (GTSRB)
NUM_LABELS = 43                             # Number of labels
BATCH_SIZE = 32                             # Size of batch
HEIGHT = 32                                 # Height of input image
WIDTH = 32                                  # Width of input image
N_CHANNEL = 3                               # Number of channels
OUTPUT_DIM = 43                             # Number of output dimension

# Set training hyperparameters
NUM_EPOCH = 100                             # Number of epoch to train
LR = 0.0001                                 # Learning rate

INPUT_SHAPE = (HEIGHT, WIDTH, N_CHANNEL)  # Input shape of model
IMG_SHAPE = (HEIGHT, WIDTH, N_CHANNEL)

# Load the images and labels
imgs_train, labels_train, imgs_val, labels_val, imgs_test, labels_test = load_dataset_GTSRB(n_channel=N_CHANNEL)

# Display the shapes of train, validation, and test datasets
print('Images train shape: {} - Labels train shape: {}'.format(imgs_train.shape, labels_train.shape))
print('Images validation shape: {} - Labels validation shape: {}'.format(imgs_val.shape, labels_val.shape))
print('Images test shape: {} - Labels test shape: {}'.format(imgs_test.shape, labels_test.shape))

# Display the range of images (to make sure they are in the [0, 1] range)
print('\nMax pixel value', np.max(imgs_train))
print('Min pixel value', np.min(imgs_train))
print('Average pixel value', np.mean(imgs_train))
print('Data type', imgs_train[0].dtype)

# Read the sign names
signnames = pd.read_csv(DATA_DIR + 'signnames.csv').values[:, 1]

# Plot a few images to check if the data makes sense (note that the quality of some of the images is pretty low)
plt.figure(figsize=(16, 10))
for n in range(9):
    i = np.random.randint(0, len(imgs_train), 1)
    ax = plt.subplot(3, 3, n+1)
    plt.imshow(imgs_train[i[0]])
    plt.title('Label:' + str(signnames[labels_train[i[0]]]))
    plt.axis('off')

# Convert the labels to one-hot encoding (to input to the models)
labels_train = keras.utils.to_categorical(labels_train, NUM_LABELS)
labels_test = keras.utils.to_categorical(labels_test, NUM_LABELS)
labels_val = keras.utils.to_categorical(labels_val, NUM_LABELS)

print('Labels train shape: {}'.format(labels_train.shape))

### Creating the model and training

resnet = Sequential([
  ResNet50(weights='imagenet', include_top=False, input_tensor=Input(shape=(WIDTH, HEIGHT, N_CHANNEL))),
  Dense(NUM_LABELS, activation='softmax')
])
resnet.layers[0].trainable = False

resnet.compile(optimizer='adam', loss=keras.losses.CategoricalCrossentropy(from_logits=True), metrics=['acc'])
resnet.fit(x=imgs_train, y=labels_train, batch_size=BATCH_SIZE, epochs=NUM_EPOCH, validation_data=(imgs_val, labels_val))

# predictions = resnet.predict(x=imgs_test)

这是数据：https ://drive.google.com/drive/folders/1dbr0bYFX40HoAJetAIDSh7Zsc-O5JbGS?usp=sharing

如果有人读到这个，谢谢。我很感激你的时间。

标签： pythontensorflowkeras

Answer 1

我在 ResNet50 和 Dense 层之间添加了一个 Flatten() 层，错误消失了！这是现在创建模型的代码：

resnet = Sequential(
    [
     ResNet50(weights='imagenet', include_top=False, input_tensor=Input(shape=(WIDTH, HEIGHT, N_CHANNEL))),
     Flatten(),
     Dense(NUM_LABELS, activation='softmax')
    ]
)