新手搭建双输入图像分类模型遇ValueError问题求助

Hey there, sorry you've been stuck on this for 5 days—frustration is totally understandable when you're just starting out with deep learning. Let's break down what's going wrong and get your model up and running!

核心问题分析 #

The ValueError: setting an array element with a sequence error pops up because your data format doesn't match what your model expects. Let's break down the key mismatches:

1. 你的模型是单输入设计，但你要传入两张图像
   Your current AlexNet is built for a single input image (shape=[None, width, height, 1]), but you're trying to feed it a 2D array of nested images (X.shape=(1200,2)). TensorFlow/TFLearn can't parse this nested array structure as valid input.

2. 数据预处理后的格式不规范

   • You're reshaping individual images inside a nested array, which leaves X as an array of arrays (not a contiguous tensor the model can process).
   • Your Y is an array of lists—you need to convert this to a proper one-hot encoded numpy array of floats.
   • You also have a critical typo in preprocessing: X[i][1] = X[i][0].reshape(...) should be X[i][1].reshape(...) (you're overwriting the second image with the first one right now!).

3. 模型输入输出逻辑不匹配
   Your model expects a single input tensor, but you need a dual-input model that processes both images, combines their features, and outputs the 9-class prediction.

分步解决方案 #

1. 修正数据预处理 #

First, let's restructure your data into two separate input tensors (one for each image) and fix your labels:

train = np.load('train_data.npy')

# Split into two input arrays and labels
X1 = np.array([item[0] for item in train])
X2 = np.array([item[1] for item in train])
Y = np.array([item[2] for item in train])

# Reshape inputs to match model input shape (batch, height, width, channels)
# Normalize pixel values to 0-1 to speed up training
X1 = X1.reshape(-1, 480, 720, 1).astype('float32') / 255.0
X2 = X2.reshape(-1, 480, 720, 1).astype('float32') / 255.0

# Convert labels to float32 (model expects float type for categorical crossentropy)
Y = Y.astype('float32')

# Verify shapes (these should match our expectations)
print(X1.shape)  # Should be (1200, 480, 720, 1)
print(X2.shape)  # Same as X1
print(Y.shape)   # Should be (1200, 9)

2. 修改AlexNet为双输入模型 #

We need to adjust the model to accept two inputs, extract features from each, combine them, and then output the prediction:

import tflearn
from tflearn.layers.conv import conv_2d, max_pool_2d
from tflearn.layers.core import input_data, dropout, fully_connected, merge
from tflearn.layers.estimator import regression
from tflearn.layers.normalization import local_response_normalization

def dual_input_alexnet(width, height, lr, output):
    # Define two input layers for the two images
    input1 = input_data(shape=[None, width, height, 1], name='input1')
    input2 = input_data(shape=[None, width, height, 1], name='input2')

    # Shared feature extraction network (reuse AlexNet layers for both inputs)
    def alexnet_feature_extractor(input_layer):
        network = conv_2d(input_layer, 96, 11, strides=4, activation='relu')
        network = max_pool_2d(network, 3, strides=2)
        network = local_response_normalization(network)
        network = conv_2d(network, 256, 5, activation='relu')
        network = max_pool_2d(network, 3, strides=2)
        network = local_response_normalization(network)
        network = conv_2d(network, 384, 3, activation='relu')
        network = conv_2d(network, 384, 3, activation='relu')
        network = conv_2d(network, 256, 3, activation='relu')
        network = max_pool_2d(network, 3, strides=2)
        network = local_response_normalization(network)
        return network

    # Extract features from both inputs
    features1 = alexnet_feature_extractor(input1)
    features2 = alexnet_feature_extractor(input2)

    # Combine features (try 'concat' or 'elementwise_sum' to see what works better)
    combined_features = merge([features1, features2], mode='concat', axis=1)

    # Add fully connected layers
    network = fully_connected(combined_features, 4096, activation='tanh')
    network = dropout(network, 0.5)
    network = fully_connected(network, 4096, activation='tanh')
    network = dropout(network, 0.5)
    network = fully_connected(network, output, activation='softmax')

    # Define regression layer
    network = regression(network, optimizer='momentum', loss='categorical_crossentropy', 
                         learning_rate=lr, name='targets')

    model = tflearn.DNN(network, checkpoint_path='model_dual_alexnet', 
                        max_checkpoints=1, tensorboard_verbose=2, tensorboard_dir='log')
    return model

3. 修正模型训练代码 #

Now when you call model.fit(), you need to pass both inputs as a dictionary matching the input layer names:

WIDTH = 480
HEIGHT = 720
LR = 1e-3
EPOCHS = 10
OUTPUT_SHAPE= 9
MODEL_NAME = 'dual_alexnet_image_classifier'

# Initialize the dual-input model
model = dual_input_alexnet(WIDTH, HEIGHT, LR, OUTPUT_SHAPE)

# Train with both inputs and labels
model.fit({'input1': X1, 'input2': X2}, {'targets': Y}, 
          n_epoch=EPOCHS, validation_split=0.1, 
          snapshot_step=500, show_metric=True, run_id=MODEL_NAME)

额外注意事项 #

• Normalization: I added /255.0 to scale pixel values between 0 and 1—this helps the model train faster and more stably.
• Shared vs Separate Extractors: The code uses a shared feature extractor (weight sharing) for both images. If you want separate extractors, just duplicate the layers instead of using the helper function.
• Label Validation: Double-check that your training output is a proper one-hot list (e.g., [0,1,0,...]) with exactly 9 elements. If you're using integer labels instead, convert them to one-hot with tflearn.data_utils.to_categorical(Y, 9).

内容的提问来源于stack exchange，提问作者jundo