使用TFLearn处理Kaggle fer2013数据集时遇张量形状不匹配错误求助

Hey there, let's get this shape mismatch issue sorted out for you! That error message tells us exactly what's going wrong: your model is expecting a 1-dimensional tensor of labels (shape (?,)), but you're feeding it a 2-dimensional one-hot encoded tensor (shape (64,7)). Here's how to fix it, depending on which approach you want to take:

Option 1: Keep using one-hot encoded labels (adjust your model) #

If you want to stick with the one-hot encoded labels you've already prepared, you need to make sure your model's output layer and loss function are set up to handle 7-dimensional targets:

• Set the output layer correctly: Add a fully connected layer with n_units=7 (matching your 7 emotion classes) and use the softmax activation function (perfect for multi-class classification).
• Use the right loss function: In the regression layer, specify loss='categorical_crossentropy'—this is designed to work with one-hot encoded labels.

Here's a quick code snippet example for your model structure:

# Input layer (assuming your fer2013 images are 48x48 grayscale)
net = tflearn.input_data(shape=[None, 48, 48, 1])

# Add your existing convolutional/pooling layers here
# ... (your existing layers like conv_2d, max_pool_2d, etc.)

# Output layer tailored for one-hot labels
net = tflearn.fully_connected(net, 7, activation='softmax')
# Regression layer with correct loss function
net = tflearn.regression(net, optimizer='adam', 
                         loss='categorical_crossentropy', 
                         metric='accuracy')

model = tflearn.DNN(net)

Now when you run model.fit(X_train, Y_one_hot, batch_size=64, ...), your Y_one_hot (shape (batch_size,7)) will match what the model expects.

Option 2: Convert one-hot labels back to integer labels (adjust your data) #

If you'd rather not tweak your model, you can convert your one-hot encoded labels back to 1-dimensional integer labels (each value is 0-6, representing the emotion class). Use NumPy's argmax function for this:

import numpy as np

# Convert one-hot labels (shape (N,7)) to integer labels (shape (N,))
Y_train = np.argmax(Y_one_hot_train, axis=1)
Y_test = np.argmax(Y_one_hot_test, axis=1)

Then, make sure your model's regression layer is set up to handle integer labels. You can either:

• Use loss='sparse_categorical_crossentropy' (designed for integer class labels), or
• Keep loss='categorical_crossentropy' and add to_one_hot=True, n_classes=7 to the regression layer, which tells TFLearn to automatically convert your integer labels to one-hot under the hood.

Example of the regression layer for integer labels:

net = tflearn.fully_connected(net, 7, activation='softmax')
net = tflearn.regression(net, optimizer='adam', 
                         loss='sparse_categorical_crossentropy', 
                         metric='accuracy')

Now feeding the integer labels (shape (batch_size,)) will align with the model's expected (?,) tensor shape.

Quick Check to Avoid Future Issues #

Double-check the shapes of your training data before fitting:

• Run print(Y_train.shape) to confirm it matches what your model expects (either (num_samples,7) for one-hot, or (num_samples,) for integers).
• Ensure your output layer's unit count always matches the number of classes in your dataset.

内容的提问来源于stack exchange，提问作者Mahmoud S. Ahmed