含3个特征的线性回归TensorFlow代码运行报ValueError错误求助

Hey Vijay, let's break down what's causing that error and get your model working properly step by step.

Core Error Cause #

The immediate problem is with your INPUT_XOR dataset: every sample except the last one has 4 elements (like [1,72,50,33.6]), but the final entry is [96,54,0] which only has 3. This makes your input list an irregular sequence, and TensorFlow can't convert it into a valid tensor—hence the "setting an array element with a sequence" error.

Additional Fixes for a Functional Regression Model #

Beyond fixing the input data, there are a few other tweaks to make your linear regression model behave as expected:

• Flexible Placeholder Shapes: Don't hardcode the batch size in your placeholders. Using [None, 4] for x and [None, 1] for y lets your model handle any batch size, not just 10 samples.
• Remove Sigmoid from Output Layer: Sigmoid squashes values to the 0-1 range, which is terrible for regression tasks where your outputs are large continuous numbers (like 148, 85, etc.). We'll use a linear activation (no activation function) for the output layer instead.
• Consistent Epoch Count: Your original code defines hm_epochs=10000 but loops up to 100000—let's align these for clarity.

Corrected Full Code #

import tensorflow as tf

# Fixed input data: last sample now has 4 elements (matches others)
INPUT_XOR = [
    [1,72,50,33.6], [1,66,31,26.6], [1,64,32,23.3], 
    [1,66,21,28.1], [1,40,33,43.1], [1,74,30,25.6], 
    [1,50,26,31.0], [1,0,29,35.3], [1,70,53,30.5], 
    [1,96,54,0]  # Fixed: added the leading 1 to match other samples
]
OUTPUT_XOR = [[148],[85],[183],[89],[137],[116],[78],[115],[197],[125]]

n_nodes_hl1 = 10
n_nodes_hl2 = 1
batch_size = 100

# Flexible placeholders (None = any batch size)
x = tf.placeholder('float', [None, 4])
y = tf.placeholder('float', [None, 1])

def train_neural_network(x):
    hidden_1_layer = {
        'weights': tf.Variable(tf.random_uniform([4, n_nodes_hl1], -1.0, 1.0)),
        'biases': tf.Variable(tf.zeros([n_nodes_hl1]))
    }
    output_layer = {
        'weights': tf.Variable(tf.random_uniform([n_nodes_hl1, n_nodes_hl2], -1.0, 1.0)),
        'biases': tf.Variable(tf.zeros([n_nodes_hl2]))
    }

    l1 = tf.add(tf.matmul(x, hidden_1_layer['weights']), hidden_1_layer['biases'])
    l1 = tf.nn.relu(l1)

    # Removed sigmoid activation for regression output
    output = tf.add(tf.matmul(l1, output_layer['weights']), output_layer['biases'])
    prediction = output

    cost = tf.reduce_mean(tf.squared_difference(prediction, y))
    optimizer = tf.train.GradientDescentOptimizer(0.01).minimize(cost)

    hm_epochs = 100000  # Aligned with loop range
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        for epoch in range(hm_epochs):
            sess.run(optimizer, feed_dict={x: INPUT_XOR, y: OUTPUT_XOR})
            if epoch % 10000 == 0:
                c = sess.run(cost, feed_dict={x: INPUT_XOR, y: OUTPUT_XOR})
                print(f'Epoch: {epoch} completed out of {hm_epochs}, Cost: {c:.4f}')

train_neural_network(x)

Key Changes Explained #

1. Fixed Input Data: The last sample in INPUT_XOR now has 4 elements, matching the rest of the dataset—this resolves the original ValueError.
2. Flexible Placeholders: Using None for the first dimension lets you train with different batch sizes later if needed.
3. Linear Output: Removing the sigmoid ensures your model can predict the full range of your target values instead of being stuck between 0 and 1.
4. Consistent Epochs: The epoch count is now aligned between the variable and the loop, making the code easier to read.

内容的提问来源于stack exchange，提问作者Vijay Prabakaran