TensorFlow Conv1D核尺寸维度错误：一维卷积维度不足问题咨询

Hey, let's break down why you're hitting that dimension error and fix it right away!

The Root Cause #

Your input shaping is flipped for how tf.layers.conv1d works. Let's look at your code:

inputs_ = tf.placeholder(tf.float32 ,(None, 45), name='inputs')
x1 = tf.expand_dims(inputs_, axis=1)  # Results in shape (None, 1, 45)

tf.layers.conv1d expects inputs in the format [batch_size, sequence_length, num_features] — and it performs convolution along the sequence_length dimension (the second axis here).

When you set kernel_size=1, the sequence length (1) was just enough to fit the kernel. But bumping it to 3 means you're trying to slide a 3-length kernel over a 1-length sequence — which is physically impossible, hence the dimension error.

The Fix #

Adjust the axis you use for expand_dims to move the feature dimension to the end, making your sequence length 45 instead of 1:

inputs_ = tf.placeholder(tf.float32 ,(None, 45), name='inputs')
x1 = tf.expand_dims(inputs_, axis=-1)  # Now shape is (None, 45, 1)
x1 = tf.layers.conv1d(x1, filters=64, kernel_size=3, strides=1, padding='valid')

This gives the convolution kernel enough space (45 units) to slide over, using your 3-sized kernel as intended.

Bonus: Padding Options #

• With padding='valid', your output sequence length will be 45 - 3 + 1 = 43 (shape: (None, 43, 64)), since the kernel only slides over valid positions without padding.
• If you want the output sequence length to match the input (45), switch to padding='same' — TensorFlow will automatically add zero-padding to the ends of your sequence so the kernel can cover every input position.

内容的提问来源于stack exchange，提问作者bmbigbang