Keras中合并辅助输入是否保留排序？洗牌批次时拼接排序可靠吗？

1. Does merging auxiliary inputs in Keras preserve the original order? #

Absolutely—all merging operations in Keras (like Concatenate, Add, Multiply, etc.) strictly preserve the sample order across inputs.

Here's why: Every tensor passed to a merging layer has a batch dimension (usually axis=0, representing individual samples in the batch). Merging layers operate per-sample: the i-th sample in the main input tensor will be merged with the i-th sample in the auxiliary input tensor, no exceptions. This is a core design of Keras' tensor operations—layers don't reorder samples unless you explicitly add a layer or preprocessing step that does so.

2. Can we rely on merging to maintain correct pairing when batches are shuffled? #

Yes—as long as your main data and auxiliary data are paired correctly and shuffled synchronously.

Let's use your photo/day-of-week example to break this down:

• Suppose you have a dataset where each entry is a tuple: (photo_of_event, day_of_week_of_event).
• When you shuffle your dataset (either via Keras' ImageDataGenerator, a custom tf.data.Dataset, or manual indexing), you must shuffle the entire paired dataset—not the photos and day-of-week values separately.

If you do this, every photo will stay linked to its correct day_of_week value through the shuffle. When the data reaches your model:

1. The photo goes through convolution + flattening, resulting in a feature vector for each sample.
2. The day_of_week parameter is a 1D tensor for each sample.
3. The Concatenate layer will take the i-th feature vector from the photo branch and the i-th day_of_week value, and merge them into a single feature vector for the i-th sample.

Critical Caveat #

If you shuffle the photos and day_of_week values independently (e.g., shuffle photos first, then shuffle day_of_week separately), you will get incorrect pairings. To avoid this, always shuffle using a shared random seed or index permutation. For example:

import numpy as np

# Sample data
photos = np.random.rand(100, 224, 224, 3)  # 100 sample photos
day_of_week = np.random.randint(0, 7, size=(100, 1))  # Corresponding day values

# Create a single permutation of indices to shuffle both datasets identically
shuffle_indices = np.random.permutation(len(photos))
shuffled_photos = photos[shuffle_indices]
shuffled_day_of_week = day_of_week[shuffle_indices]

In Keras' tf.data pipeline, this is handled automatically if you zip paired datasets before shuffling:

import tensorflow as tf

# Create paired dataset
dataset = tf.data.Dataset.from_tensor_slices((photos, day_of_week))
# Shuffle the entire paired dataset (preserves sample pairs)
dataset = dataset.shuffle(buffer_size=100).batch(32)

Final Takeaway #

Keras' merging operations are completely reliable for maintaining sample order—you just need to ensure your data preprocessing keeps paired samples linked during shuffling.

内容的提问来源于stack exchange，提问作者Colin Beckingham