Keras中SPP Net多尺寸图像的model.fit使用及性能计算问询

Hey there! Let's tackle your SPP-Net and Keras questions clearly, grounded in the paper's core idea and practical Keras behavior.

First, let's anchor this to the key SPP paper excerpt you referenced:

'For a single network to accept variable input sizes, we approximate it by multiple networks that share all parameters, while each ...'

The heart of this is shared parameters across different input sizes—you don't need a separate model for every unique image dimension. Here's how to adapt this to your 278 variable-size images:

• Group images by their dimensions: Keras requires all images in a single batch to have the same shape, so split your dataset into groups where each group has identical width/height. You don’t need dozens of groups—even clustering similar sizes (e.g., 200–300px, 300–400px) works if you want to cut down on overhead, but keeping exact original sizes preserves SPP's full feature-extraction benefit.
• Train with one shared model: Initialize your SPP-Net model once. Then, for each dimension group, call model.fit() with that group’s data. Every fit call updates the same set of model weights—this is exactly the "multiple networks sharing parameters" the paper describes.
• Use a custom data generator for smoother workflow: If manual grouping feels tedious, build a generator that yields batches of same-size images (shuffle your dataset and pull matching-size images on the fly). Then use model.fit() with this generator, setting steps_per_epoch to the total number of batches across all size groups. This avoids multiple explicit fit calls while still honoring SPP's variable-size requirement.

When you run model.fit() multiple times on the same model instance:

• Parameter updates: Keras never resets the model between calls—each fit session continues training from the current state of the weights. This is critical for making SPP-Net's shared-parameter approach work.
• Metrics (loss, accuracy, etc.): Each fit call tracks metrics independently for its own training run. If you capture the History object from each call (e.g., history_group1 = model.fit(...), history_group2 = model.fit(...)), you’ll get separate logs for each session. To get a combined view, just concatenate the history arrays (e.g., combined_loss = history_group1.history['loss'] + history_group2.history['loss']).
• Efficiency & performance: Keras logs epoch-level training time in the History object. Total training time is simply the sum of time across all fit calls. Batch processing speed depends on each group’s batch size and image dimensions—larger images will naturally take longer per batch, but Keras handles this automatically as long as each batch is uniform in size.

内容的提问来源于stack exchange，提问作者Hitesh