main.evaluate() gives incorrect results when validation images are not all the same size.

[bw4sz]

# Bug Report: `evaluate()` ignores `config.validation.size`

## Description

The `evaluate()` method does not use `config.validation.size` even when it is set, causing a RuntimeError when images have different dimensions. The size parameter must be passed explicitly to `evaluate()`, which is inconsistent with how other methods respect config settings.

## Expected Behavior

When `config.validation.size` is set, the `evaluate()` method should use this value by default (similar to how training and validation dataloaders use it).

## Actual Behavior

`evaluate()` ignores `config.validation.size` and requires the `size` parameter to be passed explicitly. This causes the following error when evaluating on images with different dimensions:

RuntimeError: Images in batch have different dimensions. Set validation.size in config.yaml to resize all images to a common size.

## Steps to Reproduce

```python
from deepforest import main

m = main.deepforest()
m.load_model("weecology/deepforest-tree")

# Set validation size in config (as documented)
m.config["validation"]["size"] = 800

# This fails even though config.validation.size is set
results = m.evaluate(csv_file="test.csv", root_dir="./data")
# RuntimeError: Images in batch have different dimensions...

Current Workaround

Pass size explicitly to evaluate():

# Must pass size explicitly (redundant with config setting)
results = m.evaluate(csv_file="test.csv", root_dir="./data", size=800)

Minimal Reproducible Example

See reproduce_evaluate_size_issue.py for a complete working example.

Suggested Fix

Modify the evaluate() method in main.py to use config.validation.size as default when size parameter is None:

def evaluate(self, csv_file, iou_threshold=None, root_dir=None, size=None, batch_size=None, predictions=None):
    # ...
    if size is None:
        size = self.config.validation.size
    
    if predictions is None:
        predictions = self.predict_file(csv_file, root_dir, size=size, batch_size=batch_size)
    # ...

Environment

• DeepForest version: [current main branch]
• Python version: 3.12
• OS: Linux

Additional Context

This issue affects training scripts that set config.validation.size and then call evaluate() for pre/post training evaluation. The inconsistency makes the config system less intuitive.

[jveitchmichaelis]

@bw4sz I put a larger comment in the discussion. But the key points here:

• size is ignored entirely during training, because evaluate is provided the results from the validation loop.
• Models will implicitly resize inputs if they're outside 800-1333px, unless disabled. But perhaps RetinaNet still needs the batch to be all the same?
• The effect is subtle, because you might conflicting evaluation results if your size value is outside that range.
• Bigger concern for me is avoiding situations where you train on one size (scale) image. Then when you validate on other different size images, your predictions are trash because the model treated the image as a different scale.

Setting the config value here to override None works for now. Better to deprecate size from the signature and encourage setting it in the config?

My other suggestion is to handle all resizing at the dataset level. This doesn't affect predictions, because we're already resizing. And it makes training more explicit if we put it in the augmentations. It solves the validation issue because as long as the image/window is the right scale, the absolute size doesn't matter (assuming you tile big images).

For efficiency, we could also use a Resize transform in the prediction dataset - in collate_fn? That ought to be more efficient for larger batch sizes. https://kornia.readthedocs.io/en/latest/geometry.transform.html#kornia.geometry.transform.Resize

[bw4sz]

I think the key elements here are

1. The config needs to be connected to main.evaluate().
2. We need to assert that val images of multiple sizes produce realistic results from main.evaluate(), the current code rescales the predictions, which are then compared to the ground truth of the same size (need to confirm, but appears to be the case).
3. We should expect similar box_recall and box_precision results between

trainer.validate(model)
model.evaluate()

even if resizing is happening, depending on the augmentations used and data. I am currently seeing a huge gulf in performance and I cannot tell if its due to model weights or an error in the API for main.evaluate().