[Refactor]: Refactor data pipeline and support multi-scale training.

[RangiLyu]

Refactor data pipeline

1. Add naive_collate

Replace old collate function with naive_collate. naive_collate only merge meta data dict values into a list from per-image's meta data which load from dataloader workers. Stacking image tensor is moved to the training task.

2. Add stack_batch_img

This function will pad all image tensor into same shape and then stack them into a batch tensor.

3. Support scaling image instead of resizing into a fix shape.

4. Support multi-scale training

Usage: Add multi_scale to data.train

data:
  train:
    name: CocoDataset
    img_path: coco/train2017
    ann_path: coco/annotations/instances_train2017.json
    input_size: [320,320] # [w,h]
    multi_scale: [0.6, 1.4] # scale factor range of multi-scale training

[codecov]

Codecov Report

Merging #311 (30923c5) into main (76e025b) will increase coverage by 0.29%.
The diff coverage is 86.33%.

@@            Coverage Diff             @@
##             main     #311      +/-   ##
==========================================
+ Coverage   73.19%   73.49%   +0.29%     
==========================================
  Files          60       61       +1     
  Lines        3944     4074     +130     
  Branches      649      665      +16     
==========================================
+ Hits         2887     2994     +107     
- Misses        863      881      +18     
- Partials      194      199       +5     

Flag	Coverage Δ
unittests	73.49% <86.33%> (+0.29%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
nanodet/model/head/gfl_head.py	89.12% <ø> (ø)
nanodet/trainer/dist_trainer.py	28.00% <0.00%> (-1.17%)	⬇️
nanodet/data/collate.py	53.48% <20.00%> (-4.41%)	⬇️
nanodet/trainer/trainer.py	73.26% <55.55%> (-1.19%)	⬇️
nanodet/trainer/task.py	57.41% <63.63%> (+0.47%)	⬆️
nanodet/data/transform/pipeline.py	73.91% <73.33%> (-26.09%)	⬇️
nanodet/data/batch_process.py	95.45% <95.45%> (ø)
nanodet/data/transform/warp.py	85.63% <98.38%> (+4.76%)	⬆️
nanodet/data/dataset/base.py	82.97% <100.00%> (+4.60%)	⬆️
nanodet/data/dataset/coco.py	72.52% <100.00%> (+0.93%)	⬆️
... and 2 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 76e025b...30923c5. Read the comment docs.