efficientdet hyperparameters

[avber]

Hi @leondgarse ,

is it correct to specify aspect_ratios like this?

--det_header efficientdet.EfficientDetLite4
--additional_backbone_kwargs '{"aspect_ratios": [0.44, 2.16, 6.68]}'

Is it possible to specify custom scales like [32, 64, 128, 256, 512] and not like
--additional_backbone_kwargs '{"num_scales": 3}'
?

Thanks

[leondgarse]

Currently both aspect_ratios and num_scales are not a parameter for EfficientDet. They are not affecting model structure, but data and eval requires them. Also coco.eval_func.DecodePredictions and coco.data.init_dataset both setting aspect_ratios, num_scales, grid_zero_start = [1, 2, 0.5], 3, False as fixed. May check if easy to set them custom later....

[leondgarse]

Both aspect_ratios and num_scales added a parameter for coco.eval_func.DecodePredictions / coco.data.init_dataset / coco_train_script.py / coco_eval_script.py, but be caution of their usage, for somewhere else may assuming efficientdet using num_anchors=9. Also, it requires a custom DecodePredictions, not model.decode_predictions any more...

This is not tested.

[avber]

Thank you!

These lines look suspicious.

https://github.com/leondgarse/keras_cv_attention_models/blob/main/keras_cv_attention_models/coco/eval_func.py#L358
https://github.com/leondgarse/keras_cv_attention_models/blob/main/keras_cv_attention_models/coco/eval_func.py#L365

Will look for other ones.

[avber]

And this

keras_cv_attention_models/keras_cv_attention_models/efficientdet/efficientdet.py

Line 205 in 56c222a

post_process = eval_func.DecodePredictions(backbone.input_shape[1:], pyramid_levels, anchors_mode, use_object_scores, anchor_scale)

[avber]

I guess Yolo models should also be updated

[avber]

Thank you, will try training later this week.

A few minor changes are necessary.

class COCOEvalCallback(tf.keras.callbacks.Callback):
def init():

 +  self.aspect_ratios, self.num_scales = aspect_ratios, num_scales

def build(self, input_shape, output_shape):
self.pred_decoder = DecodePredictions(

•   #input_shape, pyramid_levels, self.anchors_mode, anchor_scale=self.anchor_scale, aspect_ratios=aspect_ratios, num_scales=num_scales
  

•   input_shape, pyramid_levels, self.anchors_mode, anchor_scale=self.anchor_scale, aspect_ratios=self.aspect_ratios, num_scales=self.num_scales
  

  )

[avber]

3 issues:

It's necessary to add type=float to --aspect_ratios parsing in both train and eval scripts.

==========================
I ran training in the kecam_coco_tiny_test.ipynb. It failed with and without aspect_ratios

Traceback (most recent call last):
File "./coco_train_script.py", line 273, in
model, latest_save, _ = run_training_by_args(args)
File "./coco_train_script.py", line 257, in run_training_by_args
model, epochs, train_dataset, test_dataset, args.initial_epoch, lr_scheduler, args.basic_save_name, init_callbacks, logs=args.tensorboard_logs
File "/content/keras_cv_attention_models/keras_cv_attention_models/imagenet/train_func.py", line 250, in train
workers=8,
File "/usr/local/lib/python3.7/dist-packages/keras/utils/traceback_utils.py", line 67, in error_handler
raise e.with_traceback(filtered_tb) from None
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/eval_func.py", line 454, in on_epoch_end
detection_results = model_detection_and_decode(self.model, eval_dataset, self.pred_decoder, self.nms_kwargs, self.is_coco, self.image_id_map)
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/eval_func.py", line 261, in model_detection_and_decode
decoded_preds = pred_decoder(preds, **nms_kwargs)
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/eval_func.py", line 179, in call
return [self.decode_single(pred, **self.nms_kwargs, input_shape=input_shape) for pred in preds]
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/eval_func.py", line 179, in
return [self.decode_single(pred, **self.nms_kwargs, input_shape=input_shape) for pred in preds]
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/eval_func.py", line 151, in decode_single
bbs_decoded = anchors_func.decode_bboxes(bbs, anchors)
File "/content/keras_cv_attention_models/keras_cv_attention_models/coco/anchors_func.py", line 201, in decode_bboxes
bboxes_center = preds_center * anchors_hw + anchors_center
tensorflow.python.framework.errors_impl.InvalidArgumentError: Exception encountered when calling layer "decode_predictions_1" (type DecodePredictions).

cannot compute Mul as input #1(zero-based) was expected to be a half tensor but is a float tensor [Op:Mul]

Call arguments received by layer "decode_predictions_1" (type DecodePredictions):
• preds=tf.Tensor(shape=(16, 49104, 5), dtype=float16)
• input_shape=None
• training=False
• nms_kwargs={'score_threshold': '0.001', 'iou_or_sigma': '0.5', 'max_output_size': '100', 'method': "'gaussian'", 'mode': "'per_class'", 'topk': '5000'}

===============================

'max_output_size': '100'

I guess max_labels_per_image should be passed

[leondgarse]

OK, just fixed some of those issues. It is caused by DecodePredictions returning tf.tensor now, not np.array... Guess I need a dev branch, lol.
max_labels_per_image is for training deciding the max number of ground truth labels in a single image, not affecting other processes like eval or predict. max_output_size will take that work. Ya, they look similar.
User case in Colab kecam_coco_tiny_test.ipynb is updated using !python3 ./coco_train_script.py ... --aspect_ratios 0.44 1.16 3.33 6.68 --num_scales 5 --anchors_mode efficientdet. Currently have to specify --anchors_mode efficientdet, may check if any better way later. Just showing the basic function should work now.

[avber]

Let's say evaluation is run after a training epoch on images with the number of objects greater than max_output_size. Aren't the metrics distorted in this case?

[leondgarse]

Uh, got it. You mean pass args.max_labels_per_image as nms_max_output_size for eval_func.COCOEvalCallback in coco_train_script.py. Good idea, updated.

[avber]

A minor regression

keras_cv_attention_models/coco_train_script.py

Line 259 in 4b5ebc4

print(">>>> COCO AP eval start_epoch: {}, frequency: {}".format(start_epoch, frequency))

frequency variable was deleted.

[avber]

Done

[leondgarse]

Ya, just updated.