frozen model inference labels on custom dataset

[td43]

Hi there,

I have a question regarding the frozen_model.pb

I trained a fine-tuned model with a pre-trained checkpoint with efficientdet-d0 to make inferences on my own custom dataset. I exported the frozen model with tf2/

python inspector.py \
--mode=saved_model \
--model_name=efficientdet-d0 \
--model_dir=/home/daniel_tobon/workspace/result \
--saved_model_dir=/home/daniel_tobon/workspace/freeze_model \
--hparams=/home/daniel_tobon/workspace/tfrecords/hparams_config.yaml

Then I get in my output folder: assets efficientdet-d0_frozen.pb saved_model.pb variables

According to the tf2/README.md there are several ways to make predictions. Using a frozen graph or construct the graph from scratch.

My question is why the predictions with the frozen graph assign the labels from the COCO dataset and not my own custom labels.

python inspector.py \
--mode=infer \
--model_name=efficientdet-d0  \
--saved_model_dir=/home/daniel_tobon/workspace/freeze_model/efficientdet-d0_frozen.pb  \
--input_image=/home/daniel_tobon/workspace/dataset/IMAGENES/009076ba-42a4-451e-abfb-cb476eaec327.jpg \
--output_image_dir=/home/daniel_tobon/workspace/

result with frozen model:

with model built from scratch:

python inspector.py \
--mode=infer \
--model_name=efficientdet-d0 \
--model_dir=/home/daniel_tobon/workspace/result \
--hparams=/home/daniel_tobon/workspace/tfrecords/hparams_config.yaml \
--input_image=/home/daniel_tobon/workspace/dataset/IMAGENES/009076ba-42a4-451e-abfb-cb476eaec327.jpg \
--output_image_dir=/home/daniel_tobon/workspace/

result without frozen model:

If I set hparams with the frozen model, then I got the correct labels predictions. But, if the frozen model needs an extra file with the labels maps, then why in the online resources everybody says that the frozen model.pb is enough to make predictions or I am maybe misunderstanding something?

I would like to get some insight into this topic.

[mingxingtan]

Hi, you might need to specify hparams during your inference command line:

python inspector.py \
--mode=infer \
--model_name=efficientdet-d0  \
--saved_model_dir=/home/daniel_tobon/workspace/freeze_model/efficientdet-d0_frozen.pb  \
--input_image=/home/daniel_tobon/workspace/dataset/IMAGENES/009076ba-42a4-451e-abfb-cb476eaec327.jpg \
--output_image_dir=/home/daniel_tobon/workspace/ \
--hparams=/home/daniel_tobon/workspace/tfrecords/hparams_config.yaml   # Add this line

[td43]

Hi @mingxingtan thanks for your answer. as I mentioned at the end of the post:

If I set hparams with the frozen model, then I got the correct labels predictions. But, if the frozen model needs an extra file with the labels maps, then why in the online resources everybody says that the frozen model.pb is enough to make predictions or I am maybe misunderstanding something?

I already set hparams. My question is why the frozen model saves the coco labels and not my custom labels.