How to increase num_classes in ssd_mobilenet_v1 tensorflow - tensorflow-model-garden

Question

I am using ssd_mobilenet_v1_coco.config and
I changed the value of num_classes to 20 after adding 13 things after planning training
python model_main.py --alsologtostderr --model_dir=training/ --pipeline_config_path=training/ssd_mobilenet_v1_coco.config
I kept trying to learn with the command, but I get an error.
To increase num_classes
What should I do ?
Should I grab num_classes=100 from the beginning and start?
I need help.
model {
ssd {
num_classes: 20
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
File "/usr/local/lib/python3.6/dist-packages/tensorflow_core/python/training/saver.py", line 1326, in restore
err, "a mismatch between the current graph and the graph")
tensorflow.python.framework.errors_impl.InvalidArgumentError: Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:
Assign requires shapes of both tensors to match. lhs shape= [126] rhs shape= [84]
[[node save/Assign_56 (defined at /usr/local/lib/python3.6/dist-packages/tensorflow_core/python/framework/ops.py:1748) ]]

Answer 1

I recently had a similar issue. To solve my problem, I had to do the following:
In the train_config section of pipeline.config, have fine_tune_checkpoint point to the previous model checkpoint. eg: `fine_tune_checkpoint: "./model/model.ckpt"
In the model_main.py command call, make model_dir refer to a different folder from the previous checkpoint:
python research/object_detection/model_main.py \
--model_dir=./model/finetune0 \
--pipeline_config_path=./model/pipeline.config \
--alsologtostderr
My file structure:
+ models
-+ model
--+ checkpoint
--+ model.ckpt.index
--+ model.ckpt.meta
--+ model.ckpt.data-00000-of-00001
--+ pipeline.config
--- finetune0 (will be autogenerated)
-- data (tfrecord dataset)
-- annotations (labels)
...
Context
Looks like when you already have a checkpoint at the model_dir, the script will try to resume the training on the model provided, but the new configuration on pipeline.config won't match the current model (num_class differs).
If you provide this checkpoint in the fine_tune_checkpoint and point model_dir to a new folder, it will build the model from the checkpoint variable, tweak it to match the new config, and then start the training.