CODEWITHSUNDEEP
tensorflowobject-detection
iSeeDeadPeople
iSeeDeadPeople

Reputation: 59

Tensorflow Object Detection API - Adding more training data half way, is it ok?

Currently I have a total of 300 images for the training dataset, the current saved checkpoint is 8054. Trained for 10 hours and having accuracy of around 50%. The model is required to detect only one object.

I wish to improve the accuracy. So, I am just wondering, what if I wish to add more training images? I will need to create new, xml files, csv files and for sure, new TF records as well. Do I need to start the training all over again? Or is it just, add images, new xml, new csv, then new TF records, and just continue the training at checkpoint 8054?

I am using SSD mobilenet COCO model, batch size 5. Besides of just merely increasing the number of training dataset, what are the parameters / factors that I need to change / improve in order to increase the accuracy?

Upvotes: 3

Views: 1291

Answers (2)

Phoenix666
Phoenix666

Reputation: 182

Correct me if I am wrong but I don't believe the accepted answer is completely accurate. Every time you call train.py, your box predictors and global step get initialized from random weights and zero and it is as if you are almost training from scratch i.e. you have only managed to retrieve your feature extractors. If you don't care about the global_step, you can use the following simple and dirty hack to recover your bounding box weights:

In the file meta_architectures/ssd_meta_arch.py change the line #694 from 

if variable.op.name.startswith(self._extract_features_scope):

to this:

if variable.op.name.startswith(self._extract_features_scope) or (from_detection_checkpoint and variable.op.name.startswith("BoxPredictor_")):

Upvotes: 1

Long Hoang Nguyen
Long Hoang Nguyen

Reputation: 116

To answer your first question, yes, you can add more data (as a separate tfrecord file, for example) to the object detection API. Just edit the pipeline.config file for your model. You can actually add a whole list of .record files to the input_path parameters.

It would look roughly like this:

train_input_reader: {
  tf_record_input_reader {
    input_path: ["path/to/first.record", "path/to/second.record"]
  }
label_map_path: "..."
}

If you don't change any other parameters, training should simply continue from the checkpoint and it will load the new data.

To answer the second question, for SSD a simple thing to look at is the learning rate, it might be too high and the loss might jump a lot. You would have to analyze the loss using tensorboard yourself, though. As for other hyperparameters, I didn't have much success yet so I hope others might chime in on that one.

Upvotes: 2

Related Questions