search

solaris33 / EAST-tf2

TensorFlow2 and Keras implementation of EAST: An Efficient and Accurate Scene Text Detector
GNU General Public License v3.0
20 stars 8 forks source link

How many steps to train ? #2

Open mohammedayub44 opened 3 years ago

mohammedayub44 commented 3 years ago

Hi @solaris33, I'm using your setup to train on the ICDAR dataset link you have provided. I ran it to about a 1000 steps initially to see if I get any bounding boxes. The loss is not decreasing so much, I'm I doing something wrong here or 1000 steps is too less When run the eval on 1000'th checkpoint I don't get any boxes with the default thresholds in place, when I reduce the score_map_thresh =0.5 I start to get too many irrelevant boxes.

So my question is how many steps do I have to train to eventually start seeing good eval results with the default thresholds in place ?

Thanks !

solaris33 commented 3 years ago

Hi.

I set default train step 100,000 (https://github.com/solaris33/EAST-tf2/blob/master/train.py#L17) and use it for example images(https://github.com/solaris33/EAST-tf2/tree/master/examples)

I think at least 50,000 step need to relevant prediction result and loss is meaningless signal (it doesn't decrease after amount of step)

Thanks!

rahulhanotDTU commented 1 year ago

Does 50000 step are enough for this model to perform the best and i have just started the training process for start