search

MhLiao / MaskTextSpotter

A PyTorch implementation of Mask TextSpotter
https://github.com/MhLiao/MaskTextSpotter
413 stars 94 forks source link

how to train from the begin #33

Open ChChwang opened 4 years ago

ChChwang commented 4 years ago

The pretrain code loaded the trained model to get the model, how to train from the begin without downloaded model? the loss seems to be very small.

MhLiao commented 4 years ago

You can pre-train the model with SynthText.

Shualite commented 4 years ago

SynthText only has '.mat' for ground truth labeling. How can I know the structure of file '.txt' for gt labeling?

Shualite commented 4 years ago

@MhLiao

MhLiao commented 4 years ago

@Shualite I have provided the converted labels of SynthText here.

ChChwang commented 4 years ago

@MhLiao the cfg.SOLVER.RESUME shoule be set False when pretraining? And the model you provided was evaluated, and the F1 is 0.80, because of it's epoch is 150k? thanks a lot

sno6 commented 4 years ago

Has anybody had any luck with training a custom model completely from scratch? can I simply provide a model such as the finetuned weights and then set RESUME to false or is there more to it?