Hello, I retrained HRDA on single V100 according to the config file you provided, but MIoU missed by ten points. I compared the log files generated by your training and found some differences.
your log file:
exp = 10
name_dataset = 'gtaHR22cityscapesCFixHR_1024x1024'
name_architecture = 'mres40-0.1_daformer_sepaspp_sl3_mitb5'
name_decoder = 'mres40-0.1_daformer_sepaspp_sl3'
name_uda = 'dacs_a999_fdthings_rcs0.01-2.0_cpl7'
I retrain the generated log file:
exp = 'basic'
name_dataset = 'gtaHR2cityscapesHR_1024x1024'
name_architecture = 'hrda1-512-0.1_daformer_sepaspp_sl_mitb5'
name_decoder = 'hrda1-512-0.1_daformer_sepaspp_sl'
name_uda = 'dacs_a999_fdthings_rcs0.01-2.0_cpl2'
Is the poor training accuracy caused by these files? Or is it something else?
Attached is the complete log file generated by my training, I hope to get your reply, it is really important to me! Thank you very much!
20240227_115056.log
Hello, I retrained HRDA on single V100 according to the config file you provided, but MIoU missed by ten points. I compared the log files generated by your training and found some differences. your log file: exp = 10 name_dataset = 'gtaHR22cityscapesCFixHR_1024x1024' name_architecture = 'mres40-0.1_daformer_sepaspp_sl3_mitb5' name_decoder = 'mres40-0.1_daformer_sepaspp_sl3' name_uda = 'dacs_a999_fdthings_rcs0.01-2.0_cpl7'
I retrain the generated log file: exp = 'basic' name_dataset = 'gtaHR2cityscapesHR_1024x1024' name_architecture = 'hrda1-512-0.1_daformer_sepaspp_sl_mitb5' name_decoder = 'hrda1-512-0.1_daformer_sepaspp_sl' name_uda = 'dacs_a999_fdthings_rcs0.01-2.0_cpl2' Is the poor training accuracy caused by these files? Or is it something else? Attached is the complete log file generated by my training, I hope to get your reply, it is really important to me! Thank you very much! 20240227_115056.log