The .pth file with suffix finetuning_last, which contains the last weights of the model during fine-tuning and is updated for each epoch, for being able to resume the training at any time, didn't include the actual last model of the whole training. Now it does.
The .pth files with suffix finetuning_<epoch> are uploaded on W&B every 100 epochs and on the last epoch.
Now it should be possible to finetune CounTR on images smaller than 384x384 px, which are upscaled during data loading to fit the size requirement.
Testing
Updated test with external exemplars: now you can use different exemplars for different subsets, defining different test splits in the json split file.
New param max_s_cnt: the max number of small exemplars which is tolerated before splitting the image in a grid (i.e., if the number of small objects among the exemplars is less than max_s_cnt the image is processed as a whole, otherwise it is splitted in 9 parts).
New plots which show an approximate yet intuitive visualization of True Positives (in green - the predictions that matche with the GT), False Positives (in yellow -- the other predictions), False Negatives (in red - the missed predictions), based on color subtraction.
Fine-tuning
finetuning_last, which contains the last weights of the model during fine-tuning and is updated for each epoch, for being able to resume the training at any time, didn't include the actual last model of the whole training. Now it does.finetuning_<epoch>are uploaded on W&B every 100 epochs and on the last epoch.Testing
max_s_cnt: the max number of small exemplars which is tolerated before splitting the image in a grid (i.e., if the number of small objects among the exemplars is less thanmax_s_cntthe image is processed as a whole, otherwise it is splitted in 9 parts).