simrdwn
simrdwn copied to clipboard
avanetten
No detection during test time
I was trying to reproduce the result with COWC, however, there are no cars detected during test time while the loss curve during training time looks reasonable:
I was using yolo-v2-tiny config and weights, and I was using Tesla K80 on GCP with CUDA Version: 10.1. Any idea on what's going wrong??
It's hard to determine what might be happening without further details. What test data are you using, and what parameters are used for the test command?
@avanetten I am using the COWC data, here are my commands for training and testing:
python simrdwn/core/simrdwn.py \ --framework yolt2 \ --mode train \ --outname yolov2_cars \ --yolt_cfg_file yolov2tiny.cfg \ --weight_dir simrdwn2/yolt2/input_weights \ --weight_file yolov2-tiny.weights \ --yolt_train_images_list_file cowc_yolt_train_list.txt \ --label_map_path simrdwn2/train_data/class_labels_car.pbtxt \ --nbands 3 \ --max_batches 10000 \ --batch_size 64 \ --subdivisions 32 \ --gpu 0
python simrdwn/core/simrdwn.py \ --framework yolt2 \ --mode test \ --outname yolov2_cars \ --label_map_path 'simrdwn/train_data/class_labels_car.pbtxt' \ --train_model_path 'train_yolt2_yolov2_cars_2019_05_30_06-30-51' \ --weight_file yolov2tiny_final.weights \ --yolt_cfg_file yolov2tiny.cfg \ --testims_dir 'test_images/cowc' \ --keep_test_slices 0 \ --test_slice_sep __ \ --test_make_legend_and_title 0 \ --edge_buffer_test 1 \ --test_box_rescale_frac 1 \ --plot_thresh_str 0.01 \ --slice_sizes_str 416 \ --slice_overlap 0.2 \ --alpha_scaling 1 \ --show_labels 1
@avanetten also, sometimes there is a bug showing up during test time: pd.read_csv cannot find file "results/test_xxxxxx/cars.txt"
This bug can be fixed by just re-making yolt2. I am not sure why this is happening.
Can you print out the output of the test script? The issue in pd.read_csv() is usually caused because there was an error further upstream and the inference script did not run correctly
Yes sure, i was print it now in my issue: https://github.com/avanetten/simrdwn/issues/44#issuecomment-501442591
I was trying to reproduce the result with COWC, however, there are no cars detected during test time while the loss curve during training time looks reasonable:
I was using yolo-v2-tiny config and weights, and I was using Tesla K80 on GCP with CUDA Version: 10.1. Any idea on what's going wrong??
@avanetten I have encountered a similar situation.
I think , even though the model don't convergence correctly, there should be some wrong boxes in the output file,while there is nothing. It makes me confused.
I have the same problem. @MhmdDimassi @wendyzzzw did you manage to fix it?
Can you please put the test command that you use
I have the same problem. @MhmdDimassi @wendyzzzw did you manage to fix it?
Can you please put the test command that you use
python simrdwn/core/simrdwn.py --framework yolt2 --mode test --outname dense_cowc --label_map_path class_labels_car.pbtxt --train_model_path /simrdwn/results/train_yolt2_dense_cars_2019_11_05_16-40-12 --weight_file ave_dense_final.weights --yolt_cfg_file ave_dense.cfg --testims_dir cowc/Utah_AGRC --keep_test_slices 0 --test_slice_sep __ --test_make_legend_and_title 0 --edge_buffer_test 1 --test_box_rescale_frac 1 --plot_thresh_str 0.2 --slice_sizes_str 416 --slice_overlap 0.2 --alpha_scaling 1 --show_labels 1
This is likely an issue with YOLO and updates to architectures (see https://github.com/avanetten/simrdwn/issues/31#issuecomment-501925778). You may need to experiment with the yolt2 Makefile.
This is likely an issue with YOLO and updates to architectures (see #31 (comment)). You may need to experiment with the yolt2 Makefile.
OK thanks I will look into this. Do you have any working COWC weights that I may use to help debug, so I can avoid training all over again?