Why we get two classes mAP = 0 during training Yolov4?

we have trained models on TAO Tool kit that work fine. but now we reannotate the dataset and again we training the model but during the training, we got two class zero mAP why did we try to debug but did not understand.

First time we prepare the dataset but then we realised that it may be a problem of data annotation but now we have used the correct annotations, you can also check my previous post for wrong dataset preparation but now we have correctly annotated and getting the below result again
This is the result, the two classes are valid_balcony and invalid_balcony.

Epoch 60/80
234/234 [==============================] - 282s 1s/step - loss: 57.2428
Producing predictions: 100%|████████████████████| 67/67 [00:15<00:00,  4.24it/s]
Start to calculate AP for each class
*******************************
invalid_balconyAP    0.0
person_with_helmetAP    0.81422
person_without_helmetAP    0.70554
rail          AP    0.90052
valid_balcony AP    0.0
              mAP   0.48406
*******************************
Validation loss: 25.540452601304693
INFO: Evaluation metrics generated.

Epoch 00060: saving model to /workspace/tao-experiments/yolo_v4/experiment_dir_unpruned/weights/yolov4_resnet18_epoch_060.tlt
INFO: Training loop in progress
Epoch 61/80
234/234 [==============================] - 282s 1s/step - loss: 56.8294
INFO: Training loop in progress
Epoch 62/80
234/234 [==============================] - 283s 1s/step - loss: 56.7906
INFO: Training loop in progress
Epoch 63/80
234/234 [==============================] - 284s 1s/step - loss: 55.8155
INFO: Training loop in progress
Epoch 64/80
234/234 [==============================] - 289s 1s/step - loss: 55.0697
INFO: Training loop in progress
Epoch 65/80
234/234 [==============================] - 287s 1s/step - loss: 54.9466
INFO: Training loop in progress
Epoch 66/80
234/234 [==============================] - 284s 1s/step - loss: 54.3086
INFO: Training loop in progress
Epoch 67/80
234/234 [==============================] - 283s 1s/step - loss: 53.7069
INFO: Training loop in progress
Epoch 68/80
234/234 [==============================] - 283s 1s/step - loss: 53.2923
INFO: Training loop in progress
Epoch 69/80
234/234 [==============================] - 282s 1s/step - loss: 53.2200
INFO: Training loop in progress
Epoch 70/80
234/234 [==============================] - 282s 1s/step - loss: 53.1674
Producing predictions: 100%|████████████████████| 67/67 [00:16<00:00,  4.17it/s]
Start to calculate AP for each class
*******************************
invalid_balconyAP    0.0
person_with_helmetAP    0.82704
person_without_helmetAP    0.71231
rail          AP    0.89782
valid_balcony AP    0.0
              mAP   0.48743
*******************************
Validation loss: 23.678790747229733
INFO: Evaluation metrics generated.

what is the problem we have generated kitti label and then converted to tfrecord?

the orignal anotation is little different

Invalid_balcony      # this first letter is capital
person_with_helmet
person_without_helmet  
rail          
Valid_balcony AP      # same there first letter is capital.

how we use in train specific configuration file where we map key and value should we use the orignal anotation in value or key?

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks

Can you share training spec file and several label files?

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.