How to improve？

[niliuxi]

Calculating mAP...

       |  all  |  .50  |  .55  |  .60  |  .65  |  .70  |  .75  |  .80  |  .85  |  .90  |  .95  |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+
   box | 12.25 | 19.34 | 19.34 | 19.34 | 19.34 | 16.96 | 16.96 | 10.43 |  0.83 |  0.00 |  0.00 |
  mask | 16.04 | 19.34 | 19.34 | 19.34 | 19.34 | 16.96 | 16.96 | 16.96 | 16.96 | 15.18 |  0.00 |
-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+-------+

This is my result. How can I improve?

I only trained one kind, the result should not be higher?

[dbolya]

Sorry, can you go over exactly what your training procedure was, what data you're using, and what your parameters are?

[niliuxi]

I used the coco dataset of the date that I made myself, and the parameters were not modified.

[dbolya]

I mean what kind of dataset, what's your dataset config look like, and what training command did you use? That low mAP might indicate that you set something up wrong.

[niliuxi]

I'm using the coco dataset, which has a hundred images. my_custom_dataset = dataset_base.copy({ 'name': 'My Dataset',

'train_images': './data/coco/images/',
'train_info':   './data/coco/annotations/train2.json',

'valid_images': './data/coco/images/',
'valid_info':   './data/coco/annotations/train1.json',

'has_gt': True,
'class_names': ('zao'),
'label_map':{1:1}

}) yolact_base_config = coco_base_config.copy({ 'name': 'yolact_base',

# Dataset stuff
'dataset': my_custom_dataset,
'num_classes': len(my_custom_dataset.class_names) + 1,

I modified these.

[niliuxi]

I found this mistake. Thank you. By the way, what do I mean by B, C, m, s and t respectively? How can I get the position information of the target object?

[dbolya]

Yeah your mistake was 'class_names': ('zao'), should be 'class_names': ('zao',), (just pointing that out in case you found a different mistake).

The loss codes are as follows:

        # Loss Key:
        #  - B: Box Localization Loss
        #  - C: Class Confidence Loss
        #  - M: Mask Loss
        #  - P: Prototype Loss
        #  - D: Coefficient Diversity Loss
        #  - E: Class Existence Loss
        #  - S: Semantic Segmentation Loss

(Then new I one is Mask IoU Loss)

And where would you like to get this target information (you mean getting the x and y of the gt)?

[niliuxi]

I've found that when I use your project released in September, I get higher results, but when I use your updated project, the results are very low.

Yes, I want to get the coordinates of the boundingbox when testing. By the way, can the test results be visualized.

[niliuxi]

Hello, dbolya,I'd like to know what the following output means? Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.745 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.941 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.903 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.730 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.800 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.744 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.800 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.800 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.769 Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.837

[dbolya]

The numbers you care about there are the top one (0.745 = 74.5 in my eval.py) and the 2 ones under it (94.1 and 90.3). Since the maximum AP is 1 (or 100 in the way I report the numbers), those scores are pretty good.

Are you sure there's that much difference between the September version and the current version? That'd be worrisome.

[edwardnguyen1705]

I found this mistake. Thank you. By the way, what do I mean by B, C, m, s and t respectively? How can I get the position information of the target object?

I face a similar problem. What is the mistake you found?

[pankaja0285]

I am facing a problem with my training the model for the first time, in my case I have one class and I did set it as "class_names": ("Dauca",) and ran with image max_size = 850 and iterations = 60,

I get the following error: Scaling parameters by 0.25 to account for a batch size of 2. Per-GPU batch size is less than the recommended limit for batch norm. Disabling batch norm. train image_path=../dataset/train loading annotations into memory... Done (t=0.44s) creating index... index created! loading annotations into memory... Done (t=0.17s) creating index... index created! val image_path=../dataset/val /home/ec2-user/anaconda3/envs/tensorflow2_p36/lib/python3.6/site-packages/torch/jit/_recursive.py:165: UserWarning: 'lat_layers' was found in ScriptModule constants, but it is a non-constant submodule. Consider removing it. " but it is a non-constant {}. Consider removing it.".format(name, hint)) /home/ec2-user/anaconda3/envs/tensorflow2_p36/lib/python3.6/site-packages/torch/jit/_recursive.py:165: UserWarning: 'pred_layers' was found in ScriptModule constants, but it is a non-constant submodule. Consider removing it. " but it is a non-constant {}. Consider removing it.".format(name, hint)) /home/ec2-user/anaconda3/envs/tensorflow2_p36/lib/python3.6/site-packages/torch/jit/_recursive.py:165: UserWarning: 'downsample_layers' was found in ScriptModule constants, but it is a non-constant submodule. Consider removing it. " but it is a non-constant {}. Consider removing it.".format(name, hint)) Initializing weights... cfg max_size: 850 Start iteration: 0 Max iterations: 240.0 len(dataset): 441 args.batch_size: 2 epoch_size: 220 calculated num_epochs= 2 Begin training!

Number of epochs: 2 Contd. iter counter: 0 on check cond-> (epoch+1)*epoch_size < iteration [W TensorIterator.cpp:918] Warning: Mixed memory format inputs detected while calling the operator. The operator will output contiguous tensor even if some of the inputs are in channels_last format. (function operator()) At epoch:0 in iteration:0 [ 0] 0 || B: 5.137 | C: 5.871 | M: 6.755 | S: 0.713 | T: 18.476 || ETA: 0:37:18 || timer: 9.327 At epoch:0 in iteration:10 [ 0] 10 || B: 4.468 | C: 4.056 | M: 5.778 | S: 0.539 | T: 14.841 || ETA: 0:07:09 || timer: 0.267 At epoch:0 in iteration:20 [ 0] 20 || B: 4.380 | C: 3.590 | M: 5.581 | S: 0.386 | T: 13.936 || ETA: 0:06:46 || timer: 5.309 At epoch:0 in iteration:30 [ 0] 30 || B: 4.411 | C: 3.377 | M: 5.226 | S: 0.308 | T: 13.321 || ETA: 0:05:16 || timer: 0.270 At epoch:0 in iteration:40 [ 0] 40 || B: 4.367 | C: 3.227 | M: 5.029 | S: 0.270 | T: 12.893 || ETA: 0:04:29 || timer: 0.239 At epoch:0 in iteration:50 [ 0] 50 || B: 4.319 | C: 3.125 | M: 4.900 | S: 0.246 | T: 12.589 || ETA: 0:04:12 || timer: 0.510 At epoch:0 in iteration:60 [ 0] 60 || B: 4.242 | C: 3.032 | M: 4.866 | S: 0.220 | T: 12.360 || ETA: 0:03:57 || timer: 0.247 At epoch:0 in iteration:70 [ 0] 70 || B: 4.262 | C: 2.977 | M: 4.783 | S: 0.209 | T: 12.231 || ETA: 0:03:28 || timer: 0.247 At epoch:0 in iteration:80 [ 0] 80 || B: 4.275 | C: 2.928 | M: 4.728 | S: 0.198 | T: 12.129 || ETA: 0:03:13 || timer: 4.119 Traceback (most recent call last): File "/home/ec2-user/anaconda3/envs/tensorflow2_p36/lib/python3.6/multiprocessing/queues.py", line 234, in _feed obj = _ForkingPickler.dumps(obj) File "/home/ec2-user/anaconda3/envs/tensorflow2_p36/lib/python3.6/multiprocessing/reduction.py", line 51, in dumps cls(buf, protocol).dump(obj) _pickle.PicklingError: Can't pickle <class 'numpy.core._exceptions.UFuncTypeError'>: it's not the same object as numpy.core._exceptions.UFuncTypeError

Please help. If go below 60 for max_iterations I get output but the mAP values are very poor.