Got Bad result after inference command

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

TLT Version → docker_tag: v3.21.08-py3
Network Type → Yolov4
training spec file →
spec.txt (2.5 KB)

Hi,

After Training till 100 epochs I got 96 map. But still getting bad result.

After running inference part I am getting bad result like multiple bounding boxes. There are only 3 classes images I have in my dataset with proper annotation.1536 images in the training set and 384 images in the validation dataset. I am not able to understand why the model is behaving like this. Find the attached spec file for your reference.

I have also attached the inference result of two images below. Looking forward for the help from your side.

Inference Results →
Image 1 →
2

Image 2 →
1

2

May I know how did you run inference? Using "tao yolov4 inference xxx "?

3

Yes,

My inference commad is :> tao yolo_v4 inference -i /workspace/tao-experiments/test_images -o /workspace/tao-experiments/specs -e /workspace/tao-experiments/specs/spec.txt -m /workspace/tao-experiments/results/weights/yolov4_resnet18_epoch_100.tlt -k KEY

4

To narrow down, can you run inference against your val dataset?
/workspace/tao-experiments/val/images

5

I tried with the validation dataset and as well as tried on some other images by downloading
from the google but I am got the same bad result after inference command.

6

Hi,

I am waiting for your reply

7

So, you mean you get bad result even with validation dataset, right?
That is not expected because you mention that the training got 96% mAP. The mAP is evaluated via your validation dataset.
Could you double check your training log?

8

yeah, I got the bad result even with validation dataset. map is evaluated via validation dataset. in the training log at 90 epochs I got 98% and at 100 epoch I got 100% map. So i just used the model file of 90 epochs in order to run inference but got the bad result. Do i need to change something in the nms_config in order to get good result (having single bounding box)?

9

So, you mentioned with below.
“tao yolov4 evaluate xxx” against your val dataset: according to your training result, it is 98%.
But with
“tao yolov4 inference xxx” against your val dataset: the result is bad.

Could you double check? It does not make sense.

10

in both the cases i got bad result whether i am using validation dataset (images which are present in the dataset) for inference part or not.

11

Could you share your training log?

12

Here is the training log

training log :->
yolov4_training_log_resnet18.csv (6.2 KB)

13

To narrow down, could you run "tao evaluate xxx " against the last .tlt model ?

14

I ran the inference part by using the last tlt model file i.e which comes at 100th epoch. As in the log file its 100% map of the last model file.

15

It is very strange that you get different result between evaluation in the training and “tao yolo_v4 evaluate xxx”. Actually they should be the same.

Can you share all the command and log when you run “tao evaluate xxx”?

16

log file i have already shared to u. I am sharing few commands that i have used

first I had calculated the anchor box using this command and put those shape values inside the spec file.

Command to get anchor shape :-
tao yolo_v4 kmeans -l “/workspace/tao-experiments/labels” -i “/workspace/tao-experiments/images” -x 416 -y 416

then used training command,training went upto 100 epcoh as i mentioned in the spec file.
Training command->
tao yolo_v4 train --gpus 1 -e /workspace/tao-experiments/specs/spec.txt -r /workspace/tao-experiments/results -k KEY

after training I got 100% accuracy at 100th epoch. After that i ran the evaluation command
Evaluation command →
tao yolo_v4 evaluate -e /workspace/tao-experiments/specs/spec.txt -m /workspace/tao-experiments/results/weights/yolov4_resnet18_epoch_100.tlt -k KEY

after running the evalutation command I got output like this :-

Start to calculate AP for each class

biscuits AP 1.0
chips AP 1.0
chocolate AP 1.0
mAP 1.0

then at last I just ran the inference command :-

Inference command → tao yolo_v4 inference -i /workspace/tao-experiments/val/images -o /workspace/tao-experiments/results -e /workspace/tao-experiments/specs/spec.txt -m /workspace/tao-experiments/results/weights/yolov4_resnet18_epoch_100.tlt -k KEY

then I am able to get that bad result as output images which i already shared to u

17

So, according to your result, the "tao yolo_v4 evaluate xxx " is good, right?
This matches the evaluation result during the training.

18

right…the only thing i am not able to understand is the multiple bounding appears in the output images. how to overcome from this multiple bounding box issue?

19

Can you resize several images of val dataset( /workspace/tao-experiments/val/images ) to 416x416 and re-run "tao yolo_v4 inference xxx ?

20

the below result I am getting after converting images into 416*416

1000

Still getting bad result