# 0 mAP over 50 Epochs while training TLT DetectNet\_v2 MobileNet\_v2

**URL:** <https://forums.developer.nvidia.com/t/0-map-over-50-epochs-while-training-tlt-detectnet-v2-mobilenet-v2/155116>\
**Category:** TAO Toolkit\
**Created:** [September 23, 2020, 10:45am UTC](https://forums.developer.nvidia.com/t/0-map-over-50-epochs-while-training-tlt-detectnet-v2-mobilenet-v2/155116 "2020-09-23T10:45:06Z")\
**Posts on this page:** 1\
**Showing post:** 6

<div class="post-metadata">

**Author:** ![sukyoung.cho](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@sukyoung.cho](https://forums.developer.nvidia.com/u/sukyoung.cho)\
**Post date:** [September 24, 2020, 4:10am UTC](https://forums.developer.nvidia.com/t/0-map-over-50-epochs-while-training-tlt-detectnet-v2-mobilenet-v2/155116/6 "2020-09-24T04:10:11Z")

</div>

Yes I rezsized images / labels to 960x544. I have used almost same config file with the config file which i used to train TrafficCamModel (one that I successfully trained and exported).

## New update is that, I could train mobileNet\_v2 using the config file which this guy uploaded ([Retraining with pretrained tlt models - #7 by harryhsl8c](https://forums.developer.nvidia.com/t/retraining-with-pretrained-tlt-models/110030/7)). I wonder why the mAP after 200 epoch is only about 69%. Even TrafficCamNet resulted in a similar mAP result, which is a way lower than the benchmark you provided and what I have expected. It would be appreciated if you could tell me what could be the possible reason behind low mAP issue.

Not sure if this is what you mean by one label file.

# Label File 1

car 0 3 -1 46 358 195 449 -1 -1 -1 -1 -1 -1 -1  
car 0 3 -1 366 291 426 335 -1 -1 -1 -1 -1 -1 -1  
car 0 3 -1 418 238 451 262 -1 -1 -1 -1 -1 -1 -1  
car 0 3 -1 478 238 511 258 -1 -1 -1 -1 -1 -1 -1  
car 0 3 -1 477 216 499 233 -1 -1 -1 -1 -1 -1 -1  
car 0 3 -1 379 214 423 245 -1 -1 -1 -1 -1 -1 -1

# Lable File 2

kitti\_config {  
root\_directory\_path: “/workspace/data/day\_all”  
image\_dir\_name: “day\_all\_jpg”  
label\_dir\_name: “day\_all\_txt”  
image\_extension: “.jpg”  
partition\_mode: “random”  
num\_partitions: 2

# split percentage of train\_validation

val\_split: 30

# how many make tfrecord shards?

num\_shards: 1 }

---

_[View the full topic](https://forums.developer.nvidia.com/t/0-map-over-50-epochs-while-training-tlt-detectnet-v2-mobilenet-v2/155116)._