squeezenas_train

Introduction:

This is a training script for the SqueezeNAS models on the RailSem19.

How to start with the script:

In order to start with the script you have to first download the SqueezeNAS repository and then download and exctract this repository in the same folder where you have your SqueezeNAS scripts. There is a helper repository HERE which is also needed to be downloaded into the same directory. Our script is written specifically for RailSem so you can also download this from the provided link into the directory and start with the training.

Training:

Using rs_train.py you can train your models. By default the training is performed on the pretrained (on CityScapes dataset) weights of SqueezeNAS models which are located in the weights directory of that repository.

Inference:

1. Latency: We are measuring the inference time in rs_latency.py using cuda events.
2. IoU-Score: Using rs_iou_visual.py you can calculate the IoU-Values of the predictions and also visualize the results.

Results:

This is our results for evaluation of our trained models on RailSem19 dataset:

Architecture	mIOU	Latency Values(ms)
SqueezeNAS MAC Small	36.62	34.36
SqueezeNAS MAC Large	41.48	76.23
SqueezeNAS MAC XLarge	44	178.78
SqueezeNAS LAT Small	40.49	39.79
SqueezeNAS LAT Large	42.12	116.68
SqueezeNAS LAT XLarge	46.76	180.18

Here is an example of the prediction made by lat_xlarge model: