Open yhalk opened 7 years ago
Pilot meeting II - Aug 14
Attendees : Sen,Boris,Iris,Ioannis
General updates:
The meeting started with feedback on the preparation tasks each member had undertook. Iris and Ioannis had worked on the German Traffic sign dataset, and trained a simple CNN in Keras. There was some discussion regarding the pretrained models in Keras and how they can be used/finetuned on a new dataset (probably to be continued next time on a blackboard). Boris experimented with Faster RCNN, a model that identifies objects' locations in the image and classifies them.
Discussion then revolved around the task that the team should try for the current week. It was decided to keep working on Boris' experiments and after training FRCNN on the German Traffic dataset, try to put it on the Jetson (Teun's input needed)
However, it was stressed that at this point team members should become familiar with the tools and experiment with model training procedure.
Meeting III - Aug 16
Attendees : Sen,Boris,Teun,Ioannis
Teun was brought up to date with organizing arrangements and task, and suggested the YOLO detector and one-shot learning as possible ideas to look into.
Boris gave updates on his progress with Faster RCNN.
Next meeting on Monday and in the meantime joint work will be organized through Slack/fb channels.
Meeting IV - Aug 22
Attendees : Sen,Boris,Teun,Iris,Ioannis
The need to identify subtasks once we know the main task was stressed. Furthermore, it is important to be able to connect to the board for debugging; best way by ssh if Jetson and laptops are connected to the same router.
Alternatively, tried setting up an access point on the Jetson but did not work - Ubuntu froze - so a usb WiFi dongle might be needed (+ dongle with multiple USB ports).
Then Boris' model was tried on the Jetson and inference was carried out on a single image, but execution was too slow. Afterwards TF failed to load with cuDNN-related error. Also tried to get images from the onboard camera and ROS to feed to the model - Iris would write a script to do so.
Current work is focused on inference on the Jetson and having it control the Turtlebot.
Pilot meeting - Aug 9
Attendees : Sen,Boris,Iris,Ioannis
The discussion started with possible topics that would be asked for during the challenge. Since VW is organizing it is highly likely that the concept will be autonomous driving, and as NVIDIA is actively involved, vision and therefore cameras will be definitely available. Possibly laser scanners too, but it was pointed out that the nature of the data such sensors provide is not suited for deep learning.
Identified topics were again:
There were concerns about whether we will have multiple tasks (even not related to vehicles) and we will be asked to choose. In that case we should probably opt for a vehicle-related task since we will be preparing in that context.
Regarding the preparation during run-up, it was agreed that the task at the challenge is not something to worry about too much at the moment. Preparation should focus around more basic things, especially given the time constraints of the team.
Time constraints and expectations of team members were also discussed. The availability of the team members so far is:
Regarding expectations, the shared feeling was that the main motivation is getting maximum knowledge about and experience on deep learning methods. Given the little available time for preparation and lack of background knowledge on deep learning specifically for autonomous vehicles it was suggested that we probably cannot expect a top result. Nevertheless, we should try to do our best since we decided to invest time in the first place (Teun's input on expectations from the challenge has to be heard in the next meeting).
The meeting then progressed to more practical issues.
Regarding equipment, since Iris was about to submit specifications for her laptop, she agreed to get one that will be suitable for use at the challenge (with good GPU). Boris also suggested asking Helen Hastie to borrow a VR laptop from her lab. She will be contacted on that, however she's on leave until 14/8 (so is Tessa Berg).
Next week will be the major preparation period for the whole team as we will be able to meet all together.The task has yet to be fully defined, but should cover data preprocessing and preparation for the network, deployment of the network on the Jetson and testing.
Everyone who can prepare from today, should install Keras, TF, select a dataset from those suggested by Sen and get a network running on it. It was also pointed out that we should get acquainted with TF/Keras APIs for established datasets, e.g. ImageNet, in case we need to use one of those.
The meeting concluded with the suggestion to meet on Monday to define the task and schedule. Since Teun is not available on Monday the schedule will have to be revised on Tuesday.