Training data for Time Contrastive Network experiments

To train a contrastive network, you need alike and dissimilar examples.

Here are some I am releasing from a work in progress, Dobot TCN.

v2/ contains images of a single-axis servo "robot arm" rotating to randomly chosen locations, accompanied by a photo from two cameras (called rpi and rpzw).

There are four folders, corresponding to four different camera positions.

Each folder contains 200 observations.

Each observation consists of two jpeg images and one .json file.

For example:

data/position1$ file 40-rpi.jpg
40-rpi.jpg: JPEG image data, Exif standard: [TIFF image data, big-endian, direntries=10, height=0, 
	manufacturer=RaspberryPi, model=RP_ov5647, xresolution=156, yresolution=164, resolutionunit=2, 
	datetime=2020:08:23 20:49:10, width=0], baseline, precision 8, 640x480, components 3
data/position1$ file 40-rpzw.jpg
40-rpzw.jpg: JPEG image data, Exif standard: [TIFF image data, big-endian, direntries=10, height=0, 
	manufacturer=RaspberryPi, model=RP_imx219, xresolution=156, yresolution=164, resolutionunit=2, 
	datetime=2020:08:23 20:49:11, width=0], baseline, precision 8, 640x480, components 3
data/position2$ cat 40-joints.json
{"from": {"1": 5}, "to": {"1": -59}, "goal": {"1": -60}, 
	"exp_config": {"SWEEP_RANGE": 90, "AXIS_REPS": 2, "N_EXAMPLES": 200, "JOINT_RANGES": {"1": [-60, 60], "2": [-60, 60]}}}

The to value here specifies the position of the joint are shown in the video.

from specifies where the joint was positioned before the move and goal is where we wanted it to go. I don't think either are interesting from the standpoint of the actual observation.

exp_config specifies the configuration of this experiment. The v2/ dataset should all be the same in this regard. (The unreleased v1/ was improperly controlled.)

In addition, ALL.json in each folder gives you all the data at once.

Members of the AI Hackers Discord server requested access to the training data I was attempting to use for my experiments

That's up to you. The joint JSON should give you enough if you want to try to train an image regression model.