running this demo with plot field set to true (like it is in the demo) imageio is required as an additional dependency.
If you're in a hurry do the following to run the demo:
NOTE: DUE TO MULTIPROCESSING PACKAGE THE CURRENT IMPLEMENTATION ONLY WORKS ON UNIX/LINUX OPERATING SYSTEMS TO RUN ON WINDOWS FOLLOW THIS GUIDE
conda create --name demo_test python=3.6
conda activate demo_test
To create and activate a new environment. If conda activate demo_test fails please try source activate demo_test.
conda install scikit-image scipy numpy matplotlib
conda install -c conda-forge pydoe gpy
To install the required packages
Then in the parent folder of AutoDot:
python
To open a shell
from AutoDot.tune import tune_with_playground_from_file
tune_with_playground_from_file('AutoDot/mock_device_demo_config.json')To run (if "plot":true you must close plots manually)
Outputs: Plots (gifs) are saved in the folder 'mock_device_demo' and raw data it saved in a pickle file called 'tuning.pkl'. To open and read the pickle file run the following.
import pickle
with open("mock_device_demo/tuning.pkl","rb") as handle:
data_dict = pickle.load(handle)data_dict then contains all important output, for example data_dict['vols_pinchoff'] contains a list of all observed pinch off values and boundary points.
If you're not in a hurry do the following sections might be helpful
Provided already is a demo config json that will run 50 iterations of the algorithm as used in the paper with mock versions of the investigation stage steps in the 3D environment plotted below. The environment is specified to be 3D using the "ndim" field and the shape of the environment is specified using primitive shapes that are defined in shapes.py. The example used in the demo combines a standard Crosstalk_box and Leakage shape (note leakage does not refer to leaky gates but refers to alterative undesired current pathways from source to drain). A 3D Crosstalk_box is the type of shape one should expect to observe for three barrier gates. A Leakage demonstrates the pruning method and would be expected if one gate was required to define the current path (like a large top gate or V_1 in the original paper).
The distribution of desired features can also be controlled by using the config. In the supplied demo the "investigation" field contains the following:
"diag_trace":{
"func":"mock_measurement",
"condition":"mock_peak_check",
"a":[0,0],
"b":[-1000,-1000],
"verbose":true},
Fields "a" and "b" define a box in gates 2 and 3 inside of which mock_peak_check will return true. For mock_score_func a Euclidian distance to "target" is used to define a score. Custom condition functions and measurement functions can be created by adding to condition_functions.py and measurement_functions.py. To see the configuration of a real experiments investigation stage see pygor_demo_config.json.
If the "plot" flag in "playground" is set to True the algorithm will first attempt to treat the environment as a 3D space and extract an isosurface (note this could not be done on a real device due to the time required to measure such a plot):
During running the algorithm will print many outputs. An example output for an iteration is:
============### ITERATION 29 ###============
GPR: True GPC: True prune: True GPR1: True GPC1: True Optim: False
START
True
0.19766634805822517
Score thresh: 0.36156802062375365
STOP
dvec pinches: [ True True True]
There are 30 training examples for model 0 and 26 are positive
There are 26 training examples for model 1 and 17 are positive
conditional_idx: 2
vols_pinchoff: [-1146.48524481 -813.5532 -647.01667969]
detected: True
r_vals: 1285.0125867654601
Below is a breakdown of what each line means:
GPR: True GPC: True prune: True GPR1: True GPC1: True Optim: False
Denotes whether the gpr training, gpc training, gpr selection, gpc selection and gp optimisation is running for a given iteration
START
Brownian motion sampling has started running in parallel with algorithm
True
Result of mock_peak_check (only printed if "verbose":true)
0.19766634805822517
Result of mock_score_func.
Score thresh: 0.36156802062375365
Current value of the score threshold.
STOP
Brownian motion sampling has stopped running in parallel with algorithm
dvec pinches: [ True True True]
If pruning is active this shows which gates are showing a pinch off
There are 30 training examples for model 0 and 26 are positive
There are 26 training examples for model 1 and 17 are positive
Shows how many positives and training examples there are for the conditional gpc models (in the paper model 0 models if a pinch off will be observed and model 1 if a coulomb peak will be observed)
conditional_idx: 2
Denotes how many stages in the investigation stage were passed. In the paper 0: no pinch, 1: no coulomb peak, 2: coulomb peak but poor regime, 3: nice regime. (NOTE only printed if "conditional_idx" included in top level "verbose")
vols_pinchoff: [-1146.48524481 -813.5532 -647.01667969]
Point of pinch off in voltage space. (NOTE only printed if "vols_pinchoff" included in top level "verbose")
detected: True
If pinch off was observed. (NOTE only printed if "detected" included in top level "verbose")
r_vals: 1285.0125867654601
Distance from origin to the pinch off
All results are saved in the save directory ("save_dir") in a file called tuning.pkl. The saved fields are specified in "track". Useful fields that you might want to look at are "vols_pinchoff" (list of all pinch-off's or points where the algorithm hit the boundary), "detected" (if the point is a pinch-off or boundary point), "conditional_idx" (denotes how successful a given iteration was at passing characterisation tests), "r_vals" (distance from origin to the pinch-off/boundary), and "extra_measure" (contains list of dicts that contain all information gathered by the investigation stage including stability diagrams if running on a real device).
If the "plot" flag in "playground" is set to True the algorithm will finally attempt to treat the fit Gaussian process prediction of the hypersurface (gpr) as a 3D space and extract an isosurface. It will also plot the Gaussian process classification (gpc) as the colour map of the gpr. The points used for training of these models will also be plotted:
Finally the improvement of the empirical probability of observing the situation that the gpc models is plotted as well:
