|
|
import sys import operator from os import listdir, system import subprocess import re from collections import defaultdict
from random import randrange from ale_py import ALEInterface, SDL_SUPPORT, Action from PIL import Image from matplotlib import pyplot as plt import cv2 import pickle import queue from dataclasses import dataclass, field
from sklearn.cluster import KMeans, DBSCAN
from enum import Enum
from copy import deepcopy
import numpy as np
import logging logger = logging.getLogger(__name__)
#import readchar
from sample_factory.algo.utils.tensor_dict import TensorDict from query_sample_factory_checkpoint import SampleFactoryNNQueryWrapper
import time
tempest_binary = "/home/spranger/projects/tempest-devel/ranking_release/bin/storm" rom_file = "/home/spranger/research/Skiing/env/lib/python3.10/site-packages/AutoROM/roms/skiing.bin"
def tic(): import time global startTime_for_tictoc startTime_for_tictoc = time.time()
def toc(): import time if 'startTime_for_tictoc' in globals(): return time.time() - startTime_for_tictoc
class Verdict(Enum): INCONCLUSIVE = 1 GOOD = 2 BAD = 3
verdict_to_color_map = {Verdict.BAD: "200,0,0", Verdict.INCONCLUSIVE: "40,40,200", Verdict.GOOD: "00,200,100"}
def convert(tuples): return dict(tuples)
@dataclass(frozen=True) class State: x: int y: int ski_position: int #velocity: int def default_value(): return {'action' : None, 'choiceValue' : None}
@dataclass(frozen=True) class StateValue: ranking: float choices: dict = field(default_factory=default_value)
def exec(command,verbose=True): if verbose: print(f"Executing {command}") system(f"echo {command} >> list_of_exec") return system(command)
num_tests_per_cluster = 50 factor_tests_per_cluster = 0.2 num_ski_positions = 8
def input_to_action(char): if char == "0": return Action.NOOP if char == "1": return Action.RIGHT if char == "2": return Action.LEFT if char == "3": return "reset" if char == "4": return "set_x" if char == "5": return "set_vel" if char in ["w", "a", "s", "d"]: return char
def saveObservations(observations, verdict, testDir): testDir = f"images/testing_{experiment_id}/{verdict.name}_{testDir}_{len(observations)}" if len(observations) < 20: logger.warn(f"Potentially spurious test case for {testDir}") testDir = f"{testDir}_pot_spurious" exec(f"mkdir {testDir}", verbose=False) for i, obs in enumerate(observations): img = Image.fromarray(obs) img.save(f"{testDir}/{i:003}.png")
ski_position_counter = {1: (Action.LEFT, 40), 2: (Action.LEFT, 35), 3: (Action.LEFT, 30), 4: (Action.LEFT, 10), 5: (Action.NOOP, 1), 6: (Action.RIGHT, 10), 7: (Action.RIGHT, 30), 8: (Action.RIGHT, 40) }
#def run_single_test(ale, nn_wrapper, x,y,ski_position, velocity, duration=50): def run_single_test(ale, nn_wrapper, x,y,ski_position, duration=50): #print(f"Running Test from x: {x:04}, y: {y:04}, ski_position: {ski_position}", end="") #testDir = f"{x}_{y}_{ski_position}_{velocity}" testDir = f"{x}_{y}_{ski_position}" for i, r in enumerate(ramDICT[y]): ale.setRAM(i,r) ski_position_setting = ski_position_counter[ski_position] for i in range(0,ski_position_setting[1]): ale.act(ski_position_setting[0]) ale.setRAM(14,0) ale.setRAM(25,x) ale.setRAM(14,180) # TODO
all_obs = list() speed_list = list() resized_obs = cv2.resize(ale.getScreenGrayscale(), (84,84), interpolation=cv2.INTER_AREA) for i in range(0,4): all_obs.append(resized_obs) for i in range(0,duration-4): resized_obs = cv2.resize(ale.getScreenGrayscale(), (84,84), interpolation=cv2.INTER_AREA) all_obs.append(resized_obs) if i % 4 == 0: stack_tensor = TensorDict({"obs": np.array(all_obs[-4:])}) action = nn_wrapper.query(stack_tensor) ale.act(input_to_action(str(action))) else: ale.act(input_to_action(str(action))) speed_list.append(ale.getRAM()[14]) if len(speed_list) > 15 and sum(speed_list[-6:-1]) == 0: saveObservations(all_obs, Verdict.BAD, testDir) return Verdict.BAD saveObservations(all_obs, Verdict.GOOD, testDir) return Verdict.GOOD
def computeStateRanking(mdp_file, iteration): logger.info("Computing state ranking") tic() try: command = f"{tempest_binary} --prism {mdp_file} --buildchoicelab --buildstateval --build-all-labels --prop 'Rmax=? [C <= 1000]'" result = subprocess.run(command, shell=True, check=True) print(result) except Exception as e: print(e) sys.exit(-1) exec(f"mv action_ranking action_ranking_{iteration:03}") logger.info(f"Computing state ranking - DONE: took {toc()} seconds")
def fillStateRanking(file_name, match=""): logger.info(f"Parsing state ranking, {file_name}") tic() state_ranking = dict() try: with open(file_name, "r") as f: file_content = f.readlines() for line in file_content: if not "move=0" in line: continue ranking_value = float(re.search(r"Value:([+-]?(\d*\.\d+)|\d+)", line)[0].replace("Value:","")) if ranking_value <= 0.1: continue stateMapping = convert(re.findall(r"([a-zA-Z_]*[a-zA-Z])=(\d+)?", line)) #print("stateMapping", stateMapping) choices = convert(re.findall(r"[a-zA-Z_]*(left|right|noop)[a-zA-Z_]*:(-?\d+\.?\d*)", line)) choices = {key:float(value) for (key,value) in choices.items()} #print("choices", choices) #print("ranking_value", ranking_value) #state = State(int(stateMapping["x"]), int(stateMapping["y"]), int(stateMapping["ski_position"]), int(stateMapping["velocity"])) state = State(int(stateMapping["x"]), int(stateMapping["y"]), int(stateMapping["ski_position"])) value = StateValue(ranking_value, choices) state_ranking[state] = value logger.info(f"Parsing state ranking - DONE: took {toc()} seconds") return state_ranking
except EnvironmentError: print("Ranking file not available. Exiting.") toc() sys.exit(1) except: toc()
def createDisjunction(formulas): return " | ".join(formulas)
def clusterFormula(cluster): formula = "" #states = [(s[0].x,s[0].y, s[0].ski_position, s[0].velocity) for s in cluster] states = [(s[0].x,s[0].y, s[0].ski_position) for s in cluster] skiPositionGroup = defaultdict(list) for item in states: skiPositionGroup[item[2]].append(item)
first = True #todo add velocity here for skiPosition, group in skiPositionGroup.items(): formula += f"ski_position={skiPosition} & (" yPosGroup = defaultdict(list) for item in group: yPosGroup[item[1]].append(item) for y, y_group in yPosGroup.items(): if first: first = False else: formula += " | " sorted_y_group = sorted(y_group, key=lambda s: s[0]) formula += f"( y={y} & (" current_x_min = sorted_y_group[0][0] current_x = sorted_y_group[0][0] x_ranges = list() for state in sorted_y_group[1:-1]: if state[0] - current_x == 1: current_x = state[0] else: x_ranges.append(f" ({current_x_min}<= x & x<={current_x})") current_x_min = state[0] current_x = state[0] x_ranges.append(f" ({current_x_min}<= x & x<={sorted_y_group[-1][0]})") formula += " | ".join(x_ranges) formula += ") )" formula += ")" return formula
def createBalancedDisjunction(indices, name): #logger.info(f"Creating balanced disjunction for {len(indices)} ({indices}) formulas") if len(indices) == 0: return f"formula {name} = false;\n" else: while len(indices) > 1: indices_tmp = [f"({indices[i]} | {indices[i+1]})" for i in range(0,len(indices)//2)] if len(indices) % 2 == 1: indices_tmp.append(indices[-1]) indices = indices_tmp disjunction = f"formula {name} = " + " ".join(indices) + ";\n" return disjunction
def createUnsafeFormula(clusters): label = "label \"Unsafe\" = Unsafe;\n" formulas = "" indices = list() for i, cluster in enumerate(clusters): formulas += f"formula Unsafe_{i} = {clusterFormula(cluster)};\n" indices.append(f"Unsafe_{i}") return formulas + "\n" + createBalancedDisjunction(indices, "Unsafe") + label
def createSafeFormula(clusters): label = "label \"Safe\" = Safe;\n" formulas = "" indices = list() for i, cluster in enumerate(clusters): formulas += f"formula Safe_{i} = {clusterFormula(cluster)};\n" indices.append(f"Safe_{i}")
return formulas + "\n" + createBalancedDisjunction(indices, "Safe") + label
def updatePrismFile(newFile, iteration, safeStates, unsafeStates): logger.info("Creating next prism file") tic() initFile = f"{newFile}_no_formulas.prism" newFile = f"{newFile}_{iteration:03}.prism" exec(f"cp {initFile} {newFile}", verbose=False) with open(newFile, "a") as prism: prism.write(createSafeFormula(safeStates)) prism.write(createUnsafeFormula(unsafeStates))
logger.info(f"Creating next prism file - DONE: took {toc()} seconds")
ale = ALEInterface()
#if SDL_SUPPORT: # ale.setBool("sound", True) # ale.setBool("display_screen", True)
# Load the ROM file ale.loadROM(rom_file)
with open('all_positions_v2.pickle', 'rb') as handle: ramDICT = pickle.load(handle) y_ram_setting = 60 x = 70
nn_wrapper = SampleFactoryNNQueryWrapper()
experiment_id = int(time.time()) init_mdp = "velocity_safety" exec(f"mkdir -p images/testing_{experiment_id}", verbose=False)
markerSize = 1
imagesDir = f"images/testing_{experiment_id}"
def drawOntoSkiPosImage(states, color, target_prefix="cluster_", alpha_factor=1.0): markerList = {ski_position:list() for ski_position in range(1,num_ski_positions + 1)} for state in states: s = state[0] #marker = f"-fill 'rgba({color}, {alpha_factor * state[1].ranking})' -draw 'rectangle {s.x-markerSize},{s.y-markerSize} {s.x+markerSize},{s.y+markerSize} '" marker = f"-fill 'rgba({color}, {alpha_factor * state[1].ranking})' -draw 'point {s.x},{s.y} '" markerList[s.ski_position].append(marker) for pos, marker in markerList.items(): command = f"convert {imagesDir}/{target_prefix}_{pos:02}_individual.png {' '.join(marker)} {imagesDir}/{target_prefix}_{pos:02}_individual.png" exec(command, verbose=False)
def concatImages(prefix, iteration): exec(f"montage {imagesDir}/{prefix}_*_individual.png -geometry +0+0 -tile x1 {imagesDir}/{prefix}_{iteration}.png", verbose=False) exec(f"sxiv {imagesDir}/{prefix}_{iteration}.png&", verbose=False)
def drawStatesOntoTiledImage(states, color, target, source="images/1_full_scaled_down.png", alpha_factor=1.0): """
Useful to draw a set of states, e.g. a single cluster """
markerList = {1: list(), 2:list(), 3:list(), 4:list(), 5:list(), 6:list(), 7:list(), 8:list()} logger.info(f"Drawing {len(states)} states onto {target}") tic() for state in states: s = state[0] marker = f"-fill 'rgba({color}, {alpha_factor * state[1].ranking})' -draw 'rectangle {s.x-markerSize},{s.y-markerSize} {s.x+markerSize},{s.y+markerSize} '" markerList[s.ski_position].append(marker) for pos, marker in markerList.items(): command = f"convert {source} {' '.join(marker)} {imagesDir}/{target}_{pos:02}_individual.png" exec(command, verbose=False) exec(f"montage {imagesDir}/{target}_*_individual.png -geometry +0+0 -tile x1 {imagesDir}/{target}.png", verbose=False) logger.info(f"Drawing {len(states)} states onto {target} - Done: took {toc()} seconds")
def drawClusters(clusterDict, target, iteration, alpha_factor=1.0): for ski_position in range(1, num_ski_positions + 1): source = "images/1_full_scaled_down.png" exec(f"cp {source} {imagesDir}/{target}_{ski_position:02}_individual.png", verbose=False) for _, clusterStates in clusterDict.items(): color = f"{np.random.choice(range(256))}, {np.random.choice(range(256))}, {np.random.choice(range(256))}" drawOntoSkiPosImage(clusterStates, color, target, alpha_factor=alpha_factor) concatImages(target, iteration)
def drawResult(clusterDict, target, iteration): for ski_position in range(1, num_ski_positions + 1): source = "images/1_full_scaled_down.png" exec(f"cp {source} {imagesDir}/{target}_{ski_position:02}_individual.png") for _, (clusterStates, result) in clusterDict.items(): color = "100,100,100" if result == Verdict.GOOD: color = "0,200,0" elif result == Verdict.BAD: color = "200,0,0" drawOntoSkiPosImage(clusterStates, color, target, alpha_factor=0.7) concatImages(target, iteration)
def _init_logger(): logger = logging.getLogger('main') logger.setLevel(logging.INFO) handler = logging.StreamHandler(sys.stdout) formatter = logging.Formatter( '[%(levelname)s] %(module)s - %(message)s') handler.setFormatter(formatter) logger.addHandler(handler)
def clusterImportantStates(ranking, iteration): logger.info(f"Starting to cluster {len(ranking)} states into clusters") tic() #states = [[s[0].x,s[0].y, s[0].ski_position * 10, s[0].velocity * 10, s[1].ranking] for s in ranking] states = [[s[0].x,s[0].y, s[0].ski_position * 30, s[1].ranking] for s in ranking] #kmeans = KMeans(n_clusters, random_state=0, n_init="auto").fit(states) dbscan = DBSCAN(eps=15).fit(states) labels = dbscan.labels_ n_clusters = len(set(labels)) - (1 if -1 in labels else 0) logger.info(f"Starting to cluster {len(ranking)} states into clusters - DONE: took {toc()} seconds with {n_clusters} cluster") clusterDict = {i : list() for i in range(0,n_clusters)} for i, state in enumerate(ranking): if labels[i] == -1: continue clusterDict[labels[i]].append(state) drawClusters(clusterDict, f"clusters", iteration) return clusterDict
if __name__ == '__main__': _init_logger() logger = logging.getLogger('main') logger.info("Starting") n_clusters = 40 testAll = False
safeStates = list() unsafeStates = list() iteration = 0 while True: updatePrismFile(init_mdp, iteration, safeStates, unsafeStates) computeStateRanking(f"{init_mdp}_{iteration:03}.prism", iteration) ranking = fillStateRanking(f"action_ranking_{iteration:03}") sorted_ranking = sorted( (x for x in ranking.items() if x[1].ranking > 0.1), key=lambda x: x[1].ranking) clusters = clusterImportantStates(sorted_ranking, iteration)
if testAll: failingPerCluster = {i: list() for i in range(0, n_clusters)} clusterResult = dict() for id, cluster in clusters.items(): num_tests = int(factor_tests_per_cluster * len(cluster)) num_tests = 1 logger.info(f"Testing {num_tests} states (from {len(cluster)} states) from cluster {id}") randomStates = np.random.choice(len(cluster), num_tests, replace=False) randomStates = [cluster[i] for i in randomStates]
verdictGood = True for state in randomStates: x = state[0].x y = state[0].y ski_pos = state[0].ski_position #velocity = state[0].velocity result = run_single_test(ale,nn_wrapper,x,y,ski_pos, duration=50) #result = run_single_test(ale,nn_wrapper,x,y,ski_pos, velocity, duration=50) result = Verdict.BAD # TODO REMOVE ME!!!!!!!!!!!!!! if result == Verdict.BAD: if testAll: failingPerCluster[id].append(state) else: clusterResult[id] = (cluster, Verdict.BAD) verdictGood = False unsafeStates.append(cluster) break if verdictGood: clusterResult[id] = (cluster, Verdict.GOOD) safeStates.append(cluster) logger.info(f"Iteration: {iteration:03} -\tSafe Results : {sum([len(c) for c in safeStates])} -\tUnsafe Results:{sum([len(c) for c in unsafeStates])}") if testAll: drawClusters(failingPerCluster, f"failing", iteration)
drawResult(clusterResult, "result", iteration) iteration += 1
|