diff --git a/rom_evaluate.py b/rom_evaluate.py index 726601d..5b6907a 100644 --- a/rom_evaluate.py +++ b/rom_evaluate.py @@ -70,24 +70,9 @@ def exec(command,verbose=True): 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 model_to_actual(ski_position): - if ski_position == 1: - return 1 - elif ski_position in [2,3]: - return 2 - elif ski_position in [4,5]: - return 3 - elif ski_position in [6,7]: - return 4 - elif ski_position in [8,9]: - return 5 - elif ski_position in [10,11]: - return 6 - elif ski_position in [12,13]: - return 7 - elif ski_position == 14: - return 8 def input_to_action(char): if char == "0": @@ -117,9 +102,20 @@ def drawImportantStates(important_states): command = f"convert images/1_full_scaled_down.png {' '.join(draw_commands[i])} first_try_{i:02}.png" exec(command) +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, duration=200): #print(f"Running Test from x: {x:04}, y: {y:04}, ski_position: {ski_position}", end="") + 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] @@ -134,8 +130,9 @@ def run_single_test(ale, nn_wrapper, x,y,ski_position, duration=200): first_action_set = False first_action = 0 for i in range(0,duration): - resized_obs = cv2.resize(ale.getScreenGrayscale() , (84,84), interpolation=cv2.INTER_AREA) - all_obs.append(resized_obs) + resized_obs = cv2.resize(ale.getScreenGrayscale(), (84,84), interpolation=cv2.INTER_AREA) + for i in range(0,4): + all_obs.append(resized_obs) if len(all_obs) >= 4: stack_tensor = TensorDict({"obs": np.array(all_obs[-4:])}) action = nn_wrapper.query(stack_tensor) @@ -147,9 +144,10 @@ def run_single_test(ale, nn_wrapper, x,y,ski_position, duration=200): ale.act(Action.NOOP) speed_list.append(ale.getRAM()[14]) if len(speed_list) > 15 and sum(speed_list[-6:-1]) == 0: - return (Verdict.BAD, first_action) - #time.sleep(0.005) - return (Verdict.INCONCLUSIVE, first_action) + saveObservations(all_obs, Verdict.BAD, testDir) + return Verdict.BAD + saveObservations(all_obs, Verdict.GOOD, testDir) + return Verdict.GOOD def optimalAction(choices): return max(choices.items(), key=operator.itemgetter(1))[0] @@ -245,19 +243,16 @@ x = 70 nn_wrapper = SampleFactoryNNQueryWrapper() iteration = 0 -id = int(time.time()) -init_mdp = "velocity" -exec(f"mkdir -p images/testing_{id}") -exec(f"cp 1_full_scaled_down.png images/testing_{id}/testing_0000.png") +experiment_id = int(time.time()) +init_mdp = "velocity_safety" +exec(f"mkdir -p images/testing_{experiment_id}") +exec(f"cp 1_full_scaled_down.png images/testing_{experiment_id}/testing_0000.png") exec(f"cp {init_mdp}.prism {init_mdp}_000.prism") markerSize = 1 #markerList = {1: list(), 2:list(), 3:list(), 4:list(), 5:list(), 6:list(), 7:list(), 8:list()} -def f(n): - if n >= 1.0: - return True - return False +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)} @@ -266,12 +261,13 @@ def drawOntoSkiPosImage(states, color, target_prefix="cluster_", alpha_factor=1. 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 images/testing_{id}/{target_prefix}_{pos:02}.png {' '.join(marker)} images/testing_{id}/{target_prefix}_{pos:02}.png" + command = f"convert {imagesDir}/{target_prefix}_{pos:02}.png {' '.join(marker)} {imagesDir}/{target_prefix}_{pos:02}.png" exec(command, verbose=False) def concatImages(prefix): - exec(f"montage images/testing_{id}/{prefix}_*png -geometry +0+0 -tile x1 images/testing_{id}/{prefix}.png", verbose=False) + exec(f"montage {imagesDir}/{prefix}_*png -geometry +0+0 -tile x1 {imagesDir}/{prefix}.png", verbose=False) + exec(f"sxiv {imagesDir}/{prefix}.png&") def drawStatesOntoTiledImage(states, color, target, source="images/1_full_scaled_down.png", alpha_factor=1.0): """ @@ -285,20 +281,32 @@ def drawStatesOntoTiledImage(states, color, target, source="images/1_full_scaled 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)} images/testing_{id}/{target}_{pos:02}.png" + command = f"convert {source} {' '.join(marker)} {imagesDir}/{target}_{pos:02}.png" exec(command, verbose=False) - exec(f"montage images/testing_{id}/{target}_*png -geometry +0+0 -tile x1 images/testing_{id}/{target}.png", verbose=False) + exec(f"montage {imagesDir}/{target}_*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, 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} images/testing_{id}/{target}_{ski_position:02}.png") + exec(f"cp {source} {imagesDir}/{target}_{ski_position:02}.png") 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, f"clusters") - concatImages("clusters") + drawOntoSkiPosImage(clusterStates, color, target, alpha_factor=alpha_factor) + concatImages(target) +def drawResult(clusterDict, target): + 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}.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) def _init_logger(): logger = logging.getLogger('main') @@ -308,7 +316,7 @@ def _init_logger(): handler.setFormatter(formatter) logger.addHandler(handler) -def clusterImportantStates(ranking, n_clusters=10): +def clusterImportantStates(ranking, n_clusters=40): logger.info(f"Starting to cluster {len(ranking)} states into {n_clusters} cluster") tic() states = [[s[0].x,s[0].y, s[0].ski_position * 10, s[1].ranking] for s in ranking] @@ -324,36 +332,39 @@ if __name__ == '__main__': _init_logger() logger = logging.getLogger('main') logger.info("Starting") + n_clusters = 40 + testAll = False while True: #computeStateRanking(f"{init_mdp}_{iteration:03}.prism") ranking = fillStateRanking("action_ranking") sorted_ranking = sorted( (x for x in ranking.items() if x[1].ranking > 0.1), key=lambda x: x[1].ranking) - print(type(sorted_ranking)) - clusters = clusterImportantStates(sorted_ranking) + clusters = clusterImportantStates(sorted_ranking, n_clusters) + + 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 + result = run_single_test(ale,nn_wrapper,x,y,ski_pos, duration=50) + if result == Verdict.BAD: + if testAll: + failingPerCluster[id].append(state) + else: + clusterResult[id] = (cluster, Verdict.BAD) + verdictGood = False + break + if verdictGood: + clusterResult[id] = (cluster, Verdict.GOOD) + if testAll: drawClusters(failingPerCluster, f"failing") + drawResult(clusterResult, "result") - sys.exit(1) - #for i, state in enumerate(sorted_ranking): - # print(state) - # if i % 10 == 0: - # input("") - #print(len(sorted_ranking)) - - """ - for important_state in ranking[-100:-1]: - optimal_choice = optimalAction(important_state[1].choices) - #print(important_state[1].choices, f"\t\tOptimal: {optimal_choice}") - x = important_state[0].x - y = important_state[0].y - ski_pos = model_to_actual(important_state[0].ski_position) - result = run_single_test(ale,nn_wrapper,x,y,ski_pos, duration=50) - #print(f".... {result}") - marker = f"-fill 'rgba({verdict_to_color_map[result[0]],0.7})' -draw 'rectangle {x-markerSize},{y-markerSize} {x+markerSize},{y+markerSize} '" - markerList[ski_pos].append(marker) - populate_fixed_actions(important_state[0], result[1]) - for pos, marker in markerList.items(): - command = f"convert images/testing_{id}/testing_0000.png {' '.join(marker)} images/testing_{id}/testing_{iteration+1:03}_{pos:02}.png" - exec(command, verbose=False) - exec(f"montage images/testing_{id}/testing_{iteration+1:03}_*png -geometry +0+0 -tile x1 images/testing_{id}/{iteration+1:03}.png", verbose=False) - iteration += 1 - """ update_prism_file(f"{init_mdp}_{iteration-1:03}.prism", f"{init_mdp}_{iteration:03}.prism")