def generatePlot(exp_paths):
exp = ExperimentModel.load(exp_paths[0])
results = loadResults(exp, 'step_return.csv')
results_dict = splitOverParameter(results, 'tile-coder')
# two new streams over results
mine = results_dict['mine']
tiles3 = results_dict['tiles3']
# figure out what values of "tiles" we swept
tiles = splitOverParameter(mine, 'tiles').keys()
# figure out what values of "tilings" we swept
tilings = splitOverParameter(mine, 'tilings').keys()
f, axes = plt.subplots(len(tiles), len(tilings))
# create an empty (2, tiles, tilings) shape array
data = createEmptyList([2, len(tiles), len(tilings)])
for tiles_idx, num_tiles in enumerate(tiles):
for tilings_idx, num_tilings in enumerate(tilings):
mine_result = whereParametersEqual(mine, { 'tiles': num_tiles, 'tilings': num_tilings })
mine_result = getBest(mine_result, prefer='big')
tiles3_result = whereParametersEqual(tiles3, { 'tiles': num_tiles, 'tilings': num_tilings })
tiles3_result = getBest(tiles3_result, prefer='big')
mine_data = [np.mean(curve) for curve in mine_result.load()]
tiles3_data = [np.mean(curve) for curve in tiles3_result.load()]
data[0][tiles_idx][tilings_idx] = mine_data
data[1][tiles_idx][tilings_idx] = tiles3_data
min_perf = np.min(data)
max_perf = np.max(data)
for rep, rep_name in enumerate(['Mine', 'Tiles3']):
for tiles_idx, num_tiles in enumerate(tiles):
for tilings_idx, num_tilings in enumerate(tilings):
performance = data[rep][tiles_idx][tilings_idx]
color = 'blue' if rep == 0 else 'red'
kde = gaussian_kde(performance)
lo = 0.95 * min_perf
hi = 1.05 * max_perf
dist_space = np.linspace(lo, hi, FIDELITY)
dist = kde(dist_space)
# dist = minMaxScale(kde(dist_space)) * DIST_HEIGHT
axes[tiles_idx][tilings_idx].plot(dist_space, dist, label=rep_name, linewidth=2.0, color=color)
axes[tiles_idx][tilings_idx].fill_between(dist_space, np.zeros(FIDELITY), dist, color=color, alpha=0.2)
axes[tiles_idx][tilings_idx].set_xlim((0.95 * min_perf, 1.05 * max_perf))
# axes[tiles_idx][tilings_idx].set_xlabel('Reward')
title = f'({num_tiles} tiles, {num_tilings} tilings)'
axes[tiles_idx][tilings_idx].set_title(title)
plt.legend()
return f, axes
def generatePlot(ax, exp_paths):
for alg in ALG_ORDER:
exp_path = findExpPath(exp_paths, alg)
exp = ExperimentModel.load(exp_path)
results = loadResults(exp, 'step_return.h5')
plotBest(results, ax, {
'color': basicControlColors.get(alg),
'label': alg,
'prefer': 'big',
})
def generatePlot(ax, exp_paths, bestBy):
reducer = getCurveReducer(bestBy)
for alg in ALG_ORDER:
exp_path = findExpPath(exp_paths, alg)
exp = ExperimentModel.load(exp_path)
results = loadResults(exp, 'returns.csv')
best = getBest(results, reducer)
plot(best, ax, {
'label': alg,
'color': colors[alg],
'width': 0.75,
})
def generatePlot(ax, exp_paths):
exp = ExperimentModel.load(exp_paths[0])
results = loadResults(exp, 'step_return.csv')
results_dict = splitOverParameter(results, 'tile-coder')
# two new streams over results
mine = results_dict['mine']
tiles3 = results_dict['tiles3']
# figure out what values of "tiles" we swept
tiles = splitOverParameter(mine, 'tiles').keys()
# figure out what values of "tilings" we swept
tilings = splitOverParameter(mine, 'tilings').keys()
for num_tiles in tiles:
for num_tilings in tilings:
mine_result = whereParametersEqual(mine, {
'tiles': num_tiles,
'tilings': num_tilings
})
mine_result = getBest(mine_result, prefer='big')
tiles3_result = whereParametersEqual(tiles3, {
'tiles': num_tiles,
'tilings': num_tilings
})
tiles3_result = getBest(tiles3_result, prefer='big')
d_curves = []
mine_curves = mine_result.load()
tiles3_curves = tiles3_result.load()
min_len = min(len(mine_curves), len(tiles3_curves))
print(min_len)
for i in range(min_len):
d = mine_curves[i] - tiles3_curves[i]
d_curves.append(d)
mean = np.mean(d_curves, axis=0)
stderr = np.std(d_curves, axis=0, ddof=1) / np.sqrt(len(d_curves))
lineplot(ax,
mean,
stderr=stderr,
label=f'({num_tiles}, {num_tilings})')
ax.set_xlim(0, 500)
ax.axhline(0, color='black', linestyle='--', alpha=0.4)
ax.set_ylabel(f'd = Mine - Tiles3 \n Bigger is better')
plt.legend()
def printStats(exp_paths, metric):
print(f'-------------{metric}-------------')
exp = ExperimentModel.load(exp_paths[0])
results = loadResults(exp, f'{metric}.csv')
results_dict = splitOverParameter(results, 'tile-coder')
# two new streams over results
mine = results_dict['mine']
tiles3 = results_dict['tiles3']
# figure out what values of "tiles" we swept
tiles = splitOverParameter(mine, 'tiles').keys()
# figure out what values of "tilings" we swept
tilings = splitOverParameter(mine, 'tilings').keys()
for num_tiles in tiles:
for num_tilings in tilings:
mine_results = list(
whereParametersEqual(mine, {
'tiles': num_tiles,
'tilings': num_tilings
}))
tiles3_results = list(
whereParametersEqual(tiles3, {
'tiles': num_tiles,
'tilings': num_tilings
}))
mine_means = []
tiles3_means = []
# loop over each value of alpha
# this way we just get 3x as many samples of timing
for i in range(len(mine_results)):
mine_mean = mine_results[i].mean()[0]
tiles3_mean = tiles3_results[i].mean()[0]
mine_means.append(mine_mean)
tiles3_means.append(tiles3_mean)
mine_mean = np.mean(mine_means)
tiles3_mean = np.mean(tiles3_means)
# TODO: this is covering up a bug in results. Rerun results
if metric == 'feature_utilization':
mine_mean = mine_mean / (num_tilings * num_tiles**2)
tiles3_mean = tiles3_mean / (num_tilings * num_tiles**2)
print(
f'({num_tiles}, {num_tilings}) -- {mine_mean}, {tiles3_mean}')
def generatePlot(exp_paths):
exp = ExperimentModel.load(exp_paths[0])
results = loadResults(exp, 'step_return.csv')
results_dict = splitOverParameter(results, 'tile-coder')
# two new streams over results
mine = results_dict['mine']
tiles3 = results_dict['tiles3']
# figure out what values of "tiles" we swept
tiles = splitOverParameter(mine, 'tiles').keys()
# figure out what values of "tilings" we swept
tilings = splitOverParameter(mine, 'tilings').keys()
f, axes = plt.subplots(len(tiles), len(tilings))
for tiles_idx, num_tiles in enumerate(tiles):
for tilings_idx, num_tilings in enumerate(tilings):
mine_result = whereParametersEqual(mine, {
'tiles': num_tiles,
'tilings': num_tilings
})
mine_result = getBest(mine_result, prefer='big')
tiles3_result = whereParametersEqual(tiles3, {
'tiles': num_tiles,
'tilings': num_tilings
})
tiles3_result = getBest(tiles3_result, prefer='big')
plotBest(mine_result,
axes[tiles_idx][tilings_idx],
color='blue',
label='Mine')
plotBest(tiles3_result,
axes[tiles_idx][tilings_idx],
color='red',
label='Tiles3')
axes[tiles_idx][tilings_idx].set_title(
f'Tiles: {num_tiles} Tilings: {num_tilings}')
axes[0][0].legend()
return f, axes
def gatherMissing(experiment_paths, runs, groupSize, cores, total_hours):
out = {}
approximate_cost = np.zeros(2)
for path in experiment_paths:
exp = Experiment.load(path)
indices = detectMissingIndices(exp, runs, 'step_return.h5')
indices = sorted(indices)
out[path] = indices
approximate_cost += estimateUsage(indices, groupSize, cores,
total_hours)
# figure out how many indices to expect
size = exp.numPermutations() * runs
# log how many are missing
print(path, f'{len(indices)} / {size}')
return out, approximate_cost
# prints a progress bar
def printProgress(size, it):
for i, _ in enumerate(it):
print(f'{i + 1}/{size}', end='\r')
if i - 1 == size:
print()
yield _
# ----------------
# Scheduling logic
# ----------------
for path in experiment_paths:
print(path)
# load the experiment json file
exp = Experiment.load(path)
# load the slurm config file
slurm = Slurm.fromFile(slurm_path)
if exp.agent in SLOW_ALGS:
slurm.sequential = 1
# figure out how many indices to use
size = exp.numPermutations() * runs
paths = listResultsPaths(exp, runs)
res_path = first(paths)
data = []
data_path = f'{res_path}/returns.csv'
if os.path.exists(data_path):
if len(sys.argv) < 2:
print('run again with:')
print('python3 src/main.py <path/to/description.json> <idx>')
exit(1)
# try to detect if we are on a Cedar server
prod = 'cdr' in socket.gethostname()
# if we are local, then turn on info level logging
# otherwise, keep the console clear to improve performance and decrease clutter
if not prod:
logging.basicConfig(level=logging.INFO)
torch.set_num_threads(1)
exp = ExperimentModel.load(sys.argv[1])
idx = int(sys.argv[2])
max_steps = exp.max_steps
run = exp.getRun(idx)
collector = Collector()
# set random seeds accordingly
np.random.seed(run)
torch.manual_seed(run)
Problem = getProblem(exp.problem)
problem = Problem(exp, idx)
agent = problem.getAgent(collector)
