def get_metric_data(fp):
#for use with queue and delay data
#sort all metric data from same tsc_id
if not os.path.exists(fp):
assert 0, 'Supplied path ' + str(fp) + ' does not exist.'
tsc_data = {
tsc_id: sorted(os.listdir(fp + '/' + tsc_id))
for tsc_id in os.listdir(fp)
}
sim_runs_data = []
#until all data has been popped
k = list(tsc_data.keys())[0]
while len(tsc_data[k]) > 0:
#get file path for each intersection from same sim run
same_run_data = [
fp + '/' + tsc_id + '/' + tsc_data[tsc_id].pop(0)
for tsc_id in tsc_data
]
same_run_data = [load_data(f) for f in same_run_data]
#sum across time axis, each element of array
#represents the sum of all tsc_id metric
sim_runs_data.append(np.sum(same_run_data, axis=0))
return np.stack(sim_runs_data)
def load_weights(self, path):
path += '.p'
if os.path.exists(path):
weights = load_data(path)
self.set_weights(weights, 'online')
else:
#raise not found exceptions
assert 0, 'Failed to load weights, supplied weight file path '+str(path)+' does not exist.'
def get_folder_data(fp):
#all the travel times can be
#grouped together by extending list
if not os.path.exists(fp):
assert 0, 'Supplied path '+str(fp)+' does not exist.'
data = []
for f in os.listdir(fp):
data.extend(load_data(fp+'/'+f))
return np.array(data)
def load_replays(self):
for _id in self.agent_ids:
replay_fp = self.replay_fp+_id+'.p'
if os.path.isfile(replay_fp):
data = load_data(replay_fp)
rewards = []
for traj in data:
for exp in traj:
rewards.append(abs(exp['r']))
self.exp_replay[_id].append(traj)
# find largest reward to reward normalization
print('mean '+str(np.mean(rewards))+' std '+str(np.std(rewards))+' median '+str(np.median(rewards)))
self.rl_stats[_id]['r_max'] = max(rewards)
print(str(self.idx)+' LARGEST REWARD '+str(self.rl_stats[_id]['r_max']))
print('SUCCESSFULLY LOADED REPLAY FOR '+str(_id))
else:
print('WARNING, tried to load experience replay at '+str(replay_fp)+' but it does not exist, continuing without loading...')
def graph_hyper_params(labels, colours, fp, save_dir):
tsc = os.listdir(fp)
tsc_hp = {}
#get data
for t in tsc:
tsc_fp = fp+t+'/'
data = [ load_data(tsc_fp+f) for f in os.listdir(tsc_fp)]
tsc_hp[t] = np.stack([ [np.mean(d), np.std(d)] for d in data]).T
#create appropriate graph
n = len(tsc)
if n == 1:
f, axes = plt.subplots()
axes = [axes]
else:
nrows = 2
ncols = int(n/nrows) if n%nrows == 0 else int((n+1)/nrows)
f, axes = plt.subplots(nrows=nrows,ncols=ncols)
axes = axes.flat
if n%nrows != 0:
f.delaxes(axes[-1])
XTITLE = 'Mean\nTravel Time '+r"$(s)$"
YTITLE = ('Standard\nDeviation\nTravel Time '+r"$(s)$", 80)
#graph each tsc hyperparemeter
for ax, t, i in zip(axes, tsc, range(len(tsc))):
#order hp performance from low to high
#w.r.t mean+std
mean_data = tsc_hp[t][0]
std_data = tsc_hp[t][1]
data = sorted([ (m+s, m, s) for m,s in zip(mean_data, std_data) ], key = lambda x:x[0] )
data = np.stack([ [d[1], d[2]] for d in data]).T
mean_data = data[0]
std_data = data[1]
#rainbow_colours = mp.cm.rainbow(np.linspace(0, 1, len(mean_data)))
rg_colours = mp.cm.brg(np.linspace(1.0, 0.5, len(mean_data)))
if i%ncols == 0 and i >= len(tsc)/2:
xtitle = XTITLE
ytitle = YTITLE
elif i%ncols == 0:
xtitle = ''
ytitle = YTITLE
elif i >= len(tsc)/2:
xtitle = XTITLE
ytitle = ''
else:
xtitle = ''
ytitle = ''
#graph each tsc hp performance
graph( ax, mean_data, scatter( ax, mean_data, std_data, rg_colours, ['']*len(mean_data)),
xtitle=xtitle,
ytitle_pad = ytitle,
title=str(labels[t]),
xlim = [0.0, max(mean_data)*1.05],
ylim= [0.0, max(std_data)*1.05],
grid=True)
#axis colourbar
cax = f.add_axes([0.915, 0.1, 0.05, 0.85])
cmap = mp.cm.brg
cm = LinearSegmentedColormap.from_list('rg', rg_colours, N=rg_colours.shape[0])
norm = mp.colors.Normalize(vmin=0.5, vmax=1.0)
cb = mp.colorbar.ColorbarBase(cax, cmap=cm,
norm=norm,
orientation='vertical')
#color bar axis text
#print([ l._text for l in cb.ax.get_yticklabels()])
#cb_labels = ['']*rg_colours.shape[0]
cb_labels = [ l._text for l in cb.ax.get_yticklabels()]
cb_labels[0] = 'Best'
cb_labels[-1] = 'Worst'
cb.ax.set_yticklabels(cb_labels)
f.suptitle('Hyperparameter Performance')
save_graph(f, save_dir+'tsc_hp.pdf', 600, 14, 24.9)
plt.show()
#now compare all tsc hp sets together in one graph
#prepare data
data_order = sorted(tsc_hp.keys())
#tsc_color = colours[:len(data_order)]
mean_data, std_data, colors, tsc_labels = [], [], [], []
for d in data_order:
n = len(tsc_hp[d][0])
mean_data.extend(tsc_hp[d][0])
std_data.extend(tsc_hp[d][1])
tsc_labels.extend(labels[d])
#colors.extend([c]*len(tsc_hp[d][0]))
colors.extend( [colours[d]]*n )
#graph all hp data all together
f, ax = plt.subplots(1,1)
graph( ax, mean_data, scatter( ax, mean_data, std_data, colors, ['']*len(mean_data)),
xtitle=XTITLE,
ytitle_pad = YTITLE,
title='Traffic Signal Control\nHyperparameter Comparison',
xlim = [0.0, 200.0],
ylim= [0.0, 200.0],
#xlim = [0.0, max(mean_data)*1.05],
#ylim= [0.0, max(std_data)*1.05],
#legend=(0.82, 0.72),
#colours=colours,
grid=True)
#colorbar
#add legend manually because we only
#want one for each tsc
patches = []
for d in data_order:
c = colours[d]
patches.append( mpatches.Patch(color=c, label=labels[d]) )
plt.legend(handles=patches, framealpha=1.0)
save_graph(f, save_dir+'hp.pdf', 600, 14, 24.9)
plt.show()
def stack_folder_files(fp):
data = [ load_data(fp+f) for f in os.listdir(fp)]
return np.stack(data)
def get_hp_results(fp):
travel_times = []
for f in os.listdir(fp):
travel_times.extend(load_data(fp+f))
return travel_times
