def eval_model(args):
"""Evaluate model on training data."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
outdir = 'models/%s/gate_expert' % uid
outname = 'gate_expert_model.pt'
if KLLOSS:
outname = 'gate_expert_kldiv_model.pt'
if args.warm:
outname = outname.replace('.pt', '_warm.pt')
mdl_path = os.path.join(outdir, outname)
gate_expert = GateExpertNet(mdl_path, args.argmax)
eval_fun = gate_expert.get_y
data = npload(cfg['file_path'], uid)
datax = data[cfg['x_name']]
datay = data[cfg['y_name']]
evaly = eval_fun(datax)
print(np.histogram(evaly[:, 48]))
fig, ax = pld.get3dAxis()
ax.scatter(datax[:, 0], datax[:, 1], evaly[:, 48])
loss = l1loss(evaly, datay)
err_norm = np.mean(loss, axis=1)
fig, ax = plt.subplots()
ax.hist(err_norm)
plt.show()
def eval_final_label(args):
"""Evaluation of labels on the training set."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
outdir = 'models/%s/gate_expert' % uid
outname = 'gate_expert_model.pt'
if KLLOSS:
outname = 'gate_expert_kldiv_model.pt'
if args.warm:
outname = outname.replace('.pt', '_warm.pt')
mdl_path = os.path.join(outdir, outname)
gate_expert = GateExpertNet(mdl_path, False)
eval_fun = gate_expert.get_p_y
data = npload(cfg['file_path'], uid)
datax = data[cfg['x_name']]
p, v = eval_fun(datax)
label = np.argmax(p, axis=1)
if args.draw:
fig, ax = plt.subplots()
n_expert = np.amax(label) + 1
for i in range(n_expert):
mask = label == i
ax.scatter(datax[mask, 0], datax[mask, 1])
plt.show()
label_name = 'data/pen/gate_expert_label.npy'
if KLLOSS:
label_name = label_name.replace('_label', '_kldiv_label')
if args.warm:
label_name = label_name.replace('.npy', '_warm.npy')
np.save(label_name, label)
def eval_model_on_valid(args):
"""Evaluate model on the validation set."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
outdir = 'models/%s/gate_expert' % uid
outname = 'gate_expert_model.pt'
if KLLOSS:
outname = 'gate_expert_kldiv_model.pt'
if args.warm:
outname = outname.replace('.pt', '_warm.pt')
mdl_path = os.path.join(outdir, outname)
gate_expert = GateExpertNet(mdl_path, False)
eval_fun = gate_expert.get_y
valid_set = np.load(cfg['valid_path'])
valid_x = valid_set[cfg['x_name']]
valid_y = valid_set[cfg['y_name']]
predy = eval_fun(valid_x)
# dump output into some file
valid_name = 'data/%s/gate_expert_valid_data.npz' % uid
if KLLOSS:
valid_name = valid_name.replace('_valid', '_kldiv_valid')
if args.warm:
valid_name = valid_name.replace('.npz', '_warm.npz')
np.savez(valid_name, x=valid_x, y=predy)
def main():
# pen, car, drone specifies which dataset to use
# the next five clustering approach specifies which data to load
args = getArgs('pen', 'car', 'drone', 'kmean', 'pcakmean', 'speuclid',
'spdydx', 'spvio', 'neighbor10', 'umap', 'pca')
cfg, lbl_name = get_label_cfg_by_args(args)
if args.umap: # I am done using this
draw_umap_label(cfg, lbl_name)
elif args.pca:
draw_pca_label(cfg, lbl_name)
else:
check_cluster_label(cfg, lbl_name, args.neighbor)
def eval_model_on_valid(args):
"""Evaluate model on the validation set."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
outdir='models/%s/mom' % uid
outname='mom_model.pt'
eval_fun = momLoader(os.path.join(outdir, outname), args.prob, False)
valid_set = np.load(cfg['valid_path'])
valid_x = valid_set[cfg['x_name']]
valid_y = valid_set[cfg['y_name']]
predy = eval_fun(valid_x)
# dump output into some file
np.savez('data/%s/mom_valid_data.npz' % uid, x=valid_x, y=predy)
def main():
# num shows dataset size
# drawpca draws pca results for datasets
# umap draws scatter plot of umap projection
# penscatter draws scatter plot on pendulum dataset, we draw scatter plot of x0
args = util.get_args('num', 'drawpca', 'umap', 'penscatter')
cfg, lbl = util.get_label_cfg_by_args(args)
if args.num:
show_dataset_size(cfg)
if args.drawpca:
drawPCA(cfg)
if args.umap:
drawUMap(cfg)
if args.penscatter:
drawPenScatter(cfg)
def eval_final_label(args):
"""Evaluation of labels on the training set."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
outdir = 'models/%s/moe' % uid
outname = 'moe_model.pt'
mdl_path = os.path.join(outdir, outname)
eval_fun = get_moe_loader(mdl_path, True)
data = npload(cfg['file_path'], uid)
datax = data[cfg['x_name']]
p, v = eval_fun(datax)
label = np.argmax(p, axis=1)
np.save('data/pen/moe_label.npy', label)
def train_model_warm(args):
"""Train that model warmly by loading some pre-trained models.
We say warm we mean we pretrain gate or experts. Not sure which I should use instead.
Maybe I really need a function to modify weights of GaoNet to consider xScale and yScale
"""
cfg, lbl_name = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
result = util.get_clus_reg_by_dir('models/%s/%s' % (uid, lbl_name))
cls, regs = result[5] # we know this is bad but could possibly work
cls_data = torch.load(cls)
print(cls_data.keys())
cls_net = cls_data['model']
xmean, xstd = cls_data['xScale']
print('xmean', xmean, 'xstd', xstd)
# cls_net.extendXYScale((xmean, xstd))
expert = Experts([[2, 60, 75]] * 5)
run_the_training(args, cls_net, expert)
def eval_model(args):
"""Evaluate a model."""
cfg, lbl = util.get_label_cfg_by_args(args)
eval_rst = run_the_training(args)
y = eval_rst['y']
if args.prob:
x = eval_rst['x']
p, predy = eval_rst['predy']
print(p.shape, predy.shape)
argmax_idx = np.argmax(p, axis=1)
n_model = np.amax(argmax_idx) + 1
fig, ax = plt.subplots()
for i in range(n_model):
mask = argmax_idx == i
ax.scatter(*x[mask].T)
plt.show()
else:
predy = eval_rst['predy']
n_data = y.shape[0]
# get error
error_y = y - predy
error_norm = np.linalg.norm(error_y, axis=1)
# get histogram
fig, ax = plt.subplots()
ax.hist(error_norm, bins='auto')
ax.set_xlabel('Prediction Error')
ax.set_ylabel('Count')
diff_angle = y[:, 48] - predy[:, 48]
fig, ax = plt.subplots()
ax.hist(diff_angle, bins='auto')
ax.set_xlabel(r'$\theta_f$ Error')
ax.set_ylabel('Count')
# show difference
error_order = np.argsort(error_norm)
for i in range(n_data // 10):
fig, ax = plt.subplots()
for j in range(10):
idx = -1 - (i * 10 + j)
ax.plot(y[idx], color='C%d' % j)
ax.plot(predy[idx], color='C%d' % j, ls='--')
plt.show()
def run_the_training(args):
"""Run the MoE training without using any clustering information but let it find it on its own."""
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
data = npload(cfg['file_path'], uid)
data_feed = {'x': data[cfg['x_name']], 'y': data[cfg['y_name']]}
dimx = data_feed['x'].shape[1]
dimy = data_feed['y'].shape[1]
n_model = args.k
# create the network
net = MoMNet([dimx, 100, n_model], [[dimx, int(np.ceil(300.0 / n_model)), dimy] for _ in range(n_model)])
net.argmax = False
config = genTrainConfig(outdir='models/%s/mom' % uid, outname='mom_model.pt', overwrite=False)
if args.eval:
mdl_path = os.path.join(config['outdir'], config['outname'])
eval_fun = momLoader(mdl_path, withclus=args.prob, argmax=False)
predy = eval_fun(data_feed['x'])
return {'x': data_feed['x'], 'y': data_feed['y'], 'predy': predy}
trainOne(config, data_feed, net=net)
def main():
# pen, car, drone still means which problem we want to look into
# pcakmean specifies the clustering approach we intend to use
# error means we calculate evaluation error on the validation set
# constr means we evaluate constraint violation
# eval just evaluates data on validation set and save into a npz file for rollout validation
# snn means we evaluate SNN network
# roll means look into rollout results and extract useful information. It turns out they all fail, somehow
args = util.get_args('debug', 'error', 'constr', 'eval', 'snn', 'roll')
global DEBUG
if args.debug:
DEBUG = True
cfg, lbl_name = util.get_label_cfg_by_args(args)
if args.error:
eval_valid_error(cfg, lbl_name, args)
if args.constr:
eval_valid_constr_vio(cfg, lbl_name, args)
if args.eval:
eval_on_valid(cfg, lbl_name, args)
if args.roll:
check_rollout_results(cfg, lbl_name, args)
def main():
# pen, car, drone still means which problem we want to look into
# pcakmean specifies the clustering approach we intend to use
# clas means we train classifier
# debug enables debug mode
# miss trains models that are missing, but why?
args = util.get_args('debug', 'clas', 'miss')
global DEBUG
if args.debug:
DEBUG = True
cfg, lbl_name = util.get_label_cfg_by_args(args)
if args.pen:
def archi_fun(ratio):
return gen_archi_fun([2, 300, 75], ratio, 20)
clas = [2, 100, -1]
elif args.car:
def archi_fun(ratio):
return gen_archi_fun([4, 200, 200, 149], ratio, 20)
clas = [4, 500, -1]
elif args.drone:
def archi_fun(ratio):
return gen_archi_fun([7, 1000, 1000, 317], ratio, 20)
clas = [7, 1000, -1]
elif args.dtwo:
def archi_fun(ratio):
return gen_archi_fun([11, 2000, 2000, 317], ratio, 30)
clas = [11, 1000, -1]
elif args.done:
def archi_fun(ratio):
return gen_archi_fun([7, 1000, 1000, 317], ratio, 30)
clas = [7, 1000, -1]
else:
print('You have to choose one dataset')
raise SystemExit
if args.clas:
train_model_classifier(cfg, lbl_name, clas)
elif args.miss:
train_missing_models(cfg, lbl_name, archi_fun)
else:
train_model_by_data(cfg, lbl_name, archi_fun)
def run_the_training(args, clus=None, expert=None):
"""Run the MoE training without using any clustering information but let it find it on its own."""
# load data
cfg, lbl = util.get_label_cfg_by_args(args)
uid = cfg['uniqueid']
print('We are playing with %s' % uid)
data = npload(cfg['file_path'], uid)
data_feed = {'x': data[cfg['x_name']], 'y': data[cfg['y_name']]}
dimx = data_feed['x'].shape[1]
dimy = data_feed['y'].shape[1]
# create gate and expert
if clus is None:
n_model = 5
clus = GaoNet([dimx, 100, n_model])
expert = Experts([[dimx, 60, dimy]] * n_model)
# cuda it
clus.cuda()
expert.cuda()
# set data loader
xname, yname = 'x', 'y'
factory = KeyFactory(data_feed, xname, yname, scalex=True, scaley=True)
factory.shuffle(None)
draw_clus_region(clus, data_feed['x'], factory)
# create two sets
trainsize = 0.8
trainSet = SubFactory(factory, 0.0, trainsize)
testSet = SubFactory(factory, trainsize, 1.0)
batch_size = 32
test_batch_size = -1
trainLder = DataLoader(trainSet, batch_size=batch_size, shuffle=False)
testLder = DataLoader(testSet, batch_size=test_batch_size, shuffle=False)
# set up file output
outname = 'gate_expert_model.pt'
outdir = 'models/pen/gate_expert'
if KLLOSS:
outname = 'gate_expert_kldiv_model.pt'
if args.warm:
outname = outname.replace('.pt', '_warm.pt')
# set optimizer
lr = 1e-3
opt_G = torch.optim.Adam(clus.parameters(), lr=lr)
opt_E = torch.optim.Adam(expert.parameters(), lr=lr)
# set other training stuff
n_epoch = 500
back_check_epoch = 8
best_test_loss = np.inf
best_test_loss_expert = np.inf
best_test_epoch = 0
def get_mean_error(g_y, exp_y, feedy):
"""Calculate two loss"""
error_traj = torch.mean((exp_y - feedy.expand_as(exp_y))**2, dim=2).t()
g = f.softmax(g_y)
log_g = f.log_softmax(g_y)
posterior = g * torch.exp(
-0.5 * error_traj) # b by r probability, not scaled to 1
traj_prob = torch.mean(-torch.log(torch.sum(posterior, dim=1)))
if KLLOSS:
posterior_scale = Variable(
(posterior / torch.sum(posterior, dim=1, keepdim=True)
).data) # do not use gradient of it
div_error = f.kl_div(log_g, posterior_scale)
return traj_prob, div_error
else:
Og = torch.sum(exp_y * g.t().unsqueeze(2), dim=0)
traj_error = f.smooth_l1_loss(Og, feedy)
return traj_prob, traj_error
# start training
for epoch in range(n_epoch):
sum_train_loss = 0
sum_train_loss_prob = 0
for idx, batch_data in enumerate(trainLder):
feedy = Variable(batch_data[yname], requires_grad=False).cuda()
feedx = Variable(batch_data[xname], requires_grad=False).cuda()
# train experts
opt_E.zero_grad()
opt_G.zero_grad()
exp_y = expert(feedx)
g_y = clus(feedx)
g = f.softmax(g_y) # this is prior
log_g = f.log_softmax(g_y)
error_traj = torch.mean((exp_y - feedy.expand_as(exp_y))**2,
dim=2).t()
posterior = g * torch.exp(
-0.5 * error_traj) # b by r probability, not scaled to 1
posterior_scale = Variable(
(posterior / torch.sum(posterior, dim=1, keepdim=True)
).data) # do not use gradient of it
lossi = torch.mean(-torch.log(torch.sum(posterior, dim=1)))
lossi.backward(retain_graph=True)
sum_train_loss_prob += lossi.cpu().data.numpy() * feedx.size()[0]
opt_E.step()
# update h by regression error
all_pred = exp_y
if KLLOSS:
error = f.kl_div(log_g, posterior_scale)
else:
Og_before = all_pred * g.t().unsqueeze(2)
Og = torch.sum(Og_before, dim=0)
error = f.smooth_l1_loss(Og, feedy)
sum_train_loss += error.cpu().data.numpy() * feedx.size()[0]
error.backward()
opt_G.step()
# val = clus.printWeights(3)
mean_train_loss = sum_train_loss / trainLder.getNumData()
mean_train_loss_prob = sum_train_loss_prob / trainLder.getNumData()
# evaluate on test data
sum_test_loss_gate = 0
sum_test_loss_expert = 0
n_test_data = testLder.getNumData()
for idx, batch_data in enumerate(testLder):
feedy = Variable(batch_data[yname], volatile=True).cuda()
feedx = Variable(batch_data[xname], volatile=True).cuda()
exp_y = expert(feedx)
g_y = clus(feedx)
traj_prob, div_error = get_mean_error(g_y, exp_y, feedy)
sum_test_loss_gate += div_error.cpu().data.numpy() * feedx.size(
)[0]
sum_test_loss_expert += traj_prob.cpu().data.numpy() * feedx.size(
)[0]
mean_test_loss_gate = sum_test_loss_gate / n_test_data
mean_test_loss_expert = sum_test_loss_expert / n_test_data
print('epoch %d gate loss %f expert loss %f test gate loss %f expert loss %f' \
% (epoch, mean_train_loss, mean_train_loss_prob, mean_test_loss_gate, mean_test_loss_expert))
if mean_test_loss_gate < best_test_loss:
best_test_loss = mean_test_loss_gate
best_test_epoch = epoch
if mean_test_loss_expert < best_test_loss_expert:
best_test_loss_expert = mean_test_loss_expert
best_test_epoch = epoch
if epoch > best_test_epoch + back_check_epoch:
break
print('Save model now')
# draw region for classifier
draw_clus_region(clus, data_feed['x'], factory)
clus.cpu()
expert.cpu()
model = {
'gate': clus,
'expert': expert,
'xScale': [trainLder.xmean, trainLder.xstd],
'yScale': [trainLder.ymean, trainLder.ystd]
}
if not os.path.exists(outdir):
os.mkdir(outdir)
torch.save(model, os.path.join(outdir, outname))
def main():
args = util.get_args('show')
cfg, lbl_name = util.get_label_cfg_by_args(args)
show_picky_states(cfg, lbl_name, args)
def main():
args = util.get_args()
cfg, lbl = util.get_label_cfg_by_args(args)
show_picky_examples(cfg)
