def _fit(network, name, va):
tr = Frame(np.load(os.path.join("trained_models", name, "tr.npz")))
va = Frame(va)
est = network(name=name)
est.fit(tr, va)
pr = est.predict(va)
rmse = ((va["T"] - pr["Y"])**2).mean()**0.5
print("network", name, "trained with", rmse, "validation rmse")
return rmse
def merge(self, merge_video, frame_nr, start_merge_sec=0, end_merge_sec=0):
file_name = os.path.splitext(os.path.basename(merge_video))[0]
temp_video = os.path.join('tmp/merge_{}.avi'.format(file_name))
probe = ffmpeg.probe(merge_video)
video_info = next(s for s in probe['streams'] if s['codec_type'] == 'video')
width = int(video_info['width'])
height = int(video_info['height'])
fps = eval(video_info['avg_frame_rate'])
x = time.strptime(video_info['tags']['DURATION'].split('.')[0], '%H:%M:%S')
duration = int(datetime.timedelta(minutes=x.tm_min, seconds=x.tm_sec).total_seconds() + 1)
num_frames = int(duration * fps)
end_merge_sec = end_merge_sec if end_merge_sec != 0 else duration
process1 = (
ffmpeg
.input(merge_video)
.output(temp_video, r=fps, b='500000k', g=num_frames, pix_fmt='yuv420p',
keyint_min=999999,
ss=start_merge_sec, to=end_merge_sec)
.run(overwrite_output=True)
)
print('Done making p-frame video')
merge_input = open(temp_video, 'rb')
self.get_frames()
print('Got frames of first vid: {}'.format(len(self.frames)))
for i in range(0, frame_nr):
self.write_frame(self.frames[i])
in_file_bytes = merge_input.read()
# 0x30306463 which is ASCII 00dc signals the end of a frame. '0x' is a common way to say that a number is in hexidecimal format.
frames = [Frame(d) for d in in_file_bytes.split(bytes.fromhex('30306463'))]
for frame in frames:
if frame.is_delta_frame():
self.write_frame(frame)
def predict_raw(self, D):
""" returns dictionary of raw results """
network = self.network_shim()
prY = np.squeeze(network.predict(D)['Y'])
pr = Frame(D.items(), Y=prY)
result = defaultdict(list)
splits = int(np.ceil(pr.N / 5000))
for i, idx in enumerate(np.array_split(np.arange(pr.N), splits)):
print("\rworking on split {i} of {n}...".format(i=i + 1, n=splits),
end="")
memberships = self.cluster_est.predict(
self._pipe_transform(pr[idx]))
result["memberships"] += [memberships]
print()
pr.update([(k, np.concatenate(v)) for k, v in result.items()])
return pr
def predict(self, D, coverage=0.95): prY = self.network_shim().predict(D)['Y'] pr = Frame(D.items(), Y=prY) alpha = 1.0 - coverage # opt: SQP or Cobyla quantiles = self.est.compute_quantile(self._pipe_transform(pr), np.array([alpha/2, 1.0-alpha/2]), do_optim=True, opt_method='Cobyla') # print(quantiles.shape) # (N, 2) return prY, quantiles[:,0], quantiles[:,1]
def fit(self, tr, va): network = self.network_shim().fit(tr, va) self.ref = Frame(va.items(), Y=network.predict(va)['Y']) # maybe include the tr set, if prY isn't part of the input # self.ref = Frame([(k, np.concatenate([ref[k],tr[k]])) for k in ref.keys()]) self._pipe_fit(self.ref, self.keys) self.est.fit(self._pipe_transform(self.ref), self.ref['T']) return self
def fit(self, trD, vaD):
network = self.network_shim().fit(trD, vaD)
data = []
if "tr" in self.sets:
data += [(trD, np.squeeze(network.predict(trD)['Y']))]
if "va" in self.sets:
data += [(vaD, np.squeeze(network.predict(vaD)['Y']))]
self.ref = Frame([(k, np.concatenate([D[k] for D, prY in data]))
for k in trD.keys()])
self.ref['Y'] = np.squeeze(np.concatenate([prY for D, prY in data]))
self.ref_results = Frame(R=self.ref['T'] - self.ref['Y']) # residuals
self._pipe_fit(self.ref, self.keys)
self.knn = NearestNeighbors(n_neighbors=self.neighbors,
n_jobs=self.n_jobs)
self.knn.fit(self._pipe_transform(self.ref))
return self
def _predict(self, D, ops_dict, is_validation_run=False):
if self._initialize:
raise RuntimeError('Session is uninitialized')
if self._session._closed:
raise RuntimeError('Session has already been closed.')
# need the original set for D, not prefabricated batches
if isinstance(D, list):
keys = D[0].keys()
data = Frame((k, np.concatenate([d[k] for d in D])) for k in keys)
elif not isinstance(D, Frame) and isinstance(D, dict):
data = Frame(**D)
else:
data = D
prE = ops_dict["E"]
Es = []
I = np.arange(len(data))
# predict once for the function value using global normalization
# chunks = self._split(D, 1000)
# E = np.concatenate([self._session.run([prE], feed_dict=self._fill_feed(batch, is_training=False))[0] for batch in chunks])
# Es += [E]
# predict many times for the uncertainty using additional batch normalization
if not is_validation_run:
for i in range(self._n_predictions):
I = np.random.permutation(I)
chunks = self._split(D[I], self._prediction_batch_size)
E = np.concatenate([
self._session.run([prE],
feed_dict=self._fill_feed(
batch, is_training=True))[0]
for batch in chunks
])
Es += [E[np.argsort(I)]]
return dict(
E=np.mean(Es, axis=0),
U=np.var(Es, axis=0),
)
def load_simple(filename="data/benzene.npz", skip=10e3, N=None, sizes=[10e3, 2e3, 8e3], verbose=True): N = N or sum(sizes) skip, N = int(np.round(skip)), int(np.round(N)) data_raw = Frame(np.load(filename)) data_raw['index'] = np.arange(len(data_raw)) data_raw = data_raw[skip:] # leak information about your testset, decrease your rmse! data_raw = data_raw[np.random.choice(np.arange(len(data_raw)), size=N, replace=False)] if verbose: print(data_raw.keys(), len(data_raw["E"])) preproc = MolecularDataPreprocessor(expand_method=None, target_key="E", dont_norm='R Z E index'.split()) sets = split_sets(data_raw['D E R Z index'.split()].copy(), sizes, start=0) preproc = preproc.fit(sets[0]) sets = map(preproc.transform, sets) sets = [Frame(s.items()) for s in sets] if verbose: print([len(s) for s in sets], 'sets') for k,v in sets[0].items(): for s in sets: print(k, s[k].shape, s[k].mean(), s[k].std()) return preproc, sets
def fit(self, tr, va):
network = self.network_shim().fit(tr, va)
self.ref = Frame(va.items(), Y=network.predict(va)['Y'])
self._pipe_fit(self.ref, self.keys)
print("fitting normal", end="\r")
self.est.fit(self._pipe_transform(self.ref), self.ref['T'])
print("fitting lower ", end="\r")
self.est_min.fit(self._pipe_transform(self.ref), self.ref['T'])
print("fitting upper", end="\r")
self.est_max.fit(self._pipe_transform(self.ref), self.ref['T'])
print("fitting done")
return self
def predict(self, D, coverage=None):
if coverage is not None:
print(
"WARNING: GradientBoostingUncertainty requires the desired coverage as an init parameter, otherwise it will be ignored."
)
prY = self.network_shim().predict(D)['Y']
pr = Frame(D.items(), Y=prY)
prY = self.est.predict(self._pipe_transform(pr))
minY = self.est_min.predict(self._pipe_transform(pr))
maxY = self.est_max.predict(self._pipe_transform(pr))
return prY, minY, maxY
r=fps,
b='500000k',
g=fps * (end_sec - start_sec),
pix_fmt='yuv420p',
keyint_min=999999,
ss=start_sec,
to=end_sec).run(overwrite_output=True))
in_file = open(temp_video, 'rb')
out_file = open(output_mosh, 'wb')
# because we used 'rb' above when the file is read the output is in byte format instead of Unicode strings
in_file_bytes = in_file.read()
# 0x30306463 which is ASCII 00dc signals the end of a frame. '0x' is a common way to say that a number is in hexidecimal format.
frames = [Frame(d) for d in in_file_bytes.split(bytes.fromhex('30306463'))]
def write_frame(frame):
out_file.write(frame + bytes.fromhex('30306463'))
pic1, err1 = (ffmpeg.input(input_png, s='{}x{}'.format(width, height)).trim(
start_frame=0,
end_frame=1).output('pipe:',
format='avi',
pix_fmt='yuv420p',
s='{}x{}'.format(width,
height)).run(capture_stdout=True))
png_frame = pic1.split(bytes.fromhex('30306463'))[1]
def predict(self, data):
cache = Frame(**np.load(self.cachefile))
cacheindex = cache['index'] # should be sorted
I = np.argsort(cacheindex)
return cache[I[np.searchsorted(cacheindex[I], data['index'])]]
def fit(self, trD, vaD):
network = self.network_shim().fit(trD, vaD)
data = []
if "tr" in self.sets:
data += [(trD, np.squeeze(network.predict(trD)['Y']))]
if "va" in self.sets:
data += [(vaD, np.squeeze(network.predict(vaD)['Y']))]
self.ref = Frame([(k, np.concatenate([D[k] for D, prY in data]))
for k in trD.keys()])
self.ref['Y'] = np.squeeze(np.concatenate([prY for D, prY in data]))
self.ref_results = Frame(R=self.ref['T'] - self.ref['Y']) # residuals
self._pipe_fit(self.ref, self.keys)
n_clusters = int((len(self.ref) / 2)**0.5) # heuristic from pevec2013
a = 1.0 - self.coverage
R = self.ref_results["R"]
if self.method == 'kmeans' or self.method == 'kmeans-mahalanobis':
self.cluster_est = KMeans(n_clusters=n_clusters,
n_jobs=self.n_jobs)
memberships = self.cluster_est.fit_predict(
self._pipe_transform(self.ref))
self.ref_results["memberships"] = memberships
self.clusterbounds = np.empty((n_clusters, 2))
for c in range(n_clusters):
cdf, ppf = ecdf(R[memberships == c])
self.clusterbounds[c, :] = ppf(a / 2), ppf(1 - a / 2)
elif self.method == "cmeans":
self.cluster_est = CMeans(n_clusters=n_clusters,
error=1e-5,
max_iter=10 * self.ref.N)
memberships = self.cluster_est.fit_predict(
self._pipe_transform(self.ref)) # (N, clusters)
asc_residuals = np.argsort(R)
# self.clusterbounds = np.empty((n_clusters, 2))
# for c in range(n_clusters):
# cumsum = np.cumsum(memberships.T[c][asc_residuals])
# lb = np.argwhere(cumsum < (a/2)*cumsum[-1])[-1,0]
# ub = np.argwhere(cumsum > (1-a/2)*cumsum[-1])[0,0]
# self.clusterbounds[c,:] = R[asc_residuals][lb], R[asc_residuals][ub]
cumsum = np.cumsum(memberships[asc_residuals],
axis=0).T # (clusters, N)
self.clusterbounds = np.vstack([
R[asc_residuals][np.argmax(cumsum >= (a / 2) * cumsum[:, -1:],
axis=1)], R[asc_residuals]
[np.argmin(cumsum <= (1 - a / 2) * cumsum[:, -1:], axis=1) - 1]
]).T
elif self.method == 'density':
self.cluster_est = GaussianDensity(n_jobs=self.n_jobs)
self.cluster_est.fit(self._pipe_transform(self.ref))
density = self.cluster_est.predict(self._pipe_transform(self.ref))
self.max_density = density.max()
else:
raise ValueError("Method {} is not a valid option.".format(
self.method))
return self
def predict(est,
data_dir=None,
calibrate=False,
coverage=0.95,
pretrained=False,
savename=None):
''' predicts '''
print(est, "(loaded estimator)")
print("loading datasets")
if not isinstance(data_dir,
str): # directory containing tr, va, te{i} files
raise ValueError(
"data must a path to directory containing tr.npz va.npz te.npz files"
)
# load data from file
sets = dict()
for name in "tr va te de special".split():
filename = os.path.join(data_dir, "{}.npz".format(name))
if os.path.exists(filename):
print("loading file", filename)
sets[name] = Frame(**np.load(filename))
# fit
if pretrained:
print("pretrained model, training skipped")
else:
print("fitting...")
tstart = time.time()
est.fit(sets['tr'], sets['va'])
mins = (time.time() - tstart) / 60
print("fit in {}h{}m".format(int(mins / 60), int(mins % 60)))
if calibrate:
prY, lo, hi = est.predict(sets['va'], coverage=coverage)
factor = est.calibrate(
sets['va']['T'], prY, (hi - lo) / 2
) # factor is a global scaling factor, but pointwise scaling will be used instead
# predict
print("predicting and saving...")
has_index = 'index' in sets['tr'] and 'index' in sets['va']
# predict and save reference sets
prs = dict()
for dname, d in sets.items():
print("predicting", dname.upper(), "for coverage", coverage)
prY, minY, maxY = est.predict(d, coverage=coverage)
pr = dict(T=d["T"], Y=prY, minY=minY, maxY=maxY)
print(((pr['T'] - pr['Y'])**2).mean()**0.5, 'rmse')
prs[dname] = Frame(pr)
if has_index: pr["index"] = d["index"].copy()
path = os.path.join(data_dir, savename)
os.makedirs(path, exist_ok=True)
np.savez(os.path.join(path, "pr_{}.npz".format(dname)), **pr)
if calibrate: # save calibrated predictions, if desired
prY, minY, maxY = est.scale(pr['T'],
pr['Y'],
pr['minY'],
pr['maxY'],
coverage=coverage)
pr = dict(pr.items(), Y=prY, minY=minY, maxY=maxY)
dname_cp = "{}_{:02d}cp".format(dname, int(100 * coverage))
prs[dname_cp] = Frame(pr)
np.savez(os.path.join(path, "pr_{}.npz".format(dname_cp)), **pr)
print("predict done")
return prs
def predict_raw(self, D, skip_closest=0):
""" returns dictionary of raw results """
network = self.network_shim()
prY = np.squeeze(network.predict(D)['Y'])
pr = Frame(D.items(), Y=prY)
if not hasattr(self, "_pipe"):
raise RuntimeError(
"Estimator is not fitted yet. The neural network shim should decide if training is necessary or not."
)
X = self._pipe_transform(pr)
refX = self._pipe_transform(self.ref)
if self.method == 'mahalanobis': # takes a long time
if not self.silent: print("caching inverse covariance matrix ...")
self.mahalanobis_uncertainty(X[:2],
refX[:2],
refX,
n_jobs=self.n_jobs)
k = self.neighbors
knn = self.knn
knn.set_params(n_neighbors=k)
result = defaultdict(list)
splits = int(np.ceil(len(pr) / 5000))
for i, idx in enumerate(np.array_split(np.arange(len(pr)), splits)):
if not self.silent:
print("working on split {i} of {n}...".format(i=i + 1,
n=splits),
end="\r")
distances, neighbors = (A[:, skip_closest:]
for A in knn.kneighbors(X[idx]))
knnR = self.ref_results['R'][
neighbors] # signed residuals in neighborhood
knnY = self.ref['Y'][neighbors] # predictions in neighborhood
knnT = self.ref['T'][neighbors] # labels (targets) in neighborhood
# print(knnY.shape, "knnY shape") # (n_points_in_split, n_neighbors)
result['avgDist'] += [distances.mean(axis=1)]
result['meanR'] += [knnR.mean(axis=1)] # for stdR
result['meanY'] += [knnY.mean(axis=1)]
result['meanT'] += [knnT.mean(axis=1)]
result['stdR'] += [
(((knnR - result['meanR'][-1][:, None])**2).sum(axis=1) /
(k - 1))**0.5
]
result['absR'] += [np.abs(knnR).mean(axis=1)]
result['varT'] += [np.var(knnT, axis=1, ddof=1)] # cs_knnV
result['stdT'] += [np.var(knnT, axis=1,
ddof=1)**0.5] # cs_knnV**0.5
result['sqR'] += [(knnR**2).mean(axis=1)] # cs_knnE
if self.method == 'mahalanobis': # takes a long time, exceeds memory limits if not split like this
VI = self._mahalanobis_params['VI'] # precomputed above
dists = Parallel(self.n_jobs, 'threading', verbose=0)(
delayed(run)(self._mahalanobis_uncertainty_job,
X[idx[i]][None, :],
refX[neighbors[i]],
VI=VI) for i in range(len(idx)))
result['mn-dist'] += [np.concatenate(dists)]
if not self.silent: print()
pr.update([(k, np.concatenate(v)) for k, v in result.items()])
return pr
def _predict(network, name, data):
data = Frame(data)
est = network(name=name)
pr = est.predict(data)
return pr["Y"]
def create_instances(data_dir="trained_models/benzene/",
folder_prefix="benzene/",
n_jobs=1,
only_est=None,
gpus=(1, ),
mu_max=10):
'''
returns a list of (name,constructor) tuples of uncertainty estimators to train
and predict with in the evaluation loop.
'''
fp = folder_prefix
if fp[:len("trained_models")] == "trained_models":
# all models are implied to be saved under trained_models/, no need to
# explicitely prefix it
fp = fp[len("trained_models/"):]
instances = []
dtnn = lambda: DTNNShim(name=os.path.join(fp, "dtnn"))
normal_ppf = scipy.stats.norm.ppf
t_ppf = lambda df: lambda x: scipy.stats.t.ppf(x, df=df)
t50_ppf = t_ppf(50)
ensemble_members = 24
days = ceil(ensemble_members / sum(gpus)) * 72 / 24
print("ensemble training will take approx.", days, "days")
tr = Frame(**load(os.path.join(data_dir, "tr.npz")))
params["tr_mean_std"] = Frame(
**load(os.path.join(data_dir, "tr_mean_std.npz")))
params["point"] = tr[0]
params["n_tr_points"] = len(tr)
params["mu_max"] = mu_max
# basic baseline, nothing should perform worse than these
instances += [("dummy-const",
lambda: DummyUncertainty(dtnn, method="const"))]
# these rely on the points being sorted (from one trajectory)
instances += [("dummy-f'", lambda: DummyUncertainty(dtnn, method="f'"))]
instances += [("dummy-f''", lambda: DummyUncertainty(dtnn, method="f''"))]
instances += [("dummy-1/f'", lambda: DummyUncertainty(dtnn, method="1/f'"))
]
instances += [("dummy-1/f''",
lambda: DummyUncertainty(dtnn, method="1/f''"))]
# dropout variance ~ gal2015
instances += [('doho60', lambda: UncertaintyWrapper(
HoShim60, ppf=t50_ppf, name=os.path.join(fp, 'doho60')))]
instances += [('doho70', lambda: UncertaintyWrapper(
HoShim70, ppf=t50_ppf, name=os.path.join(fp, 'doho70')))]
instances += [('doho80', lambda: UncertaintyWrapper(
HoShim80, ppf=t50_ppf, name=os.path.join(fp, 'doho80')))]
instances += [('doho90', lambda: UncertaintyWrapper(
HoShim90, ppf=t50_ppf, name=os.path.join(fp, 'doho90')))]
instances += [('dohe60', lambda: UncertaintyWrapper(
HeShim60, ppf=t50_ppf, name=os.path.join(fp, 'dohe60')))]
instances += [('dohe70', lambda: UncertaintyWrapper(
HeShim70, ppf=t50_ppf, name=os.path.join(fp, 'dohe70')))]
instances += [('dohe80', lambda: UncertaintyWrapper(
HeShim80, ppf=t50_ppf, name=os.path.join(fp, 'dohe80')))]
instances += [('dohe90', lambda: UncertaintyWrapper(
HeShim90, ppf=t50_ppf, name=os.path.join(fp, 'dohe90')))]
# conditional density estimation [bishop1995]
instances += [("density", lambda: UncertaintyWrapper(
DensityShim, ppf=normal_ppf, name=os.path.join(fp, "density")))]
instances += [("density_full", lambda: UncertaintyWrapper(
FullDensityShim, ppf=normal_ppf, name=os.path.join(fp, "density_full"))
)]
instances += [
("density_stepped",
lambda: UncertaintyWrapper(SteppedDensityShim,
ppf=normal_ppf,
name=os.path.join(fp, "density_stepped")))
]
instances += [("density_half", lambda: UncertaintyWrapper(
HalfDensityShim, ppf=normal_ppf, name=os.path.join(fp, "density_half"))
)]
instances += [("mixture", lambda: UncertaintyWrapper(
MixtureShim, ppf=normal_ppf, name=os.path.join(fp, "mixture")))]
instances += [("mixture_half", lambda: UncertaintyWrapper(
HalfMixtureShim, ppf=normal_ppf, name=os.path.join(fp, "mixture_half"))
)]
instances += [("mixture_full", lambda: UncertaintyWrapper(
FullMixtureShim, ppf=normal_ppf, name=os.path.join(fp, "mixture_full"))
)]
instances += [
("mixture_stepped",
lambda: UncertaintyWrapper(SteppedMixtureShim,
ppf=normal_ppf,
name=os.path.join(fp, "mixture_stepped")))
]
instances += [("mixture_re", lambda: UncertaintyWrapper(
ReMixtureShim, ppf=normal_ppf, name=os.path.join(fp, "mixture_re")))]
# print("skipping quantile regression")
# if False:
# instances += [("quantile-regression", lambda: GradientBoostingUncertainty(DTNNShim, keys="DY", n_estimators=100, max_depth=300, learning_rate=.1, min_samples_leaf=9, min_samples_split=9))]
# est = lambda: QuantileForestUncertainty(DTNNShim, keys="DY", n_estimators=500, max_depth=30, n_jobs=n_jobs)
# # let's try it how the org. authors suggested as well (e.i. not include prY)
# instances += [("quantile-regression-no-y", lambda: GradientBoostingUncertainty(DTNNShim, keys="D", n_estimators=100, max_depth=300, learning_rate=.1, min_samples_leaf=9, min_samples_split=9))]
# instances += [("quantile-forest-no-y", lambda: QuantileForestUncertainty(DTNNShim, keys="D", n_estimators=500, max_depth=30, n_jobs=n_jobs))]
# fuzzy c-means clustering [pevec2013 shrestha2006]
# LOCAL NEIGHBORHOOD
# instances += [("fuzzy-cmeans", lambda: ClusterUncertainty(dtnn, method="cmeans", coverage=cp, n_jobs=n_jobs))]
# instances += [("kmeans", lambda: ClusterUncertainty(dtnn, method="kmeans", coverage=cp, n_jobs=n_jobs))]
# mahalanobis distance to data set center [toplak2014]
# instances += [("kmeans-mahalanobis", lambda: ClusterUncertainty(dtnn, method="kmeans-mahalanobis", coverage=cp, n_jobs=n_jobs))]
# density based estimate [bosnic2008]
# instances += [("gaussian-clusters", lambda: ClusterUncertainty(dtnn, method="density", coverage=cp, n_jobs=n_jobs))]
# CONFIVE (Variance of Error) [briesemeister2012]
# (self, network_shim, method="label_var", sets="va", keys="", neighbors=20, n_jobs=1, silent=False):
# quantile will have to be provided at evaluation time.
# it makes sense to generate quantiles when the desired coverage is known.
instances += [("knn-mse", lambda: KNNUncertainty(
dtnn, method="mse", sets='va', keys='Y', neighbors=8, n_jobs=n_jobs))]
instances += [("knn-mseq", lambda: KNNUncertainty(
dtnn, method="mse", sets='tr', keys='', neighbors=10, n_jobs=n_jobs))]
instances += [("knn-dev", lambda: KNNUncertainty(
dtnn, method="dev", sets='va', keys='Y', neighbors=716, n_jobs=n_jobs))
]
instances += [("knn-avgdist", lambda: KNNUncertainty(dtnn,
method="avgdist",
sets='tr',
keys='Y',
neighbors=34,
n_jobs=n_jobs))]
instances += [("knn-vary", lambda: KNNUncertainty(
dtnn, method="label_var", sets='va', neighbors=2653, n_jobs=n_jobs))]
# instances += [("knn-mahalanobis", lambda: KNNUncertainty(dtnn, method="mahalanobis", n_jobs=n_jobs))]
# instances += [("knn-absolute-deviation-from-avg-label", lambda: KNNUncertainty(dtnn, method="absDev", n_jobs=n_jobs))]
# instances += [("knn-mae", lambda: KNNUncertainty(dtnn, method="mae", n_jobs=n_jobs))]
# instances += [("knn-rmse", lambda: KNNUncertainty(dtnn, method="rmse", n_jobs=n_jobs))]
# avgDist [briesemeister2012]
# instances += [("knn-avg-dist", lambda: KNNUncertainty(dtnn, method="avgDist", n_jobs=n_jobs))]
# ENSEMBLES
# bagging by example pairs and residuals [tibshirani1995]
instances += [(
"ensemble-pairs",
lambda: EnsembleUncertainty(DTNNShim,
m=ensemble_members,
method="pairs",
balancing=False,
folder=os.path.join(fp, "ensemble-pairs/"),
gpus=gpus))]
instances += [("ensemble-pairs-12", lambda: EnsembleUncertainty(
DTNNShim,
m=12,
method="pairs",
balancing=False,
folder=os.path.join(fp, "ensemble-pairs-12/"),
gpus=gpus))]
instances += [("ensemble-pairs-6", lambda: EnsembleUncertainty(
DTNNShim,
m=6,
method="pairs",
balancing=False,
folder=os.path.join(fp, "ensemble-pairs-6/"),
gpus=gpus))]
instances += [("ensemble-pairs-3", lambda: EnsembleUncertainty(
DTNNShim,
m=3,
method="pairs",
balancing=False,
folder=os.path.join(fp, "ensemble-pairs-3/"),
gpus=gpus))]
instances += [("ensemble-residuals", lambda: EnsembleUncertainty(
DTNNShim,
m=ensemble_members,
method="residuals",
balancing=False,
folder=os.path.join(fp, "ensemble-residuals/"),
gpus=gpus))]
instances += [("ensemble-residuals-12", lambda: EnsembleUncertainty(
DTNNShim,
m=12,
method="residuals",
balancing=False,
folder=os.path.join(fp, "ensemble-residuals-12/"),
gpus=gpus))]
instances += [("ensemble-residuals-6", lambda: EnsembleUncertainty(
DTNNShim,
m=6,
method="residuals",
balancing=False,
folder=os.path.join(fp, "ensemble-residuals-6/"),
gpus=gpus))]
instances += [("ensemble-residuals-3", lambda: EnsembleUncertainty(
DTNNShim,
m=3,
method="residuals",
balancing=False,
folder=os.path.join(fp, "ensemble-residuals-3/"),
gpus=gpus))]
# balancing doesn't yield better estimates, unfortunately
instances += [("ensemble-pairs-balanced", lambda: EnsembleUncertainty(
DTNNShim,
m=ensemble_members,
k=ensemble_members // 6,
method="pairs",
balancing=True,
folder=os.path.join(fp, "ensemble-pairs-balanced/"),
gpus=gpus))]
instances += [("ensemble-residuals-balanced", lambda: EnsembleUncertainty(
DTNNShim,
m=ensemble_members,
k=ensemble_members // 6,
method="residuals",
balancing=True,
folder=os.path.join(fp, "ensemble-residuals-balanced/"),
gpus=gpus))]
if only_est is not None:
estimators = dict(instances)
if only_est not in estimators.keys():
raise ValueError(
"No such estimator instance: '{}'\nChoose from \n {}".format(
only_est, "\n ".join(sorted(estimators.keys()))))
else:
return [(only_est, estimators[only_est])]
return instances