def test_3():
# test using a callable metric. should get same results
model1 = LandmarkAgglomerative(n_clusters=10, n_landmarks=20, metric='euclidean')
model2 = LandmarkAgglomerative(n_clusters=10, n_landmarks=20, metric=lambda target, ref, i: np.sqrt(np.sum((target-ref[i])**2, axis=1)))
data = np.random.RandomState(0).randn(100, 2)
eq(model1.fit_predict([data])[0], model2.fit_predict([data])[0])
def test_agglom_with_metric_msm():
my_list = [_get_random_prob_dist(4) for i in range(100)]
my_flat = np.array([x.flatten() for x in my_list])
model = LandmarkAgglomerative(n_clusters=2,
metric=sym_kl_divergence_msm,
linkage='complete')
assert model.fit_predict([my_flat])[0].shape == (100, )
def test_cluster_centers():
x = [random.randn(20, 2) + 10, random.randn(20, 2)]
n_clusters = np.random.randint(2, 7)
model = LandmarkAgglomerative(n_clusters=n_clusters,
linkage='ward')
labels = model.fit_predict(x)
print(model.cluster_centers_)
assert model.cluster_centers_.shape == (n_clusters, 2)
def test_alanine_dipeptide():
# test for rmsd metric compatibility with ward clustering
# keep n_landmarks small or this will get really slow
trajectories = AlanineDipeptide().get_cached().trajectories
n_clusters = 4
model = LandmarkAgglomerative(n_clusters=n_clusters, n_landmarks=20,
linkage='ward', metric='rmsd')
labels = model.fit_predict(trajectories[0][0:100])
assert len(np.unique(np.concatenate(labels))) <= n_clusters
def test_2():
# this should be a really easy clustering problem
x = [random.randn(20,2)+10, random.randn(20,2)]
n_clusters = 2
model1 = LandmarkAgglomerative(n_clusters=n_clusters)
model2 = LandmarkAgglomerative(n_clusters=n_clusters,
landmark_strategy='random', random_state=random, n_landmarks=20)
labels1 = model1.fit_predict(x)
labels2 = model2.fit_predict(x)
assert adjusted_rand_score(np.concatenate(labels1), np.concatenate(labels2)) == 1.0
def _do_lumping(self):
"""Do the MVCA lumping.
"""
model = LandmarkAgglomerative(linkage='ward',
n_clusters=self.n_macrostates,
metric=self.metric,
n_landmarks=self.n_landmarks,
landmark_strategy=self.landmark_strategy,
random_state=self.random_state)
microstate_mapping_ = model.fit_transform([self.transmat_])[0]
self.microstate_mapping_ = microstate_mapping_
def _do_lumping(self):
"""Do the MVCA lumping.
"""
model = LandmarkAgglomerative(linkage='ward',
n_clusters=self.n_macrostates,
metric=self.metric,
n_landmarks=self.n_landmarks,
landmark_strategy=self.landmark_strategy,
random_state=self.random_state)
model.fit([self.transmat_])
if self.fit_only:
microstate_mapping_ = model.landmark_labels_
else:
microstate_mapping_ = model.transform([self.transmat_])[0]
self.microstate_mapping_ = microstate_mapping_
def test_1():
x = [random.randn(10,2), random.randn(10,2)]
n_clusters = 2
model1 = LandmarkAgglomerative(n_clusters=n_clusters)
model2 = LandmarkAgglomerative(n_clusters=n_clusters,
n_landmarks=sum(len(s) for s in x))
labels0 = clone(model1).fit(x).predict(x)
labels1 = model1.fit_predict(x)
labels2 = model2.fit_predict(x)
assert len(labels0) == 2
assert len(labels1) == 2
assert len(labels2) == 2
eq(labels0[0], labels1[0])
eq(labels0[1], labels1[1])
eq(labels0[0], labels2[0])
eq(labels0[1], labels2[1])
assert len(np.unique(np.concatenate(labels0))) == n_clusters
def test_callable_metric():
def my_euc(target, ref, i):
return np.sqrt(np.sum((target - ref[i]) ** 2, axis=1))
model1 = LandmarkAgglomerative(n_clusters=10, n_landmarks=20,
metric='euclidean')
model2 = LandmarkAgglomerative(n_clusters=10, n_landmarks=20, metric=my_euc)
data = np.random.RandomState(0).randn(100, 2)
eq(model1.fit_predict([data])[0], model2.fit_predict([data])[0])
def test_3():
# test using a callable metric. should get same results
model1 = LandmarkAgglomerative(n_clusters=10,
n_landmarks=20,
metric='euclidean')
model2 = LandmarkAgglomerative(n_clusters=10,
n_landmarks=20,
metric=lambda target, ref, i: np.sqrt(
np.sum((target - ref[i])**2, axis=1)))
data = np.random.RandomState(0).randn(100, 2)
eq(model1.fit_predict([data])[0], model2.fit_predict([data])[0])
def test_2():
# this should be a really easy clustering problem
x = [random.randn(20, 2) + 10, random.randn(20, 2)]
n_clusters = 2
model1 = LandmarkAgglomerative(n_clusters=n_clusters)
model2 = LandmarkAgglomerative(n_clusters=n_clusters,
landmark_strategy='random',
random_state=random, n_landmarks=20)
labels1 = model1.fit_predict(x)
labels2 = model2.fit_predict(x)
assert adjusted_rand_score(np.concatenate(labels1),
np.concatenate(labels2)) == 1.0
def test_1():
x = [random.randn(10, 2), random.randn(10, 2)]
n_clusters = 2
model1 = LandmarkAgglomerative(n_clusters=n_clusters)
model2 = LandmarkAgglomerative(n_clusters=n_clusters,
n_landmarks=sum(len(s) for s in x))
labels0 = clone(model1).fit(x).predict(x)
labels1 = model1.fit_predict(x)
labels2 = model2.fit_predict(x)
assert len(labels0) == 2
assert len(labels1) == 2
assert len(labels2) == 2
eq(labels0[0], labels1[0])
eq(labels0[1], labels1[1])
eq(labels0[0], labels2[0])
eq(labels0[1], labels2[1])
assert len(np.unique(np.concatenate(labels0))) == n_clusters
tops = preload_tops(meta)
totframes = meta['nframes'].sum()
def traj_load(irow):
i, row = irow
traj = md.load(row['traj_fn'], top=tops[row['top_fn']])
return i, traj
traj_dict = dict(map(traj_load, meta.iterrows()))
trajs = [traj for traj in traj_dict.values() if traj.n_frames > 1000]
print(len(trajs))
num_clust = 20
cluster = LandmarkAgglomerative(n_clusters=num_clust,
n_landmarks=int(totframes / 100),
linkage='ward',
metric='rmsd')
ctrajs = cluster.fit_transform(trajs)
# print('Fitting cluster labels for MSM')
# ctraj = {}
# count = 0
# for k, v in traj_dict.items():
# print(k, count)
# count +=1
# ctraj[k] = cluster.partial_predict(v)
#
# ctrajs = [traj for traj in ctraj.values() if traj.shape[0] > 1000]
print('Fitting MSM')
lag = 4000
def traj_load(irow):
i, row = irow
traj = md.load(row['traj_fn'], top=tops[row['top_fn']])
return i, traj
traj_dict = dict(map(traj_load, meta.iterrows()))
trajs = [traj for traj in traj_dict.values()]
# Make Pipeline
cv_iter = ShuffleSplit(n_splits=5, test_size=0.5)
estimators = [('cluster',
LandmarkAgglomerative(n_clusters=2,
n_landmarks=int(totframes / 200),
linkage='ward',
metric='rmsd')),
('msm', MarkovStateModel())]
params = {'cluster__n_clusters': [200]}
pipe = Pipeline(estimators)
pipe.set_params(msm__lag_time=999)
pipe.set_params(msm__n_timescales=20)
if __name__ == "__main__":
cvSearch = GridSearchCV(pipe, params, n_jobs=1, verbose=1, cv=cv_iter)
print("Performing grid search...")
print("pipeline:", [name for name, _ in pipe.steps])
matplotlib.use('Agg')
from matplotlib.pylab import plt
import numpy as np
import seaborn as sns
from utilities import to_dataframe
# load trajectories
feature = 'dihedrals'
meta, traj_dict= load_trajs('pruned_trajectories/{}-ftraj'.format(feature))
trajs = [traj for traj in traj_dict.values()]
nframes = int(np.max(meta['nframes'].unique()[0]))
# cluster
num_clusters=10
cluster = LandmarkAgglomerative(n_clusters=num_clusters, n_landmarks=200, linkage='ward', metric='euclidean')
cluster.fit(trajs)
ctraj = {}
for k, v in traj_dict.items():
v = v.copy(order='C')
v = cluster.partial_predict(v)
diff = nframes-v.shape[0]
v = np.append(v, np.zeros(diff)-1)
ctraj[k] = v
# Convert to DF for plotting and sampling.
df = to_dataframe(ctraj, nframes, dt=1)
# Plot trajectories
sample = df.sample(frac=0.1, axis=0)
if __name__ == '__main__':
with Pool() as pool:
trajs_dct = dict(pool.imap_unordered(traj_load, meta.iterrows()))
trajs = [traj for traj in trajs_dct.values()]
to_ns, t_max, frames_tot = get_timings(meta)
n_clusters = int(np.sqrt(frames_tot))
print(n_clusters)
# n_clusters = int(frames_tot/1000)
clusterer = LandmarkAgglomerative(n_clusters=n_clusters,
n_landmarks=n_clusters // 10,
linkage='ward',
metric='rmsd',
landmark_strategy='stride',
random_state=None,
max_landmarks=None,
ward_predictor='ward')
ctrajs = clusterer.fit_transform(trajs)
lags = (np.arange(1, 50, 1) / to_ns).astype(int)
n_timescales = 50
timescales = np.zeros((lags.shape[0], n_timescales))
eigenvalues = np.zeros((lags.shape[0], n_timescales))
for idx, lag in enumerate(lags):
msm = MarkovStateModel(lag_time=lag, n_timescales=n_timescales)
msm.fit_transform(ctrajs)
timescales[idx] = msm.timescales_
eigenvalues[idx] = msm.eigenvalues_[1:]
def traj_load(irow):
i, row = irow
traj = md.load(row['traj_fn'], top=tops[row['top_fn']])
return i, traj
traj_dict = dict(map(traj_load, meta.iterrows()))
trajs = [traj for traj in traj_dict.values()]
# cluster
print('Attempting to cluster')
num_clusters = 20
cluster = LandmarkAgglomerative(n_clusters=num_clusters,
n_landmarks=int(totframes / 100),
linkage='ward',
metric='rmsd')
cluster.fit(trajs)
#
# print('Fitting cluster labels')
# ctraj = {}
# for k, v in traj_dict.items():
# v = cluster.partial_predict(v)
# diff = nframes-v.shape[0]
# v = np.append(v, np.zeros(diff)-1)
# ctraj[k] = v
# Convert to DF for plotting and sampling.
# df = to_dataframe(ctraj, nframes, dt=1)
ctraj_path = 'ctraj-200'
if isdir(ctraj_path):
meta, all_ctrajs_dict = load_trajs(ctraj_path)
else:
def traj_load(irow):
i, row = irow
traj = md.load(row['traj_fn'], top=tops[row['top_fn']])
return i, traj
traj_dict = dict(map(traj_load, meta.iterrows()))
all_trajs = [traj for traj in traj_dict.values()]
cluster = LandmarkAgglomerative(n_clusters=200, n_landmarks=int(totframes /200), linkage='ward', metric='rmsd')
cluster.fit(all_trajs)
# TODO will this work?
args = [(k,v,cluster) for k, v in traj_dict.items()]
with Pool() as pool:
all_ctrajs_dict = dict(pool.imap_unordered(clust, args))
save_generic(cluster, 'cluster-200')
save_trajs(all_ctrajs_dict, 'ctraj-200', meta)
long_ctrajs = [np.squeeze(traj) for traj in all_ctrajs_dict.values() if traj.shape[0] > 1000]
all_ctrajs = [np.squeeze(traj) for traj in all_ctrajs_dict.values()]
lags = np.concatenate((np.arange(200, 1000, 200),np.arange(1000, 5000, 500)))
all_msms = []
