fix tram bug (#209)

deeptime/markov/msm/tram/_tram.py

@@ -238,7 +238,7 @@ def _make_tram_estimator(self, model, dataset):
                 # copy free energies along the markoc state axis to get initial biased_conf_energies
                 biased_conf_energies = np.repeat(free_energies[:, None], dataset.n_markov_states, axis=1)
             else:
-                biased_conf_energies = np.zeros((dataset.n_markov_states, dataset.n_therm_states))
+                biased_conf_energies = np.zeros((dataset.n_therm_states, dataset.n_markov_states))
 
         lagrangian_mult_log = tram.initialize_lagrangians(dataset.transition_counts)
         modified_state_counts = np.zeros_like(lagrangian_mult_log)  # intialize this as the dataset state counts???

deeptime/src/include/deeptime/markov/msm/tram/mbar.h

@@ -63,7 +63,7 @@ initialize_MBAR(BiasMatrix <dtype> biasMatrix, CountsMatrix stateCounts, std::si
                 double maxErr = 1e-6, std::size_t callbackInterval = 1, const py::object *callback = nullptr) {
     // get dimensions...
     auto nThermStates = stateCounts.shape(0);
-    auto nSamples = biasMatrix.shape(1);
+    auto nSamples = biasMatrix.shape(0);
 
     // work in log space so compute the log of the statecounts beforehand
     auto stateCountsLog = std::vector<dtype>(nThermStates);

devtools/conda-setup+build.yml

@@ -5,6 +5,7 @@ steps:
     conda config --set quiet true
   displayName: Configure conda
 - bash: |
+    conda clean --all
     conda install mamba
     mamba update --all
     mamba install boa conda-build conda-verify pip

examples/methods/plot_tram.py

@@ -76,7 +76,7 @@ def sample_trajectories(bias_functions):
     dtrajs = clustering.transform(trajectories.flatten()).reshape((len(bias_matrices), n_samples))
 
     from tqdm import tqdm
-    tram = TRAM(lagtime=1, maxiter=1000, maxerr=1e-3, progress=tqdm)
+    tram = TRAM(lagtime=1, maxiter=1000, maxerr=1e-3, progress=tqdm, init_strategy="MBAR")
 
     # For every simulation frame seen in trajectory i and time step t, btrajs[i][t,k] is the
     # bias energy of that frame evaluated in the k'th thermodynamic state (i.e. at the k'th

tests/markov/msm/test_tram.py

@@ -265,12 +265,24 @@ def test_tqdm_progress_bar():
     tram.fit(make_random_input_data(5, 5))
 
 
-def test_fit_with_dataset():
+@pytest.mark.parametrize(
+    "init_strategy", ["MBAR", None]
+)
+def test_fit_with_dataset(init_strategy):
     dataset = TRAMDataset(dtrajs=[np.asarray([0, 1, 2])], bias_matrices=[np.asarray([[1.], [2.], [3.]])])
-    tram = TRAM()
+    tram = TRAM(init_strategy=init_strategy)
     tram.fit(dataset)
 
 
+@pytest.mark.parametrize(
+    "init_strategy", ["MBAR", None]
+)
+def test_fit_with_dataset(init_strategy):
+    input_data = make_random_input_data(20, 2)
+    tram = TRAM(init_strategy=init_strategy)
+    tram.fit(input_data)
+
+
 def test_mbar_initalization():
     (dtrajs, bias_matrices) = make_random_input_data(5, 5, make_ttrajs=False)
     tram = TRAM(callback_interval=2, maxiter=0, progress=tqdm, init_maxiter=100)
@@ -296,3 +308,4 @@ def test_mbar_initialization_zero_iterations():
     model1 = tram1.fit_fetch(input_data)
     model2 = tram2.fit_fetch(input_data)
     np.testing.assert_equal(model1.biased_conf_energies, model2.biased_conf_energies)
+