def test_scheme_C2():
# Need to test it as a learning curve, not just pick one
num = 3
atom_df = pd.read_pickle('tests/test_mols/atoms.pkl')
pair_df = pd.read_pickle('tests/test_mols/pairs.pkl')
mol_df, graphs = graphin.make_graph_df(atom_df, pair_df)
total = len(graphs)
counter = 0
chosen = []
while len(chosen) < len(graphs) - num:
# Iterations continue until there are not enough graphs to select
# Each iteration we want (num*n) graphs in train, and (total-(num*n)) graphs in test
# Make the selection the same each time, we just want to select chosen from the df
# We then just grow chosen each time using the selection scheme
counter += 1
chosen = select_molecules_C2(mol_df,
atom_df,
pair_df,
prev_chosen=chosen,
num=num)
train_graphs, train_mol_df, test_graphs, test_mol_df = get_split(
chosen, mol_df, graphs)
assert len(train_graphs) == num * counter
assert len(test_graphs) == total - (num * counter)
assert len(train_mol_df.molecule_name.unique()) == len(
train_mol_df.molecule_name)
assert len(test_mol_df.molecule_name.unique()) == len(
test_mol_df.molecule_name)
assert counter < total / num
for molname in train_mol_df.molecule_name.unique():
assert not molname in test_mol_df.molecule_name.unique()
for molname1 in train_mol_df.molecule_name:
mol_df1 = atom_df.loc[(atom_df.molecule_name == molname1)]["conn"]
for molname2 in test_mol_df.molecule_name:
mol_df2 = atom_df.loc[(
atom_df.molecule_name == molname1)]["conn"]
count1 = [bond for conn in mol_df1.values
for bond in conn].count(2)
count2 = [bond for conn in mol_df2.values
for bond in conn].count(2)
assert count1 >= count2
def test_scheme_D6():
num = 3
atom_df = pd.read_pickle('tests/test_mols/atoms.pkl')
pair_df = pd.read_pickle('tests/test_mols/pairs.pkl')
mol_df, graphs = graphin.make_graph_df(atom_df, pair_df)
total = len(graphs)
counter = 0
chosen = []
while len(chosen) < len(graphs) - num:
# Iterations continue until there are not enough graphs to select
# Each iteration we want (num*n) graphs in train, and (total-(num*n)) graphs in test
# Make the selection the same each time, we just want to select chosen from the df
# We then just grow chosen each time using the selection scheme
counter += 1
chosen = select_molecules_D6(mol_df,
atom_df,
pair_df,
prev_chosen=chosen,
num=num)
train_graphs, train_mol_df, test_graphs, test_mol_df = get_split(
chosen, mol_df, graphs)
assert len(train_graphs) == num * counter
assert len(test_graphs) == total - (num * counter)
assert len(train_mol_df.molecule_name.unique()) == len(
train_mol_df.molecule_name)
assert len(test_mol_df.molecule_name.unique()) == len(
test_mol_df.molecule_name)
assert counter < total / num
for molname1 in train_mol_df.molecule_name:
mol_df1 = atom_df.loc[(
atom_df.molecule_name == molname1)]["typestr"]
count1 = 0
for type in mol_df1.values:
if type not in ['H', 'C', 'N', 'O', 'F']:
count1 += 1
for molname2 in test_mol_df.molecule_name:
mol_df2 = atom_df.loc[(
atom_df.molecule_name == molname2)]["typestr"]
count2 = 0
for type in mol_df2.values:
if type not in ['H', 'C', 'N', 'O', 'F']:
count2 += 1
assert count1 >= count2
def test_scheme_B2():
# Need to test it as a learning curve, not just pick one
num = 3
atom_df = pd.read_pickle('tests/test_mols/atoms.pkl')
pair_df = pd.read_pickle('tests/test_mols/pairs.pkl')
mol_df, graphs = graphin.make_graph_df(atom_df, pair_df)
total = len(graphs)
counter = 0
chosen = []
while len(chosen) < len(graphs) - num:
# Iterations continue until there are not enough graphs to select
# Each iteration we want (num*n) graphs in train, and (total-(num*n)) graphs in test
# Make the selection the same each time, we just want to select chosen from the df
# We then just grow chosen each time using the selection scheme
counter += 1
chosen = select_molecules_B2(mol_df,
atom_df,
pair_df,
prev_chosen=chosen,
num=num)
train_graphs, train_mol_df, test_graphs, test_mol_df = get_split(
chosen, mol_df, graphs)
assert len(train_graphs) == num * counter
assert len(test_graphs) == total - (num * counter)
assert len(train_mol_df.molecule_name.unique()) == len(
train_mol_df.molecule_name)
assert len(test_mol_df.molecule_name.unique()) == len(
test_mol_df.molecule_name)
assert counter < total / num
for molname in train_mol_df.molecule_name.unique():
assert not molname in test_mol_df.molecule_name.unique()
for graph1 in train_graphs:
for graph2 in test_graphs:
assert graph1.number_of_nodes() <= graph2.number_of_nodes()
def add_IMP_to_df(atom_df, pair_df, model_dir):
mol_df, graphs = graph_in.make_graph_df(atom_df, pair_df)
print('Making predictions:')
modelfile = model_dir + 'all_model.torch'
model = GTNmodel()
model.load_model(modelfile)
model.params['batch_size'] = 4
test_loader = model.get_input(graphs, mol_df)
graphs_out = model.predict(test_loader, progress=True)
atom_df, pair_df = model.assign_preds(graphs_out,
mol_df,
atom_df,
pair_df,
assign_to="",
progress=True)
return atom_df, pair_df
def test_scheme_I7():
num = 3
atom_df = pd.read_pickle('tests/test_mols/atoms.pkl')
pair_df = pd.read_pickle('tests/test_mols/pairs.pkl')
mol_df, graphs = graphin.make_graph_df(atom_df, pair_df)
total = len(graphs)
counter = 0
chosen = []
while len(chosen) < len(graphs) - num:
atom_df = add_randomised_FEPs(atom_df)
atom_df = add_randomised_FEP_vars(atom_df)
counter += 1
chosen = select_molecules_I7(mol_df,
atom_df,
pair_df,
prev_chosen=chosen,
num=num)
train_graphs, train_mol_df, test_graphs, test_mol_df = get_split(
chosen, mol_df, graphs)
assert len(train_graphs) == num * counter, print(
len(train_graphs), num, counter)
assert len(test_graphs) == total - (num * counter)
assert len(train_mol_df.molecule_name.unique()) == len(
train_mol_df.molecule_name)
assert len(test_mol_df.molecule_name.unique()) == len(
test_mol_df.molecule_name)
assert counter < total / num
for molname in train_mol_df.molecule_name.unique():
assert not molname in test_mol_df.molecule_name.unique()
assert counter == 2
assert len(chosen) == len(graphs) - len(graphs) % 3