我正在使用OpenAI健身库。我已经下载了一个新的环境，但是我得到了这个错误。
AssertionError: The environment must specify an action space. https://www.gymlibrary.dev/content/environment_creation/它说我没有指定动作空间。但我指定了。
class MoleculeEnv(gym.Env):
metadata = {'render.modes': ['human']}
def init(self, docking_config=dict(), data_type='zinc',ratios=dict(),reward_step_total=1,is_normalize=0,reward_type='crystal',reward_target=0.5,has_scaffold=False, has_feature=False,is_conditional=False,conditional='low',max_action=128,min_action=20,force_final=False):
'''
own init function, since gym does not support passing argument
'''
self.is_normalize = bool(is_normalize)
self.has_feature = has_feature
# init smi
self.starting_smi = 'c1([*:1])c([*:2])ccc([*:3])c1' # for hit (benzene ring)
# self.starting_smi = '[*:1]c1ccc2[nH]c(-c3cc([*:2])cc(-c4cccc([*:3])c4)c3O)cc2c1' # fa7 scaffold
# self.starting_smi = 'O=C(c1cccc(Cc2c([*:1])[nH]c(=O)c3cc([*:2])c([*:3])n23)c1)N1CCN([*:4])CC1' # parp1 scaffold
# self.starting_smi = 'C1=C([*:1])C2=NC=C(CC([*:2])C[NH+]3CCC([*:3])C([*:4])C3)[C@H]2C=C1[*:5]' # 5ht1b scaffold
self.smi = self.starting_smi
self.mol = Chem.MolFromSmiles(self.smi)
self.smile_list = []
self.smile_old_list = []
possible_atoms = ATOM_VOCAB
possible_motifs = FRAG_VOCAB
possible_bonds = [Chem.rdchem.BondType.SINGLE, Chem.rdchem.BondType.DOUBLE,
Chem.rdchem.BondType.TRIPLE, Chem.rdchem.BondType.AROMATIC]
self.atom_type_num = len(possible_atoms)
self.motif_type_num = len(possible_motifs)
self.possible_atom_types = np.array(possible_atoms)
self.possible_motif_types = np.array(possible_motifs)
self.possible_bond_types = np.array(possible_bonds, dtype=object)
self.d_n = len(self.possible_atom_types)+18
self.max_action = max_action
self.min_action = min_action
self.max_atom = 150
self.action_space = gym.spaces.MultiDiscrete([20, len(FRAG_VOCAB), 20])
self.counter = 0
self.level = 0 # for curriculum learning, level starts with 0, and increase afterwards
self.predictor = DockingVina(docking_config)
self.attach_point = Chem.MolFromSmiles('*')
self.Na = Chem.MolFromSmiles('[Na+]')
self.K = Chem.MolFromSmiles('[K+]')
self.H = Chem.MolFromSmiles('[H]')
def seed(self,seed):
np.random.seed(seed=seed)
random.seed(seed)
def normalize_adj(self,adj):
degrees = np.sum(adj,axis=2)
D = np.zeros((adj.shape[0],adj.shape[1],adj.shape[2]))
for i in range(D.shape[0]):
D[i,:,:] = np.diag(np.power(degrees[i,:],-0.5))
adj_normal = D@adj@D
adj_normal[np.isnan(adj_normal)]=0
return adj_normal
def reset_batch(self):
self.smile_list = []
def reward_batch(self):
reward = []
print('smiles list', self.smile_list)
return reward_vina(self.smile_list, self.predictor)
def reward_single(self, smile_list):
reward = []
print('smiles list', smile_list)
return reward_vina(smile_list, self.predictor)
def step(self, ac):
"""
Perform a given action
:param action:
:param action_type:
:return: reward of 1 if resulting molecule graph does not exceed valency,
-1 if otherwise
"""
ac = ac[0]
### init
info = {} # info we care about
self.mol_old = copy.deepcopy(self.mol) # keep old mol
stop = False
new = False
if (self.counter >= self.max_action) or get_att_points(self.mol) == []:
new = True
else:
self._add_motif(ac) # problems here
reward_step = 0.05
if self.mol.GetNumAtoms() > self.mol_old.GetNumAtoms():
reward_step += 0.005
self.counter += 1
if new:
reward = 0
# Only store for obs if attachment point exists in o2
if get_att_points(self.mol) != []:
mol_no_att = self.get_final_mol()
Chem.SanitizeMol(mol_no_att, sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE)
smi_no_att = Chem.MolToSmiles(mol_no_att)
info['smile'] = smi_no_att
print("smi:", smi_no_att)
self.smile_list.append(smi_no_att)
# Info for old mol
mol_old_no_att = self.get_final_mol_ob(self.mol_old)
Chem.SanitizeMol(mol_old_no_att, sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE)
smi_old_no_att = Chem.MolToSmiles(mol_no_att)
info['old_smi'] = smi_old_no_att
self.smile_old_list.append(smi_old_no_att)
stop = True
else:
stop = False
self.counter = 0
### use stepwise reward
else:
reward = reward_step
info['stop'] = stop
# get observation
ob = self.get_observation()
return ob,reward,new,info
def reset(self,smile=None):
'''
to avoid error, assume an atom already exists
:return: ob
'''
if smile is not None:
self.mol = Chem.RWMol(Chem.MolFromSmiles(smile))
Chem.SanitizeMol(self.mol, sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE)
else:
# init smi
self.smi = self.starting_smi
self.mol = Chem.MolFromSmiles(self.smi)
# self.smile_list = [] # only for single motif
self.counter = 0
ob = self.get_observation()
return ob
def render(self, mode='human', close=False):
return
def sample_motif(self):
go_on = True
while go_on:
cur_mol_atts = get_att_points(self.mol)
ac1 = np.random.randint(len(cur_mol_atts))
ac2 = np.random.randint(self.motif_type_num)
motif = FRAG_VOCAB_MOL[ac2]
ac3 = np.random.randint(len(FRAG_VOCAB_ATT[ac2]))
a = self.action_space.sample()
a[0] = ac1
a[1] = ac2
a[2] = ac3
go_on = False
return a
def _add_motif(self, ac):
cur_mol = Chem.ReplaceSubstructs(self.mol, self.attach_point, self.Na)[ac[0]]
motif = FRAG_VOCAB_MOL[ac[1]]
att_point = FRAG_VOCAB_ATT[ac[1]]
motif_atom = map_idx(ac[2], att_point, motif)
motif = Chem.ReplaceSubstructs(motif, self.attach_point, self.K)[ac[2]]
motif = Chem.DeleteSubstructs(motif, self.K)
next_mol = Chem.ReplaceSubstructs(cur_mol, self.Na, motif, replacementConnectionPoint=motif_atom)[0]
self.mol = next_mol
def get_final_smiles_mol(self):
"""
Returns a SMILES of the final molecule. Converts any radical
electrons into hydrogens. Works only if molecule is valid
:return: SMILES
"""
m = Chem.DeleteSubstructs(self.mol, Chem.MolFromSmiles("*"))
m = convert_radical_electrons_to_hydrogens(m)
return m, Chem.MolToSmiles(m, isomericSmiles=True)
def get_final_mol(self):
"""
Returns a rdkit mol object of the final molecule. Converts any radical
electrons into hydrogens. Works only if molecule is valid
:return: SMILES
"""
m = Chem.DeleteSubstructs(self.mol, Chem.MolFromSmiles("*"))
return m
def get_final_mol_ob(self, mol):
m = Chem.DeleteSubstructs(mol, Chem.MolFromSmiles("*"))
return m
def get_observation(self, expert_smi=None):
"""
ob['adj']:d_e*n*n --- 'E'
ob['node']:1*n*d_n --- 'F'
n = atom_num + atom_type_num
"""
ob = {}
if expert_smi:
mol = Chem.MolFromSmiles(expert_smi)
else:
ob['att'] = get_att_points(self.mol)
mol = copy.deepcopy(self.mol)
try:
Chem.SanitizeMol(mol)
except:
pass
smi = Chem.MolToSmiles(mol)
n = mol.GetNumAtoms()
F = np.zeros((1, self.max_atom, self.d_n))
for a in mol.GetAtoms():
atom_idx = a.GetIdx()
atom_symbol = a.GetSymbol()
if self.has_feature:
float_array = atom_feature(a, use_atom_meta=True)
else:
float_array = (atom_symbol == self.possible_atom_types).astype(float)
F[0, atom_idx, :] = float_array
d_e = len(self.possible_bond_types)
E = np.zeros((d_e, self.max_atom, self.max_atom))
for b in mol.GetBonds():
begin_idx = b.GetBeginAtomIdx()
end_idx = b.GetEndAtomIdx()
bond_type = b.GetBondType()
float_array = (bond_type == self.possible_bond_types).astype(float)
try:
assert float_array.sum() != 0
except:
print('error',bond_type)
E[:, begin_idx, end_idx] = float_array
if self.is_normalize:
E = self.normalize_adj(E)
ob_adj = adj2sparse(E.squeeze())
ob_node = torch.Tensor(F)
g = dgl.DGLGraph()
ob_len = torch.sum(torch.sum(ob_node, dim=-1).bool().float().squeeze(-2), dim=-1)
g.add_nodes(ob_len)
if ob_adj is not None and len(ob_adj[0])>0 :
g.add_edges(ob_adj[0][0], ob_adj[0][1], {'x': ob_adj[1]})
g.ndata['x'] = ob_node[:, :int(ob_len),:].squeeze(0)
ob['g'] = g
ob['smi'] = smi
return ob
def get_observation_mol(self,mol):
"""
ob['adj']:d_e*n*n --- 'E'
ob['node']:1*n*d_n --- 'F'
n = atom_num + atom_type_num
"""
ob = {}
ob['att'] = get_att_points(mol)
try:
Chem.SanitizeMol(mol)
except:
pass
smi = Chem.MolToSmiles(mol)
n = mol.GetNumAtoms()
F = np.zeros((1, self.max_atom, self.d_n))
for a in mol.GetAtoms():
atom_idx = a.GetIdx()
atom_symbol = a.GetSymbol()
if self.has_feature:
float_array = atom_feature(a, use_atom_meta=True)
else:
float_array = (atom_symbol == self.possible_atom_types).astype(float)
F[0, atom_idx, :] = float_array
d_e = len(self.possible_bond_types)
E = np.zeros((d_e, self.max_atom, self.max_atom))
for b in mol.GetBonds():
begin_idx = b.GetBeginAtomIdx()
end_idx = b.GetEndAtomIdx()
bond_type = b.GetBondType()
float_array = (bond_type == self.possible_bond_types).astype(float)
try:
assert float_array.sum() != 0
except:
print('error',bond_type)
E[:, begin_idx, end_idx] = float_array
if self.is_normalize:
E = self.normalize_adj(E)
ob_adj = adj2sparse(E.squeeze())
ob_node = torch.Tensor(F)
g = dgl.DGLGraph()
ob_len = torch.sum(torch.sum(ob_node, dim=-1).bool().float().squeeze(-2), dim=-1)
g.add_nodes(ob_len)
if ob_adj is not None and len(ob_adj[0])>0 :
g.add_edges(ob_adj[0][0], ob_adj[0][1], {'x': ob_adj[1]})
g.ndata['x'] = ob_node[:, :int(ob_len),:].squeeze(0)
ob['g'] = g
ob['smi'] = smi
return ob这是我的env。我指定动作空间：[20，len（FRAG_VOCAB），20]）中的一个或多个变量的值。
我查看了附加到错误的URL，但我仍然不知道如何解决此错误。
如何修复错误？
1条答案
按热度按时间jchrr9hc1#
我把我的gym版本降低到0.18.0。这修正了我的错误。