score = np.power(pre_scores.prod(1), 0.333)
for i in range(len(score)):
if np.isnan(score[i]):
score[i] = -1
return score
reward_fun_on = lambda x: 2.5 + scorer(x) #lambda x: reward_aromatic_rings(x)#
model, fitter1, fitter2 = train_policy_gradient(
molecules,
grammar,
EPOCHS=100,
BATCH_SIZE=batch_size,
reward_fun_on=reward_fun_on,
max_steps=max_steps,
lr_off=1e-4,
lr_on=1e-4,
drop_rate=drop_rate,
decoder_type='attention',
plot_prefix='valid ',
dashboard='valid',
save_file='policy_gradient_valid.h5',
smiles_save_file='pg_smiles_valid.h5',
on_policy_loss_type='valid',
off_policy_loss_type='mean') #,
#preload_file='policy_gradient_main.h5')
while True:
next(fitter1)
next(fitter2)
return score
reward_fun = lambda x: 3 * 2.5 + scorer(
x
) # - 0.2*np.array([0 if num is None else num for num in num_aromatic_rings(x)])# + reward_aliphatic_rings(x)# + 0.05*reward_aromatic_rings(x)#lambda x: reward_aromatic_rings(x)# #lambda x: reward_aromatic_rings(x)#
model, fitter1, fitter2 = train_policy_gradient(
molecules,
grammar,
EPOCHS=100,
BATCH_SIZE=batch_size,
reward_fun_on=reward_fun,
max_steps=max_steps,
lr_off=0.0,
lr_on=1e-4,
drop_rate=drop_rate,
decoder_type='attention', #'random',#
plot_prefix='anchor ',
dashboard='anchor', #None,#
save_file='paper/policy_gradient_anchor.h5',
smiles_save_file='paper/pg_smiles_anchor.h5',
on_policy_loss_type='best',
off_policy_loss_type='mean',
preload_file='paper/policy_gradient_baseline.h5',
anchor_file='paper/policy_gradient_baseline.h5',
anchor_weight=1e9)
#
while True:
next(fitter1)
#next(fitter2)
for i in range(len(score)):
if np.isnan(score[i]):
score[i] = -1
return score
reward_fun = lambda x: np.array([1 for _ in x])
model, fitter1, fitter2 = train_policy_gradient(molecules,
grammar,
EPOCHS=100,
BATCH_SIZE=batch_size,
reward_fun_on=reward_fun,
max_steps=max_steps,
lr_off=1e-4,
lr_on=0.0,
drop_rate=drop_rate,
decoder_type='attention',#'action',#'random',#'action',#'step,
plot_prefix='baseline ',
dashboard='baseline',#None,#
save_file='paper/policy_gradient_baseline_deleteme.h5',
smiles_save_file='paper/pg_smiles_baseline_deleteme.h5',
on_policy_loss_type='mean',
off_policy_loss_type='mean',
sanity_checks=False) # some of the database molecules don't conform to our constraints
# preload_file='policy_gradient_tmp.h5',
# anchor_file='policy_gradient_tmp.h5',
# anchor_weight=1)
#
while True:
next(fitter1)
next(fitter2)