def hill_climbing(pattern=None,
restore_agent_from='data/Prior.ckpt',
scoring_function='tanimoto',
scoring_function_kwargs=None,
save_dir=None,
learning_rate=0.0005,
batch_size=64,
n_steps=10,
num_processes=0,
use_custom_voc="data/Voc"):
voc = Vocabulary(init_from_file=use_custom_voc)
start_time = time.time()
if pattern:
Agent = scaffold_constrained_RNN(voc)
else:
Agent = RNN(voc)
logger = VizardLog('data/logs')
if torch.cuda.is_available():
Agent.rnn.load_state_dict(torch.load(restore_agent_from))
else:
Agent.rnn.load_state_dict(
torch.load(restore_agent_from,
map_location=lambda storage, loc: storage))
optimizer = torch.optim.Adam(Agent.rnn.parameters(), lr=learning_rate)
# Scoring_function
scoring_function = get_scoring_function(scoring_function=scoring_function,
num_processes=num_processes,
**scoring_function_kwargs)
# For policy based RL, we normally train on-policy and correct for the fact that more likely actions
# occur more often (which means the agent can get biased towards them). Using experience replay is
# therefor not as theoretically sound as it is for value based RL, but it seems to work well.
experience = Experience(voc)
# Log some network weights that can be dynamically plotted with the Vizard bokeh app
logger.log(Agent.rnn.gru_2.weight_ih.cpu().data.numpy()[::100],
"init_weight_GRU_layer_2_w_ih")
logger.log(Agent.rnn.gru_2.weight_hh.cpu().data.numpy()[::100],
"init_weight_GRU_layer_2_w_hh")
logger.log(Agent.rnn.embedding.weight.cpu().data.numpy()[::30],
"init_weight_GRU_embedding")
logger.log(Agent.rnn.gru_2.bias_ih.cpu().data.numpy(),
"init_weight_GRU_layer_2_b_ih")
logger.log(Agent.rnn.gru_2.bias_hh.cpu().data.numpy(),
"init_weight_GRU_layer_2_b_hh")
# Information for the logger
step_score = [[], []]
print("Model initialized, starting training...")
for step in range(n_steps):
# Sample from Agent
if pattern:
seqs, agent_likelihood, entropy = Agent.sample(pattern, batch_size)
else:
seqs, agent_likelihood, entropy = Agent.sample(batch_size)
gc.collect()
# Remove duplicates, ie only consider unique seqs
unique_idxs = unique(seqs)
seqs = seqs[unique_idxs]
agent_likelihood = agent_likelihood[unique_idxs]
entropy = entropy[unique_idxs]
# Get prior likelihood and score
smiles = seq_to_smiles(seqs, voc)
score = scoring_function(smiles)
new_experience = zip(smiles, score, agent_likelihood)
experience.add_experience(new_experience)
indexes = np.flip(np.argsort(np.array(score)))
# Train the agent for 10 epochs on hill-climbing procedure
for epoch in range(10):
loss = Variable(torch.zeros(1))
counter = 0
seen_seqs = []
for j in indexes:
if counter > 50:
break
seq = seqs[j]
s = smiles[j]
if s not in seen_seqs:
seen_seqs.append(s)
log_p, _ = Agent.likelihood(Variable(seq).view(1, -1))
loss -= log_p.mean()
counter += 1
loss /= counter
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Print some information for this step
time_elapsed = (time.time() - start_time) / 3600
time_left = (time_elapsed * ((n_steps - step) / (step + 1)))
print(
"\n Step {} Fraction valid SMILES: {:4.1f} Time elapsed: {:.2f}h Time left: {:.2f}h"
.format(step,
fraction_valid_smiles(smiles) * 100, time_elapsed,
time_left))
print(" Agent Prior Target Score SMILES")
for i in range(10):
print(" {:6.2f} {}".format(score[i], smiles[i]))
# Need this for Vizard plotting
step_score[0].append(step + 1)
step_score[1].append(np.mean(score))
# Log some weights
logger.log(Agent.rnn.gru_2.weight_ih.cpu().data.numpy()[::100],
"weight_GRU_layer_2_w_ih")
logger.log(Agent.rnn.gru_2.weight_hh.cpu().data.numpy()[::100],
"weight_GRU_layer_2_w_hh")
logger.log(Agent.rnn.embedding.weight.cpu().data.numpy()[::30],
"weight_GRU_embedding")
logger.log(Agent.rnn.gru_2.bias_ih.cpu().data.numpy(),
"weight_GRU_layer_2_b_ih")
logger.log(Agent.rnn.gru_2.bias_hh.cpu().data.numpy(),
"weight_GRU_layer_2_b_hh")
logger.log("\n".join([smiles + "\t" + str(round(score, 2)) for smiles, score in zip \
(smiles[:12], score[:12])]), "SMILES", dtype="text", overwrite=True)
logger.log(np.array(step_score), "Scores")
# If the entire training finishes, we create a new folder where we save this python file
# as well as some sampled sequences and the contents of the experinence (which are the highest
# scored sequences seen during training)
if not save_dir:
save_dir = 'data/results/run_' + time.strftime("%Y-%m-%d-%H_%M_%S",
time.localtime())
try:
os.makedirs(save_dir)
except:
print("Folder already existing... overwriting previous results")
copyfile('train_agent.py', os.path.join(save_dir, "train_agent.py"))
experience.print_memory(os.path.join(save_dir, "memory"))
torch.save(Agent.rnn.state_dict(), os.path.join(save_dir, 'Agent.ckpt'))
previous_smiles = []
with open(os.path.join(save_dir, "memory.smi"), 'w') as f:
for i, exp in enumerate(experience.memory):
try:
if Chem.MolToSmiles(
Chem.rdmolops.RemoveStereochemistry(
Chem.MolFromSmiles(
exp[0]))) not in previous_smiles:
f.write("{}\n".format(exp[0]))
previous_smiles.append(
Chem.MolToSmiles(
Chem.rdmolops.RemoveStereochemistry(
Chem.MolFromSmiles(exp[0]))))
except:
pass
def train_agent(restore_prior_from='data/Prior.ckpt',
restore_agent_from='data/Prior.ckpt',
voc_file='data/Voc',
molscore_config=None,
learning_rate=0.0005,
batch_size=64, n_steps=3000, sigma=60,
experience_replay=0):
voc = Vocabulary(init_from_file=voc_file)
start_time = time.time()
# Scoring_function
scoring_function = MolScore(molscore_config)
scoring_function.log_parameters({'batch_size': batch_size, 'sigma': sigma})
print("Building RNNs")
Prior = RNN(voc)
Agent = RNN(voc)
# By default restore Agent to same model as Prior, but can restore from already trained Agent too.
# Saved models are partially on the GPU, but if we dont have cuda enabled we can remap these
# to the CPU.
if torch.cuda.is_available():
print("Cuda available, loading prior & agent")
Prior.rnn.load_state_dict(torch.load(restore_prior_from))
Agent.rnn.load_state_dict(torch.load(restore_agent_from))
else:
print("Cuda not available, remapping to cpu")
Prior.rnn.load_state_dict(torch.load(restore_prior_from, map_location=lambda storage, loc: storage))
Agent.rnn.load_state_dict(torch.load(restore_agent_from, map_location=lambda storage, loc: storage))
# We dont need gradients with respect to Prior
for param in Prior.rnn.parameters():
param.requires_grad = False
optimizer = torch.optim.Adam(Agent.rnn.parameters(), lr=learning_rate)
# For logging purposes let's save some training parameters not captured by molscore
with open(os.path.join(scoring_function.save_dir, 'reinvent_parameters.txt'), 'wt') as f:
[f.write(f'{p}: {v}\n') for p, v in {'learning_rate': learning_rate, 'batch_size': batch_size,
'n_steps': n_steps, 'sigma': sigma,
'experience_replay': experience_replay}.items()]
# For policy based RL, we normally train on-policy and correct for the fact that more likely actions
# occur more often (which means the agent can get biased towards them). Using experience replay is
# therefore not as theoretically sound as it is for value based RL, but it seems to work well.
experience = Experience(voc)
print("Model initialized, starting training...")
for step in range(n_steps):
# Sample from Agent
seqs, agent_likelihood, entropy = Agent.sample(batch_size)
# Remove duplicates, ie only consider unique seqs
unique_idxs = unique(seqs)
seqs = seqs[unique_idxs]
agent_likelihood = agent_likelihood[unique_idxs]
entropy = entropy[unique_idxs]
# Get prior likelihood and score
prior_likelihood, _ = Prior.likelihood(Variable(seqs))
smiles = seq_to_smiles(seqs, voc)
# Using molscore instead here
try:
score = scoring_function(smiles, step=step)
augmented_likelihood = prior_likelihood + sigma * Variable(score)
except: # If anything goes wrong with molscore, write scores and save .ckpt and kill monitor
with open(os.path.join(scoring_function.save_dir,
f'failed_smiles_{scoring_function.step}.smi'), 'wt') as f:
[f.write(f'{smi}\n') for smi in smiles]
torch.save(Agent.rnn.state_dict(),
os.path.join(scoring_function.save_dir, f'Agent_{step}.ckpt'))
scoring_function.write_scores()
scoring_function.kill_dash_monitor()
raise
# Calculate loss
loss = torch.pow((augmented_likelihood - agent_likelihood), 2)
# Experience Replay
# First sample
if experience_replay and len(experience)>4:
exp_seqs, exp_score, exp_prior_likelihood = experience.sample(4)
exp_agent_likelihood, exp_entropy = Agent.likelihood(exp_seqs.long())
exp_augmented_likelihood = exp_prior_likelihood + sigma * exp_score
exp_loss = torch.pow((Variable(exp_augmented_likelihood) - exp_agent_likelihood), 2)
loss = torch.cat((loss, exp_loss), 0)
agent_likelihood = torch.cat((agent_likelihood, exp_agent_likelihood), 0)
# Then add new experience
prior_likelihood = prior_likelihood.data.cpu().numpy()
new_experience = zip(smiles, score, prior_likelihood)
experience.add_experience(new_experience)
# Calculate loss
loss = loss.mean()
# Add regularizer that penalizes high likelihood for the entire sequence
loss_p = - (1 / agent_likelihood).mean()
loss += 5 * 1e3 * loss_p
# Calculate gradients and make an update to the network weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Convert to numpy arrays so that we can print them
augmented_likelihood = augmented_likelihood.data.cpu().numpy()
agent_likelihood = agent_likelihood.data.cpu().numpy()
# Print some information for this step
time_elapsed = (time.time() - start_time) / 3600
time_left = (time_elapsed * ((n_steps - step) / (step + 1)))
print(f"\n Step {step} Fraction valid SMILES: {fraction_valid_smiles(smiles) * 100:4.1f}\
Time elapsed: {time_elapsed:.2f}h Time left: {time_left:.2f}h")
print(" Agent Prior Target Score SMILES")
for i in range(10):
print(f" {agent_likelihood[i]:6.2f} {prior_likelihood[i]:6.2f} {augmented_likelihood[i]:6.2f} {score[i]:6.2f} {smiles[i]}")
# Save the agent weights every 250 iterations ####
if step % 250 == 0 and step != 0:
torch.save(Agent.rnn.state_dict(),
os.path.join(scoring_function.save_dir, f'Agent_{step}.ckpt'))
# If the entire training finishes, write out MolScore dataframe, kill dash_utils monitor and
# save the final Agent.ckpt
torch.save(Agent.rnn.state_dict(), os.path.join(scoring_function.save_dir, f'Agent_{n_steps}.ckpt'))
scoring_function.write_scores()
scoring_function.kill_dash_monitor()
return
def train_agent(restore_prior_from='data/Prior.ckpt',
restore_agent_from='data/Prior.ckpt',
scoring_function='tanimoto',
scoring_function_kwargs=None,
save_dir=None,
learning_rate=0.0005,
batch_size=64,
n_steps=3000,
num_processes=0,
sigma=60,
experience_replay=0):
voc = Vocabulary(init_from_file="data/Voc")
start_time = time.time()
Prior = RNN(voc)
Agent = RNN(voc)
logger = VizardLog('data/logs')
# By default restore Agent to same model as Prior, but can restore from already trained Agent too.
# Saved models are partially on the GPU, but if we dont have cuda enabled we can remap these
# to the CPU.
if torch.cuda.is_available():
Prior.rnn.load_state_dict(torch.load(restore_prior_from))
Agent.rnn.load_state_dict(torch.load(restore_agent_from))
else:
Prior.rnn.load_state_dict(
torch.load(restore_prior_from,
map_location=lambda storage, loc: storage))
Agent.rnn.load_state_dict(
torch.load(restore_agent_from,
map_location=lambda storage, loc: storage))
# We dont need gradients with respect to Prior
for param in Prior.rnn.parameters():
param.requires_grad = False
optimizer = torch.optim.Adam(Agent.rnn.parameters(), lr=0.0005)
# Scoring_function
scoring_function = get_scoring_function(scoring_function=scoring_function,
num_processes=num_processes,
**scoring_function_kwargs)
# For policy based RL, we normally train on-policy and correct for the fact that more likely actions
# occur more often (which means the agent can get biased towards them). Using experience replay is
# therefor not as theoretically sound as it is for value based RL, but it seems to work well.
experience = Experience(voc)
# Log some network weights that can be dynamically plotted with the Vizard bokeh app
logger.log(Agent.rnn.gru_2.weight_ih.cpu().data.numpy()[::100],
"init_weight_GRU_layer_2_w_ih")
logger.log(Agent.rnn.gru_2.weight_hh.cpu().data.numpy()[::100],
"init_weight_GRU_layer_2_w_hh")
logger.log(Agent.rnn.embedding.weight.cpu().data.numpy()[::30],
"init_weight_GRU_embedding")
logger.log(Agent.rnn.gru_2.bias_ih.cpu().data.numpy(),
"init_weight_GRU_layer_2_b_ih")
logger.log(Agent.rnn.gru_2.bias_hh.cpu().data.numpy(),
"init_weight_GRU_layer_2_b_hh")
# Information for the logger
step_score = [[], []]
print("Model initialized, starting training...")
for step in range(n_steps):
# Sample from Agent
seqs, agent_likelihood, entropy = Agent.sample(batch_size)
# Remove duplicates, ie only consider unique seqs
unique_idxs = unique(seqs)
seqs = seqs[unique_idxs]
agent_likelihood = agent_likelihood[unique_idxs]
entropy = entropy[unique_idxs]
# Get prior likelihood and score
prior_likelihood, _ = Prior.likelihood(Variable(seqs))
smiles = seq_to_smiles(seqs, voc)
score = scoring_function(smiles)
# Calculate augmented likelihood
augmented_likelihood = prior_likelihood + sigma * Variable(score)
loss = torch.pow((augmented_likelihood - agent_likelihood), 2)
# Experience Replay
# First sample
if experience_replay and len(experience) > 4:
exp_seqs, exp_score, exp_prior_likelihood = experience.sample(4)
exp_agent_likelihood, exp_entropy = Agent.likelihood(
exp_seqs.long())
exp_augmented_likelihood = exp_prior_likelihood + sigma * exp_score
exp_loss = torch.pow(
(Variable(exp_augmented_likelihood) - exp_agent_likelihood), 2)
loss = torch.cat((loss, exp_loss), 0)
agent_likelihood = torch.cat(
(agent_likelihood, exp_agent_likelihood), 0)
# Then add new experience
prior_likelihood = prior_likelihood.data.cpu().numpy()
new_experience = zip(smiles, score, prior_likelihood)
experience.add_experience(new_experience)
# Calculate loss
loss = loss.mean()
# Add regularizer that penalizes high likelihood for the entire sequence
loss_p = -(1 / agent_likelihood).mean()
loss += 5 * 1e3 * loss_p
# Calculate gradients and make an update to the network weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
# Convert to numpy arrays so that we can print them
augmented_likelihood = augmented_likelihood.data.cpu().numpy()
agent_likelihood = agent_likelihood.data.cpu().numpy()
# Print some information for this step
time_elapsed = (time.time() - start_time) / 3600
time_left = (time_elapsed * ((n_steps - step) / (step + 1)))
print(
"\n Step {} Fraction valid SMILES: {:4.1f} Time elapsed: {:.2f}h Time left: {:.2f}h"
.format(step,
fraction_valid_smiles(smiles) * 100, time_elapsed,
time_left))
print(" Agent Prior Target Score SMILES")
for i in range(10):
print(" {:6.2f} {:6.2f} {:6.2f} {:6.2f} {}".format(
agent_likelihood[i], prior_likelihood[i],
augmented_likelihood[i], score[i], smiles[i]))
# Need this for Vizard plotting
step_score[0].append(step + 1)
step_score[1].append(np.mean(score))
# Log some weights
logger.log(Agent.rnn.gru_2.weight_ih.cpu().data.numpy()[::100],
"weight_GRU_layer_2_w_ih")
logger.log(Agent.rnn.gru_2.weight_hh.cpu().data.numpy()[::100],
"weight_GRU_layer_2_w_hh")
logger.log(Agent.rnn.embedding.weight.cpu().data.numpy()[::30],
"weight_GRU_embedding")
logger.log(Agent.rnn.gru_2.bias_ih.cpu().data.numpy(),
"weight_GRU_layer_2_b_ih")
logger.log(Agent.rnn.gru_2.bias_hh.cpu().data.numpy(),
"weight_GRU_layer_2_b_hh")
logger.log("\n".join([smiles + "\t" + str(round(score, 2)) for smiles, score in zip \
(smiles[:12], score[:12])]), "SMILES", dtype="text", overwrite=True)
logger.log(np.array(step_score), "Scores")
# If the entire training finishes, we create a new folder where we save this python file
# as well as some sampled sequences and the contents of the experinence (which are the highest
# scored sequences seen during training)
if not save_dir:
save_dir = 'data/results/run_' + time.strftime("%Y-%m-%d-%H_%M_%S",
time.localtime())
os.makedirs(save_dir)
copyfile('train_agent.py', os.path.join(save_dir, "train_agent.py"))
experience.print_memory(os.path.join(save_dir, "memory"))
torch.save(Agent.rnn.state_dict(), os.path.join(save_dir, 'Agent.ckpt'))
seqs, agent_likelihood, entropy = Agent.sample(256)
prior_likelihood, _ = Prior.likelihood(Variable(seqs))
prior_likelihood = prior_likelihood.data.cpu().numpy()
smiles = seq_to_smiles(seqs, voc)
score = scoring_function(smiles)
with open(os.path.join(save_dir, "sampled"), 'w') as f:
f.write("SMILES Score PriorLogP\n")
for smiles, score, prior_likelihood in zip(smiles, score,
prior_likelihood):
f.write("{} {:5.2f} {:6.2f}\n".format(smiles, score,
prior_likelihood))
Agent.rnn.load_state_dict(torch.load(restore_agent_from, map_location=lambda storage, loc:storage))
# We dont need gradients with respect to Prior
for param in Prior.rnn.parameters():
param.requires_grad = False
optimizer = torch.optim.Adam(Agent.rnn.parameters(), lr=0.0005)
# Scoring function
scoring_function = get_scoring_function(scoring_function=scoring_function, num_processes=num_processes,
**scoring_function_kwargs)
# For policy based RL, we normally train on-policy and correct for the fact that more likely actions
# occur more often (which means the agent can get biased towards them). Using experience replay is
# therefor not as theoretically sound as it is for value based RL, but it seems to work well.
experience = Experience(voc)
# Log some network weights that can be dynamically plotted with the vizard bokeh app
logger.log(Agent.rnn.gru_2.weight_ih.cpu().data.numpy()[::100], "init_weight_GRU_layer_2_w_ih")
logger.log(Agent.rnn.gru_2.weight_hh.cpu().data.numpy()[::100], "init_weight_GRU_layer_2_w_hh")
logger.log(Agent.rnn.embedding.weight.cpu().data.numpy()[::30], "init_weight_GRU_embedding")
logger.log(Agent.rnn.gru_2.bias_ih.cpu().data.numpy(), "init_weight_GRU_layer_2_b_ih")
logger.log(Agent.rnn.gru_2.bias_hh.cpu().data.numpy(), "init_weight_GRU_layer_2_b_hh")
# Information for the logger
step_score = [[], []]
print('Model initialized, starting training...')
for step in range(n_steps):
# sample from agent