def load_from_file(cls, file_path, evaluation_mode=False, LOG=None):
"""
Load a model from specified file path
:param file_path: model file
:param evaluation_mode: training or evaluation mode
:return:
"""
# model parameters
if torch.cuda.is_available():
save_dict = torch.load(file_path)
else:
save_dict = torch.load(file_path,
map_location=lambda storage, loc: storage)
# encoder and decoder params
encoder_params = save_dict.get("encoder_params", {})
decoder_params = save_dict.get("decoder_params", {})
# load model
model = Model(vocabulary=save_dict['vocabulary'],
tokenizer=save_dict.get('tokenizer',
mv.SMILESTokenizer()),
encoder_params=encoder_params,
decoder_params=decoder_params,
max_sequence_length=save_dict['max_sequence_length'])
model.network.encoder.load_state_dict(save_dict["encoder"])
model.network.decoder.load_state_dict(save_dict["decoder"])
if evaluation_mode:
model.network.encoder.eval()
model.network.decoder.eval()
if LOG:
LOG.info(model.network.encoder)
LOG.info(model.network.decoder)
return model
def validation_stat(self, dataloader, model, loss_compute, device, vocab):
pad = cfgd.DATA_DEFAULT['padding_value']
total_loss = 0
n_correct = 0
total_n_trg = 0
total_tokens = 0
tokenizer = mv.SMILESTokenizer()
for i, batch in enumerate(
ul.progress_bar(dataloader, total=len(dataloader))):
src, source_length, trg, src_mask, trg_mask, max_length_target, _ = batch
trg_y = trg[:, 1:].to(device) # skip start token
# number of tokens without padding
ntokens = (trg_y != pad).data.sum()
# Move to GPU
src = src.to(device)
trg = trg[:, :-1].to(device) # save start token, skip end token
src_mask = src_mask.to(device)
trg_mask = trg_mask.to(device)
# Compute loss with teaching forcing
out = model.forward(src, trg, src_mask, trg_mask)
loss = loss_compute(out, trg_y, ntokens)
total_loss += loss
total_tokens += ntokens
# Decode
max_length_target = cfgd.DATA_DEFAULT['max_sequence_length']
smiles = decode(model,
src,
src_mask,
max_length_target,
type='greedy')
# Compute accuracy
for j in range(trg.size()[0]):
seq = smiles[j, :]
target = trg[j]
target = tokenizer.untokenize(
vocab.decode(target.cpu().numpy()))
seq = tokenizer.untokenize(vocab.decode(seq.cpu().numpy()))
if seq == target:
n_correct += 1
# number of samples in current batch
n_trg = trg.size()[0]
# total samples
total_n_trg += n_trg
# Accuracy
accuracy = n_correct * 1.0 / total_n_trg
loss_epoch = total_loss / total_tokens
return loss_epoch, accuracy
def initialize_dataloader(self, data_path, batch_size, vocab, data_type):
# Read train or validation
data = pd.read_csv(os.path.join(data_path, data_type + '.csv'),
sep=",")
dataset = md.Dataset(data=data,
vocabulary=vocab,
tokenizer=mv.SMILESTokenizer(),
prediction_mode=False)
dataloader = torch.utils.data.DataLoader(
dataset,
batch_size,
shuffle=True,
collate_fn=md.Dataset.collate_fn)
return dataloader
def get_model(self, opt, vocab, device):
# Train from scratch or resume training from a given epoch
if opt.starting_epoch == 1:
model = Model.make_model(opt.num_layers, opt.layer_size, opt.cell_type, opt.embedding_layer_size, opt.dropout,
opt.bidirectional, opt.bidirect_model, opt.attn_model, cfgd.DATA_DEFAULT['max_sequence_length'],
vocab, mv.SMILESTokenizer(), self.LOG)
else:
file_name = os.path.join(self.save_path, f'checkpoint/model_{opt.starting_epoch-1}.pt')
model = Model.load_from_file(file_name)
# move to GPU
model.network.encoder.to(device)
model.network.decoder.to(device)
return model
def __init__(self, opt):
self.save_path = os.path.join('experiments', opt.save_directory, opt.test_file_name,
f'evaluation_{opt.epoch}')
global LOG
LOG = ul.get_logger(name="generate",
log_path=os.path.join(self.save_path, 'generate.log'))
LOG.info(opt)
LOG.info("Save directory: {}".format(self.save_path))
# Load vocabulary
with open(os.path.join(opt.data_path, 'vocab.pkl'), "rb") as input_file:
vocab = pkl.load(input_file)
self.vocab = vocab
self.tokenizer = mv.SMILESTokenizer()
def validation_stat(self, dataloader, model, device, vocab):
pad = cfgd.DATA_DEFAULT['padding_value']
total_loss = 0
total_tokens = 0
n_correct = 0
total_n_trg = 0
tokenizer = mv.SMILESTokenizer()
model.network.encoder.eval()
model.network.decoder.eval()
for _, batch in enumerate(ul.progress_bar(dataloader, total=len(dataloader))):
encoder_input, source_length, decoder_output, mask, _, max_length_target, _ = batch
# Move to GPU
encoder_input = encoder_input.to(device)
decoder_output = decoder_output.to(device)
source_length = source_length.to(device)
mask = torch.squeeze(mask, 1).to(device)
# Loss
with torch.no_grad():
loss_b_sq = model.loss_step(encoder_input, source_length, decoder_output, mask, max_length_target, device)
ntokens = (decoder_output != pad).data.sum()
total_tokens += ntokens
total_loss += loss_b_sq.sum()
# Sample using greedy, compute accuracy
predicted_seqs, predicted_nlls = model.greedy_sample(encoder_input, source_length, decoder_output,
mask, device)
for j, seq in enumerate(predicted_seqs):
target = tokenizer.untokenize(vocab.decode(decoder_output[j].cpu().numpy()))
smi = tokenizer.untokenize(vocab.decode(seq.cpu().numpy()))
if smi == target:
n_correct += 1
total_n_trg += decoder_output.shape[0]
accuracy = n_correct*1.0 / total_n_trg
loss = total_loss/total_tokens
return loss, accuracy
# add property name before property change; save to file
property_condition = []
for property_name in cfgd.PROPERTIES:
if property_name == 'LogD':
intervals, _ = property_change_encoder[property_name]
property_condition.extend(intervals)
else:
intervals = property_change_encoder[property_name]
for name in intervals:
property_condition.append("{}_{}".format(property_name, name))
LOG.info("Property condition tokens: {}".format(len(property_condition)))
encoded_file = pdp.save_df_property_encoded(args.input_data_path, property_change_encoder, LOG)
LOG.info("Building vocabulary")
tokenizer = mv.SMILESTokenizer()
smiles_list = pdp.get_smiles_list(args.input_data_path)
vocabulary = mv.create_vocabulary(smiles_list, tokenizer=tokenizer, property_condition=property_condition)
tokens = vocabulary.tokens()
LOG.info("Vocabulary contains %d tokens: %s", len(tokens), tokens)
# Save vocabulary to file
parent_path = uf.get_parent_dir(args.input_data_path)
output_file = os.path.join(parent_path, 'vocab.pkl')
with open(output_file, 'wb') as pickled_file:
pickle.dump(vocabulary, pickled_file)
LOG.info("Save vocabulary to file: {}".format(output_file))
# Split data into train, validation, test
train, validation, test = pdp.split_data(encoded_file, LOG)
