def main():
f = open("outdir.txt", 'r')
outdir = f.read().rstrip('\n')
#create an experiment folder tied to date and time where to save output from the model
experiment_folder = os.path.expanduser('~/data1/stratification_ILRM/experiments/') + disease_folder +\
'-'.join(map(str, list(datetime.now().timetuple()[:6])))
os.makedirs(experiment_folder)
f = open("experiment_folder.txt", 'w') ##path to the experiment folder is saved in a txt file
f.write(experiment_folder)
f.close()
##pass the size of the vocabulary to the model
with open(os.path.join(outdir, mt_to_ix_file)) as f:
rd = csv.reader(f)
next(rd)
vocab_size = 1
for r in rd:
vocab_size+=1
#set random seed for reproducible experiments
torch.manual_seed(123)
torch.cuda.manual_seed(123)
##Import data
data = myData(outdir, ehr_file)
data_generator = DataLoader(data, model_pars['batch_size'], shuffle=True, collate_fn=my_collate)
#define model and optimizer
print("cohort numerosity:{0} -- max_seq_length:{1}".format(len(data), L))
model = net.ehrEncoding(vocab_size, L, model_pars['embedding_dim'], model_pars['kernel_size'])
optimizer = torch.optim.Adam(model.parameters(), lr=model_pars['learning_rate'], weight_decay=1e-5)
#start the unsupervised training and evaluation
model.cuda()
loss_fn = net.criterion
print("Starting training for {} epochs...".format(model_pars['num_epochs']))
mrn, encoded, metrics_avg = train_and_evaluate(model, data_generator, loss_fn, optimizer, metrics, experiment_folder)
##save encoded vectors, medical record number list (to keep track of the order) and metric (loss and accuracy)
with open(experiment_folder + '/encoded_vect.csv', 'w') as f:
wr = csv.writer(f, delimiter=',')
for e in encoded:
wr.writerow(e)
with open(experiment_folder + '/mrns.csv', 'w') as f:
wr = csv.writer(f, delimiter=',')
for m in mrn:
wr.writerow([m])
with open(experiment_folder + '/metrics.txt', 'w') as f:
wr = csv.writer(f, delimiter='\t')
#for m, v in metrics_average.items():
# wr.writerow([m, v])
wr.writerow(["Mean loss:", metrics_avg['loss']])
wr.writerow(["Accuracy:", metrics_avg['accuracy']])
def main():
##pass the size of the vocabulary to the model
with open(os.path.join(data_folder, mt_to_ix_file)) as f:
rd = csv.reader(f)
vocab_size = 0
for r in rd:
vocab_size += 1
#set random seed for reproducible experiments
torch.manual_seed(123)
torch.cuda.manual_seed(123)
##Import data
data = myData(data_folder, ehr_file)
data_generator = DataLoader(data,
model_pars['batch_size'],
shuffle=True,
collate_fn=my_collate)
#define model and optimizer
print("cohort numerosity:{0} -- max_seq_length:{1}".format(len(data), L))
model = net.ehrEncoding(vocab_size, L, model_pars['embedding_dim'],
model_pars['kernel_size'])
#model = nn.DataParallel(model, device_ids=[1,2,3])
optimizer = torch.optim.Adam(model.parameters(),
lr=model_pars['learning_rate'],
weight_decay=1e-5)
#start the unsupervised training and evaluation
model.cuda()
loss_fn = net.criterion
print("Starting training for {} epochs...".format(
model_pars['num_epochs']))
mrn, encoded, metrics_avg = train_and_evaluate(model, data_generator,
loss_fn, optimizer,
experiment_folder, metrics)
#with open(experiment_folder + '/TRencoded_vect.csv', 'w') as f:
# wr = csv.writer(f, delimiter=',')
# for e in encoded_tr:
# wr.writerow(e)
#with open(experiment_folder + '/TRmrns.csv', 'w') as f:
# wr = csv.writer(f, delimiter=',')
# for m in mrn_tr:
# wr.writerow([m])
#with open(experiment_folder + '/TRmetrics.txt', 'w') as f:
# wr = csv.writer(f, delimiter = '\t')
#for m, v in metrics_average.items():
# wr.writerow([m, v])
# wr.writerow(["Mean loss:", loss_tr])
##load and evaluate best model
#print("Evaluating best model...")
#best_saved = torch.load(experiment_folder + '/best_model.pt')
#model.load_state_dict(best_saved['state_dict'])
#mrn, encoded, metrics_avg = evaluate(model, loss_fn, data_generator, metrics, best_eval=True)
with open(experiment_folder + '/encoded_vect.csv', 'w') as f:
wr = csv.writer(f, delimiter=',')
for e in encoded:
wr.writerow(e)
with open(experiment_folder + '/mrns.csv', 'w') as f:
wr = csv.writer(f, delimiter=',')
for m in mrn:
wr.writerow([m])
with open(experiment_folder + '/metrics.txt', 'w') as f:
wr = csv.writer(f, delimiter='\t')
#for m, v in metrics_average.items():
# wr.writerow([m, v])
wr.writerow(["Mean loss:", metrics_avg['loss']])
wr.writerow(["Accuracy:", metrics_avg['accuracy']])
def learn_patient_representations(indir,
test_set=False,
sampling=None,
emb_filename=None):
# experiment folder with date and time to save the representations
exp_dir = os.path.join(indir, 'encodings')
os.makedirs(exp_dir, exist_ok=True)
# get the vocabulary size
fvocab = os.path.join(os.path.join(indir), ut.dt_files['vocab'])
with open(fvocab) as f:
rd = csv.reader(f)
next(rd)
vocab = {}
for r in rd:
vocab[int(r[1])] = r[0]
vocab_size = len(vocab) + 1
print('Vocabulary size: {0}'.format(vocab_size))
# load pre-computed embeddings
if emb_filename is not None:
model = Word2Vec.load(emb_filename)
embs = model.wv
del model
print('Loaded pre-computed embeddings for {0} concepts'.format(
len(embs.vocab)))
else:
embs = None
# set random seed for experiment reproducibility
torch.manual_seed(123)
torch.cuda.manual_seed(123)
# load data
data_tr = EHRdata(os.path.join(indir), ut.dt_files['ehr-file'], sampling)
data_generator_tr = DataLoader(data_tr,
ut.model_param['batch_size'],
shuffle=True,
collate_fn=ehr_collate)
if test_set:
data_ts = EHRdata(os.path.join(indir), ut.dt_files['ehr-file-test'],
sampling)
data_generator_ts = DataLoader(data_ts,
ut.model_param['batch_size'],
shuffle=True,
collate_fn=ehr_collate)
print("Test cohort size: {0}".format(len(data_ts)))
else:
data_generator_ts = data_generator_tr
print('Training cohort size: {0}\n'.format(len(data_tr)))
print('Max Sequence Length: {0}\n'.format(ut.len_padded))
# define model and optimizer
print('Learning rate: {0}'.format(ut.model_param['learning_rate']))
print('Batch size: {0}'.format(ut.model_param['batch_size']))
print('Kernel size: {0}\n'.format(ut.model_param['kernel_size']))
model = net.ehrEncoding(vocab_size=vocab_size,
max_seq_len=ut.len_padded,
emb_size=ut.model_param['embedding_size'],
kernel_size=ut.model_param['kernel_size'],
pre_embs=embs,
vocab=vocab)
optimizer = torch.optim.Adam(model.parameters(),
lr=ut.model_param['learning_rate'],
weight_decay=ut.model_param['weight_decay'])
# training and evaluation
if torch.cuda.device_count() > 1:
print('No. of GPUs: {0}\n'.format(torch.cuda.device_count()))
model = nn.DataParallel(model)
else:
model.cuda()
print('No. of GPUs: 1\n')
# model.cuda()
loss_fn = net.criterion
print('Training for {} epochs\n'.format(ut.model_param['num_epochs']))
mrn, encoded, encoded_avg, metrics_avg = train_and_evaluate(
model, data_generator_tr, data_generator_ts, loss_fn, optimizer,
net.metrics, exp_dir)
# save results
# encoded vectors (representations)
outfile = os.path.join(exp_dir, 'convae-avg_vect.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "ENCODED-AVG"])
for m, e in zip(mrn, encoded_avg):
wr.writerow([m] + list(e))
outfile = os.path.join(exp_dir, 'convae_vect.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "ENCODED-SUBSEQ"])
for m, evs in zip(mrn, encoded):
for e in evs:
wr.writerow([m] + e)
# metrics (loss and accuracy)
outfile = os.path.join(exp_dir, 'metrics.txt')
with open(outfile, 'w') as f:
f.write('Mean loss: %.3f\n' % metrics_avg['loss'])
f.write('Accuracy: %.3f\n' % metrics_avg['accuracy'])
# ehr subseq with age in days
outfile = os.path.join(exp_dir,
'cohort-ehr-subseq{0}.csv'.format(ut.len_padded))
with open(os.path.join(os.path.join(indir), 'cohort-ehrseq.csv')) as f:
rd = csv.reader(f)
next(rd)
ehr = {}
for r in rd:
ehr.setdefault(r[0], list()).extend(r[1:])
ehr_subseq = {}
for list_m, batch in data_generator_tr:
for b, m in zip(batch, list_m):
if len(b) == 1:
ehr_subseq[m] = b.tolist()
else:
seq = []
for vec in b.tolist():
seq.extend(vec)
nseq, nleft = divmod(len(seq), ut.len_padded)
if nleft > 0:
seq = seq + [0] * \
(ut.len_padded - nleft)
for i in range(0, len(seq) - ut.len_padded + 1, ut.len_padded):
ehr_subseq.setdefault(m, list()).append(seq[i:i +
ut.len_padded])
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "EHRsubseq"])
for m, subseq in ehr_subseq.items():
for seq in subseq:
wr.writerow([m] + list(filter(lambda x: x != 0, seq)))
if test_set:
outfile = os.path.join(
exp_dir, 'cohort_test-ehr-subseq{0}.csv'.format(ut.len_padded))
ehr_subseq = {}
for list_m, batch in data_generator_ts:
for b, m in zip(batch, list_m):
if len(b) == 1:
ehr_subseq[m] = b.tolist()
else:
seq = []
for vec in b.tolist():
seq.extend(vec)
nseq, nleft = divmod(len(seq), ut.len_padded)
if nleft > 0:
seq = seq + [0] * \
(ut.len_padded - nleft)
for i in range(0,
len(seq) - ut.len_padded + 1,
ut.len_padded):
ehr_subseq.setdefault(m, list()).append(
seq[i:i + ut.len_padded])
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "EHRsubseq"])
for m, subseq in ehr_subseq.items():
for seq in subseq:
wr.writerow([m] + list(filter(lambda x: x != 0, seq)))
return
torch.cuda.manual_seed(123)
# load data
data_ts = EHRdata(os.path.join(indir), ut.dt_files['ehr-file-test'], sampling)
data_generator_ts = DataLoader(
data_ts,
ut.model_param['batch_size'], # may need to reduce this
shuffle=False,
collate_fn=ehr_collate)
print("Test cohort size: {0}".format(len(data_ts)))
# define model and optimizer
model = net.ehrEncoding(
vocab_size=vocab_size,
max_seq_len=ut.len_padded, # 32
emb_size=ut.model_param['embedding_size'], # 100
kernel_size=ut.model_param['kernel_size'], # 5
pre_embs=embs,
vocab=vocab)
optimizer = torch.optim.Adam(model.parameters(),
lr=ut.model_param['learning_rate'],
weight_decay=ut.model_param['weight_decay'])
model.cuda()
loss_fn = net.criterion
# use model from checkpoint
checkpoint = torch.load(os.path.join(indir, 'encodings', 'best_model.pt'))
model.load_state_dict(checkpoint['model_state_dict'])
def learn_patient_representations(indir,
outdir,
disease_dt,
eval_baseline=False,
sampling=None,
emb_filename=None):
# experiment folder with date and time to save the representations
exp_dir = os.path.join(
outdir, '-'.join([
disease_dt,
datetime.now().strftime('%Y-%m-%d-%H-%M-%S'), 'w2v-nobn-softplus'
]))
os.makedirs(exp_dir)
# get the vocabulary size
fvocab = os.path.join(indir, ut.dt_files['vocab'])
with open(fvocab) as f:
rd = csv.reader(f)
next(rd)
vocab = {}
for r in rd:
tkn = r[0].split('::')
tkn[1] = tkn[1].capitalize()
vocab[int(r[1])] = '::'.join(tkn)
vocab_size = len(vocab) + 1
print('Vocabulary size: {0}'.format(vocab_size))
# load pre-computed embeddings
if emb_filename is not None:
model = Word2Vec.load(emb_filename)
embs = model.wv
del model
print('Loaded pre-computed embeddings for {0} concepts'.format(
len(embs.vocab)))
else:
embs = None
# set random seed for experiment reproducibility
torch.manual_seed(123)
torch.cuda.manual_seed(123)
# load data
data = EHRdata(indir, ut.dt_files['ehr'], sampling)
data_generator = DataLoader(data,
ut.model_param['batch_size'],
shuffle=True,
collate_fn=ehr_collate)
print('Cohort Size: {0} -- Max Sequence Length: {1}\n'.format(
len(data), ut.len_padded))
# define model and optimizer
print('Learning rate: {0}'.format(ut.model_param['learning_rate']))
print('Batch size: {0}'.format(ut.model_param['batch_size']))
print('Kernel size: {0}\n'.format(ut.model_param['kernel_size']))
model = net.ehrEncoding(vocab_size=vocab_size,
max_seq_len=ut.len_padded,
emb_size=ut.model_param['embedding_size'],
kernel_size=ut.model_param['kernel_size'],
pre_embs=embs,
vocab=vocab)
optimizer = torch.optim.Adam(model.parameters(),
lr=ut.model_param['learning_rate'],
weight_decay=ut.model_param['weight_decay'])
# training and evaluation
if torch.cuda.device_count() > 1:
print('No. of GPUs: {0}\n'.format(torch.cuda.device_count()))
model = nn.DataParallel(model)
else:
model.cuda()
print('No. of GPUs: 1\n')
# model.cuda()
loss_fn = net.criterion
print('Training for {} epochs\n'.format(ut.model_param['num_epochs']))
mrn, encoded, metrics_avg = train_and_evaluate(model, data_generator,
loss_fn, optimizer,
net.metrics, exp_dir)
# save results
# encoded vectors (representations)
outfile = os.path.join(exp_dir, 'encoded_vect.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerows(encoded)
# MRNs to keep track of the order
outfile = os.path.join(exp_dir, 'mrns.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
for m in mrn:
wr.writerow([m])
# metrics (loss and accuracy)
outfile = os.path.join(exp_dir, 'metrics.txt')
with open(outfile, 'w') as f:
f.write('Mean loss: %.3f\n' % metrics_avg['loss'])
f.write('Accuracy: %.3f\n' % metrics_avg['accuracy'])
# evaluate clustering
gt_file = os.path.join(indir, ut.dt_files['diseases'])
gt_disease = clu.load_mrn_disease(gt_file)
min_clu = 2
max_clu = 10
if eval_baseline:
print('\nRunning clustering on the TF-IDF vectors')
datafile = os.path.join(indir, ut.dt_files['ehr'])
mrn_idx, svd_mtx = clu.svd_tfidf(datafile, vocab_size)
gt_disease_raw = [gt_disease[m][0] for m in mrn_idx]
clu.eval_hierarchical_clustering(svd_mtx, gt_disease_raw, min_clu,
max_clu)
print('\nRunning clustering on the encoded vectors')
gt_disease_enc = [gt_disease[m][0] for m in mrn]
clu.eval_hierarchical_clustering(encoded,
gt_disease_enc,
min_clu,
max_clu,
preproc=True)
return
def learn_patient_representations(
indir,
test_set=False,
sampling=None,
emb_filename=None
):
# encodings folder to save the representations
exp_dir = os.path.join(indir, 'encodings')
if test_set:
exp_dir = os.path.join(indir, 'encodings', 'test')
os.makedirs(exp_dir, exist_ok=True)
# get the vocabulary size
vocab_size, vocab = vocabulary.get_vocab(indir)
# load pre-computed embeddings
if emb_filename is not None:
model = Word2Vec.load(emb_filename)
embs = model.wv
del model
print('Loaded pre-computed embeddings for {0} concepts'.format(
len(embs.vocab)))
else:
embs = None
# set random seed for experiment reproducibility
torch.manual_seed(123)
torch.cuda.manual_seed(123)
# load data
data_tr = EHRdata(os.path.join(indir), ut.dt_files['ehr-file'], sampling)
data_generator_tr = DataLoader(
data_tr,
ut.model_param['batch_size'],
shuffle=True,
collate_fn=ehr_collate
)
if test_set:
data_ts = EHRdata(os.path.join(indir), ut.dt_files['ehr-file-test'], sampling)
data_generator_ts = DataLoader(
data_ts,
ut.model_param['batch_size'],
shuffle=True,
collate_fn=ehr_collate
)
print("Test cohort size: {0}".format(len(data_ts)))
else:
data_generator_ts = data_generator_tr
print('Training cohort size: {0}\n'.format(len(data_tr)))
print('Max Sequence Length: {0}\n'.format(ut.len_padded))
print('Learning rate: {0}'.format(ut.model_param['learning_rate']))
print('Batch size: {0}'.format(ut.model_param['batch_size']))
print('Kernel size: {0}\n'.format(ut.model_param['kernel_size']))
# define model and optimizer
model = net.ehrEncoding(
vocab_size=vocab_size,
max_seq_len=ut.len_padded, # 32
emb_size=ut.model_param['embedding_size'], # 100
kernel_size=ut.model_param['kernel_size'], # 5
pre_embs=embs,
vocab=vocab
)
optimizer = torch.optim.Adam(
model.parameters(),
lr=ut.model_param['learning_rate'],
weight_decay=ut.model_param['weight_decay']
)
# model.cuda()
if torch.cuda.device_count() > 1:
print('No. of GPUs: {0}\n'.format(torch.cuda.device_count()))
model = nn.DataParallel(model)
else:
model.cuda()
print('No. of GPUs: 1\n')
loss_fn = net.criterion
print('Training for {} epochs\n'.format(ut.model_param['num_epochs']))
#only train
train_and_evaluate(
model,
data_generator_tr,
data_generator_ts,
loss_fn,
optimizer,
net.metrics,
exp_dir
)
# uncomment this out to train AND evaluate
# will take a really, really long time
# training and evaluation
# results of best model are saved to outdir/best_model.pt in this function
'''
mrn, encoded, encoded_avg, metrics_avg = train_and_evaluate(
model,
data_generator_tr,
data_generator_ts,
loss_fn,
optimizer,
net.metrics,
exp_dir
)
# save encodings
# encoded vectors (representations)
outfile = os.path.join(exp_dir, 'convae_avg_vect.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "ENCODED-AVG"])
for m, e in zip(mrn, encoded_avg):
wr.writerow([m] + list(e))
outfile = os.path.join(exp_dir, 'convae_vect.csv')
with open(outfile, 'w') as f:
wr = csv.writer(f)
wr.writerow(["MRN", "ENCODED-SUBSEQ"])
for m, evs in zip(mrn, encoded):
for e in evs:
wr.writerow([m] + e)
# metrics (loss and accuracy)
outfile = os.path.join(exp_dir, 'metrics.txt')
with open(outfile, 'w') as f:
f.write('Mean loss: %.3f\n' % metrics_avg['loss'])
f.write('Accuracy: %.3f\n' % metrics_avg['accuracy'])
# chop patient sequences into fixed subsequences of length L
# L = ut.len_padded = 32
# I think that this is here for the human-readable version of how the patient records are subset
outfile = os.path.join(exp_dir, 'cohort_ehr_subseq{0}.csv'.format(ut.len_padded))
write_ehr_subseq(data_generator_tr, outfile)
if test_set:
outfile = os.path.join(exp_dir, 'test_cohort_ehr_subseq{0}.csv'.format(ut.len_padded))
write_ehr_subseq(data_generator_ts, outfile)
'''
return
