def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print("Getting the dataset...")
dataset = datasets.get_dataset(opt, exp_dir)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir, opt, dataset.dataset.input_size(),
dataset.dataset.additional_info())
# Training optimizer and stuff
criterion = torch.nn.MSELoss()
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
# The training.
print("Start training.")
#monitoring and predictions
predictions = np.zeros(
(dataset.dataset.nb_patient, dataset.dataset.nb_gene))
indices_patients = np.arange(dataset.dataset.nb_patient)
indices_genes = np.arange(dataset.dataset.nb_gene)
xdata = np.transpose([
np.tile(indices_genes, len(indices_patients)),
np.repeat(indices_patients, len(indices_genes))
])
progress_bar_modulo = len(dataset) / 100
monitoring_dic = {}
monitoring_dic['train_loss'] = []
for t in range(epoch, opt.epoch):
start_timer = time.time()
thisepoch_trainloss = []
with tqdm(dataset, unit="batch") as tepoch:
for mini in tepoch:
tepoch.set_description(f"Epoch {t}")
inputs, targets = mini[0], mini[1]
inputs = Variable(inputs, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
# Forward pass: Compute predicted y by passing x to the model
y_pred = my_model(inputs).float()
y_pred = y_pred.squeeze()
targets = torch.reshape(targets, (targets.shape[0], ))
# Compute and print loss
loss = criterion(y_pred, targets)
to_list = loss.cpu().data.numpy().reshape((1, ))[0]
thisepoch_trainloss.append(to_list)
tepoch.set_postfix(loss=loss.item())
np.save(os.path.join(exp_dir, 'pixel_epoch_{}'.format(t)),
my_model.emb_1.weight.cpu().data.numpy())
np.save(os.path.join(exp_dir, 'digit_epoch_{}'.format(t)),
my_model.emb_2.weight.cpu().data.numpy())
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
monitoring_dic['train_loss'].append(np.mean(thisepoch_trainloss))
np.save(f'{exp_dir}/train_loss.npy', monitoring_dic['train_loss'])
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print "Getting the dataset..."
dataset = datasets.get_dataset(opt)
# Creating a model
print "Getting the model..."
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir, opt, dataset.dataset.input_size())
# Training optimizer and stuff
criterion = torch.nn.MSELoss()
if not opt.cpu:
print "Putting the model on gpu..."
my_model.cuda()
# The training.
print "Start training."
progress_bar_modulo = len(dataset) / 100
for t in range(epoch, opt.epoch):
start_timer = time.time()
if opt.save_error:
outfname_g = '_'.join(['gene_epoch', str(t), 'prediction.npy'])
outfname_g = os.path.join(exp_dir, outfname_g)
outfname_t = '_'.join(['tissue_epoch', str(t), 'prediction.npy'])
outfname_t = os.path.join(exp_dir, outfname_t)
train_trace = np.zeros(
(dataset.dataset.nb_gene, dataset.dataset.nb_patient))
for no_b, mini in enumerate(dataset):
inputs, targets = mini[0], mini[1]
inputs = Variable(inputs, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
if not opt.cpu:
inputs = inputs.cuda()
targets = targets.cuda()
# Forward pass: Compute predicted y by passing x to the model
y_pred = my_model(inputs).float()
if opt.save_error:
# Log the predicted values per sample and per gene (S.L. validation)
batch_inputs = mini[0].numpy()
predicted_values = y_pred.data.cpu().numpy()
train_trace[batch_inputs[:, 0],
batch_inputs[:, 1]] = predicted_values[:, 0]
# Compute and print loss
loss = criterion(y_pred, targets)
# TODO: the logging here.
if no_b % progress_bar_modulo == 0:
print "Doing epoch {}, examples {}/{}. Loss: {}".format(
t, no_b, len(dataset), loss.data[0])
# Saving the emb
monitoring.save_everything(exp_dir, t, my_model,
dataset.dataset)
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
if opt.save_error:
monitoring.dump_error_by_tissue(train_trace, dataset.dataset.data,
outfname_t, exp_dir,
dataset.dataset.data_type,
dataset.dataset.nb_patient)
monitoring.dump_error_by_gene(train_trace, dataset.dataset.data,
outfname_g, exp_dir)
print "Saving the model..."
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
print "Done!"
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
if opt.cache == 0:
opt.cache = random.getrandbits(128)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
if opt.model == 'RNN':
print('This model is deprecated - please use TCRonly from now on')
# creating the dataset
print("Getting the dataset...")
if not 'cached_dataset' in os.listdir('.'):
os.mkdir('cached_dataset')
opt.dataset = 'binary_thome'
tenth = opt.tenth
#dataset = datasets.get_dataset(opt,exp_dir,tenth=opt.tenth)
dataset = old_datasets.get_dataset(opt, exp_dir, test=True)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir,
opt,
dataset.dataset.input_size(),
)
criterion = torch.nn.MSELoss()
# Training optimizer and stuff
if opt.loss == 'NLL' or opt.model == 'allseq_bin':
criterion = torch.nn.NLLLoss()
criterion = torch.nn.BCELoss()
if not 'tcr_embs' in os.listdir(exp_dir):
if opt.model == 'TCRonly':
os.mkdir(f'{exp_dir}/tcr_embs/')
elif opt.model == 'allseq' or opt.model == 'allseq_bin':
os.mkdir(f'{exp_dir}/tcr_embs/')
os.mkdir(f'{exp_dir}/hla_embs/')
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
loss_dict = {}
loss_dict['train_losses'] = []
def estimate_batch_accuracy(y, yhat):
return np.sum([i == j for i, j in zip(y, yhat)]) / y.shape[0]
if opt.model == 'allseq' or opt.model == 'allseq_bin':
valid_list = np.load(
'/u/trofimov/Emerson/processed_data/valid_list.npy')
loss_dict['valid_losses'] = []
# The training.
print("Getting the likelihood")
os.mkdir(f'{exp_dir}/Thome_tenth{tenth}_preds_100/')
#monitoring and predictions
for t in range(1):
loss_dict = monitoring.update_loss_dict(loss_dict, start=True)
if opt.model == 'allseq_bin':
good = 0
for no_b, mini in enumerate(dataset):
if opt.model == 'TCRonly':
y_pred, my_model, targets = training.TCRonly_batch(
mini, opt, my_model)
np.save(f'{exp_dir}/preds_100/likelihood_batch{no_b}.npy',
y_pred.data.cpu().numpy())
if no_b % 5 == 0:
print(f"Doing epoch{t},examples{no_b}/{len(dataset)}")
# Saving the emb
elif opt.model == 'allseq':
inputs_k, inputs_h1, inputs_h2, inputs_h3, inputs_h4, targets = training.allseq_batch(
mini, opt)
y_pred = my_model(inputs_k, inputs_h1, inputs_h2, inputs_h3,
inputs_h4).float()
np.save(f'{exp_dir}/preds_100/likelihood_batch{no_b}.npy',
y_pred.data.cpu().numpy())
batch_number = dataset.dataset.data[no_b]
bn = batch_number[0]
np.save(f'{exp_dir}/preds_100/likelihood_batch{bn}.npy',
y_pred.data.cpu().numpy())
if no_b % 5 == 0:
print(f"Doing epoch {t},examples{no_b}/{len(dataset)}")
elif opt.model == 'allseq_bin':
inputs_k, inputs_h1, inputs_h2, inputs_h3, inputs_h4, targets = training.binallseq_batch(
mini, opt)
y_pred = my_model(inputs_k, inputs_h1, inputs_h2, inputs_h3,
inputs_h4).float()
batch_number = dataset.dataset.data[no_b]
bn = batch_number[0]
np.save(
f'{exp_dir}/Thome_tenth{tenth}_preds_100/likelihood_batch{bn}.npy',
y_pred.data.cpu().numpy())
if no_b % 5 == 0:
print(f"Doing epoch {t},examples{no_b}/{len(dataset)}")
def main(argv=None):
"""Main."""
opt = parse_args(argv)
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print("Getting the dataset...")
dataset = datasets.get_dataset(opt, exp_dir)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(exp_dir, opt)
# Training optimizer and stuff
criterion = nn.NLLLoss()
criterion = nn.CrossEntropyLoss()
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
# The training.
print("Start training.")
monitoring_dic = {}
monitoring_dic['train_loss'] = []
monitoring_dic['valid_loss'] = []
monitoring_dic['train_accuracy'] = []
monitoring_dic['valid_accuracy'] = []
max_accuracy = 0
min_loss = 10000
patience = 5
for t in range(epoch, opt.epoch):
if patience == 0:
break
thisepoch_trainloss = []
thisepoch_trainaccuracy = []
with tqdm(dataset, unit="batch") as tepoch:
for mini in tepoch:
tepoch.set_description(f"Epoch {t}")
inputs, targets = mini[0], mini[1]
inputs = Variable(inputs, requires_grad=False).long()
targets = Variable(targets, requires_grad=False).long()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
optimizer.zero_grad()
y_pred = my_model(inputs).float()
#y_pred = torch.reshape(y_pred, (y_pred.shape[0], ))
#targets = torch.reshape(targets, (targets.shape[1], 1))
loss = criterion(y_pred, targets)
to_list = loss.cpu().data.numpy().reshape((1, ))[0]
thisepoch_trainloss.append(to_list)
loss.backward()
optimizer.step()
accuracy = np.sum(
np.argmax(y_pred.cpu().data.numpy(), axis=1) ==
targets.cpu().data.numpy()) / targets.shape[0]
thisepoch_trainaccuracy.append(accuracy)
tepoch.set_postfix(loss=loss.item(), accuracy=accuracy)
inputs = dataset.dataset.valid_inputs
inputs = torch.FloatTensor(inputs)
inputs = Variable(inputs, requires_grad=False).long()
targets = dataset.dataset.valid_targets
targets = torch.FloatTensor(targets)
targets = Variable(targets, requires_grad=False).long()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
vlosses = []
accuracies = []
split = np.sum(targets.cpu().data.numpy()) / targets.shape[0]
for i in tqdm(range(0, inputs.shape[0], 100)):
with torch.no_grad():
y_pred = my_model(inputs[i:i + 100])
vloss = criterion(y_pred, targets[i:i + 100])
accuracy = np.sum(
np.argmax(y_pred.cpu().data.numpy(), axis=1) ==
targets[i:i + 100].cpu().data.numpy()) / y_pred.shape[0]
vloss = vloss.cpu().data.numpy().reshape((1, ))[0]
vlosses.append(vloss)
accuracies.append(accuracy)
vloss = np.mean(vlosses)
accuracy = np.mean(accuracies)
if vloss < min_loss:
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
print(
f'*** new min loss*** Validation loss: epoch {t}, loss: {vloss}, accuracy: {accuracy}, split {split}'
)
min_loss = vloss
patience = 5
else:
print(
f'Validation loss: epoch {t}, loss: {vloss}, accuracy:{accuracy}, split {split}'
)
patience -= 1
monitoring_dic['valid_loss'].append(vloss)
monitoring_dic['valid_accuracy'].append(accuracy)
monitoring_dic['train_loss'].append(np.mean(thisepoch_trainloss))
monitoring_dic['train_accuracy'].append(
np.mean(thisepoch_trainaccuracy))
trainacc = np.mean(thisepoch_trainaccuracy)
print(f'Training accuracy: {trainacc}')
np.save(f'{exp_dir}/train_loss.npy', monitoring_dic['train_loss'])
np.save(f'{exp_dir}/valid_loss.npy', monitoring_dic['valid_loss'])
np.save(f'{exp_dir}/train_accuracy.npy',
monitoring_dic['train_accuracy'])
np.save(f'{exp_dir}/valid_accuracy.npy',
monitoring_dic['valid_accuracy'])
print('Done training!')
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
if opt.cache==0:
opt.cache = random.getrandbits(128)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
if opt.model == 'RNN':
print ('This model is deprecated - please use TCRonly from now on')
# creating the dataset
print ("Getting the dataset...")
if not 'cached_dataset' in os.listdir('.'):
os.mkdir('cached_dataset')
dataset = datasets.get_dataset(opt,exp_dir)
# Creating a model
print ("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(exp_dir, opt, dataset.dataset.input_size(), )
criterion = torch.nn.MSELoss()
# Training optimizer and stuff
if opt.loss == 'NLL' or opt.model=='allseq_bin':
criterion = torch.nn.NLLLoss()
criterion = torch.nn.BCELoss()
if not 'tcr_embs' in os.listdir(exp_dir):
if opt.model == 'TCRonly':
os.mkdir(f'{exp_dir}/tcr_embs/')
elif opt.model == 'allseq' or opt.model == 'allseq_bin':
os.mkdir(f'{exp_dir}/tcr_embs/')
os.mkdir(f'{exp_dir}/hla_embs/')
os.mkdir(f'{exp_dir}/predictions')
if not opt.cpu:
print ("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
loss_dict = {}
loss_dict['train_losses'] = []
def estimate_batch_accuracy(y,yhat):
return np.sum([i==j for i,j in zip(y,yhat)])/y.shape[0]
if opt.model == 'allseq' or opt.model == 'allseq_bin':
valid_list = np.load('/u/trofimov/Emerson/processed_data/valid_list.npy')
loss_dict['valid_losses'] = []
# The training.
print ("Start training.")
#monitoring and predictions
for t in range(epoch, opt.epoch):
loss_dict = monitoring.update_loss_dict(loss_dict,start = True)
if opt.model == 'allseq_bin':
good = 0
for no_b, mini in enumerate(dataset):
if opt.model == 'TCRonly':
y_pred, my_model, targets = training.TCRonly_batch(mini,opt,my_model)
loss = criterion(y_pred, targets)
loss_save = loss.data.cpu().numpy().reshape(1,)[0]
loss_dict['train_losses_epoch'].append(loss_save)
if no_b % 5 == 0:
print (f"Doing epoch{t},examples{no_b}/{len(dataset)}.Loss:{loss_save}")
# Saving the emb
np.save(os.path.join(exp_dir, 'pixel_epoch_{}'.format(t)),my_model.emb_1.weight.cpu().data.numpy())
optimizer.zero_grad()
loss.backward()
optimizer.step()
kmerembs = my_model.get_embeddings(inputs_k, inputs_s)[0].squeeze()
np.save(f'{exp_dir}/kmer_embs/kmer_embs_batch_{no_b}',kmerembs.cpu().data.numpy())
elif opt.model == 'allseq':
inputs_k,inputs_h1, inputs_h2, inputs_h3, inputs_h4, targets = training.allseq_batch(mini,opt)
y_pred = my_model(inputs_k,inputs_h1, inputs_h2, inputs_h3,
inputs_h4).float()
loss = criterion(y_pred, targets)
loss_save = loss.data.cpu().numpy().reshape(1,)[0]
if no_b in valid_list:
loss_dict['valid_losses_epoch'].append(loss_save)
print (f"Validation error {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}")
elif no_b % 5 == 0:
loss_dict['train_losses_epoch'].append(loss_save)
print (f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}")
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_number = dataset.dataset.data[no_b]
kmerembs = my_model.get_embeddings(inputs_k, inputs_h1,
inputs_h2, inputs_h3,
inputs_h4)
kmerembs1 = kmerembs[0].squeeze()
bn = batch_number[0]
np.save(f'{exp_dir}/tcr_embs/tcr_embs_batch_{bn}',kmerembs1.cpu().data.numpy())
for i in range(4):
kmerembs1 = kmerembs[i+1].squeeze()
kmerembs1 = kmerembs1[0]
np.save(f'{exp_dir}/hla_embs/hla_embs_batch_{bn}_h{i+1}',kmerembs1.cpu().data.numpy())
kmerembs1 = my_model.hla_representation
kmerembs1 = kmerembs1[0].squeeze()
np.save(f'{exp_dir}/hla_embs/ppl_embs_batch_{bn}',kmerembs1.cpu().data.numpy())
elif opt.model == 'allseq_bin':
inputs_k, inputs_h1, inputs_h2, inputs_h3, inputs_h4, targets = training.binallseq_batch(mini,opt)
y_pred = my_model(inputs_k,inputs_h1, inputs_h2, inputs_h3,
inputs_h4).float()
loss = criterion(y_pred, targets)
#nb_pos = (np.sum(np.argmax(y_pred.cpu().detach().numpy(),axis=1))/y_pred.shape[0])
#b_accuracy = (estimate_batch_accuracy(np.argmax(y_pred.cpu().detach().numpy(),axis=1),
# np.argmax(targets.cpu().detach().numpy(),axis=1)))
#if no_b % 10 == 0:
# print (f'predicted proportion: {nb_pos} - accuracy: {b_accuracy}')
#if b_accuracy>0.75:
# good+=1
loss_save = loss.data.cpu().numpy().reshape(1,)[0]
if no_b in valid_list:
loss_dict['valid_losses_epoch'].append(loss_save)
print (f"Validation error {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}")
else:
loss_dict['train_losses_epoch'].append(loss_save)
if no_b % 50 == 0:
print (f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}")
optimizer.zero_grad()
loss.backward()
optimizer.step()
batch_number = dataset.dataset.data[no_b]
bn = batch_number[0]
for newpass in range(100):
y_pred = my_model(inputs_k,inputs_h1, inputs_h2, inputs_h3,
inputs_h4).float()
loss = criterion(y_pred, targets)
loss_newpass = loss.data.cpu().numpy().reshape(1,)[0]
print (f'batch {bn} pass {newpass} loss: {loss_newpass}')
optimizer.zero_grad()
loss.backward()
optimizer.step()
preds_targets = np.hstack((y_pred.cpu().data.numpy(),
targets.cpu().data.numpy()))
np.save(f'{exp_dir}/predictions/batch_{bn}_{no_b}.npy',preds_targets)
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
batch_number = dataset.dataset.data[no_b]
kmerembs = my_model.get_embeddings(inputs_k, inputs_h1,
inputs_h2, inputs_h3,
inputs_h4)
kmerembs1 = kmerembs[0].squeeze()
bn = batch_number[0]
true_size = int(kmerembs1.shape[0]/2)
np.save(f'{exp_dir}/tcr_embs/tcr_embs_batch_{bn}',kmerembs1.cpu().data.numpy()[:true_size])
for i in range(4):
kmerembs1 = kmerembs[i+1].squeeze()
kmerembs1 = kmerembs1[0]
np.save(f'{exp_dir}/hla_embs/hla_embs_batch_{bn}_h{i+1}',kmerembs1.cpu().data.numpy()[:true_size])
kmerembs1 = my_model.hla_representation
kmerembs1 = kmerembs1[0].squeeze()
np.save(f'{exp_dir}/hla_embs/ppl_embs_batch_{bn}',kmerembs1.cpu().data.numpy()[:true_size])
print ("Saving the model...")
if opt.model=='allseq_bin' or opt.model=='allseq':
validation_scores = loss_dict['valid_losses_epoch']
else:
validation_scores = None
#print (f'number correct examples: {good}')
if opt.model == 'allseq' or opt.model=='allseq_bin':
toprint = loss_dict['valid_losses_epoch']
print (f'validation loss matrix: {toprint}')
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
monitoring.update_loss_dict(loss_dict, start=False)
monitoring.save_loss(loss_dict,exp_dir)
if t % opt.plot_frequency==0:
monitoring.plot_training_curve(exp_dir, loss_dict)
print ('Finished training! Starting evaluations')
tcr_rep_dir = f'{exp_dir}/tcr_embs'
patient_to_index = f'data/hla_for_model_eval/pt_names.csv'
original_data_dir = f'/u/trofimov/Emerson/original'
validation_scores = np.load(f'{exp_dir}/validation_loss.npy')
nb_patients = 15
validation_scores = np.load(f'{exp_dir}/validation_loss.npy')
output = evaluations.evaluate_model(opt, my_model ,exp_dir, tcr_rep_dir, patient_to_index,
original_data_dir, validation_scores, nb_patients,
train_on_index=0)
with open(f'{exp_dir}/evaluation_results.json', 'w') as json_file:
json.dump(output, json_file)
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print("Getting the dataset...")
dataset = datasets.get_dataset(opt, exp_dir)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir, opt, dataset.dataset.input_size(),
dataset.dataset.additional_info())
# Training optimizer and stuff
criterion = torch.nn.MSELoss()
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
# The training.
print("Start training.")
#monitoring and predictions
predictions = np.zeros(
(dataset.dataset.nb_patient, dataset.dataset.nb_gene))
indices_patients = np.arange(dataset.dataset.nb_patient)
indices_genes = np.arange(dataset.dataset.nb_gene)
xdata = np.transpose([
np.tile(indices_genes, len(indices_patients)),
np.repeat(indices_patients, len(indices_genes))
])
progress_bar_modulo = len(dataset) / 100
for t in range(epoch, opt.epoch):
start_timer = time.time()
if opt.save_error:
outfname_g = '_'.join(['gene_epoch', str(t), 'prediction.npy'])
outfname_g = os.path.join(exp_dir, outfname_g)
outfname_t = '_'.join(['tissue_epoch', str(t), 'prediction.npy'])
outfname_t = os.path.join(exp_dir, outfname_t)
train_trace = np.zeros(
(dataset.dataset.nb_gene, dataset.dataset.nb_patient))
### making predictions:
nb_proteins = my_model.emb_3.weight.cpu().data.numpy().shape[0]
nb_patients = my_model.emb_2.weight.cpu().data.numpy().shape[0]
predictions_protein = np.zeros((nb_patients, nb_proteins))
patient_embs = my_model.emb_2.weight.cpu().data.numpy()
for patient in np.arange(nb_patients):
new = my_model.generate_datapoint_protein(patient_embs[patient, :],
gpu=2)
new = new.cpu().data.numpy()
predictions_protein[patient, :] = new[:, 0]
np.save(f'predictions_protein_{epoch}.npy', predictions_protein)
for no_b, mini in enumerate(dataset):
inputs, targets, inputs2, targets2 = mini[0], mini[1], mini[
2], mini[3]
inputs = Variable(inputs, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
inputs2 = Variable(inputs2, requires_grad=False).float()
targets2 = Variable(targets2, requires_grad=False).float()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
inputs2 = inputs2.cuda(opt.gpu_selection)
targets2 = targets2.cuda(opt.gpu_selection)
# Forward pass: Compute predicted y by passing x to the model
y_pred = my_model([inputs, inputs2])
#if opt.save_error:
# Log the predicted values per sample and per gene (S.L. validation)
# batch_inputs = mini[0].numpy()
# predicted_values = y_pred.data.cpu().numpy()
# train_trace[batch_inputs[:,0],batch_inputs[:,1]] = predicted_values[:,0]
targets = torch.reshape(targets, (targets.shape[0], 1))
targets2 = torch.reshape(targets2, (targets2.shape[0], 1))
# Compute and print loss
loss1 = criterion(y_pred[0], targets)
loss2 = criterion(y_pred[1], targets2)
loss = loss1 + loss2
if no_b % 5 == 0:
print(
f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss.data.cpu().numpy().reshape(1,)[0]}"
)
# Saving the emb
np.save(os.path.join(exp_dir, 'pixel_epoch_{}'.format(t)),
my_model.emb_1.weight.cpu().data.numpy())
np.save(os.path.join(exp_dir, 'digit_epoch_{}'.format(t)),
my_model.emb_2.weight.cpu().data.numpy())
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
#my_model.generate_datapoint([0,0], opt.gpu_selection)
#monitoring.save_predictions(exp_dir, predictions)
# for i in range(0,xdata.shape[0],1000):
# #import pdb; pdb.set_trace()
# inputs = torch.FloatTensor(xdata[i:i+1000,:])
# inputs = Variable(inputs, requires_grad=False).float()
# if not opt.cpu:
# inputs = inputs.cuda(opt.gpu_selection)
# y_pred = my_model(inputs).float()
# predictions[inputs.data.cpu().numpy()[:,1].astype('int32'),inputs.data.cpu().numpy()[:,0].astype('int32')] = y_pred.data.cpu().numpy()[:,0]
# monitoring.dump_error_by_tissue(train_trace, dataset.dataset.data, outfname_t, exp_dir, dataset.dataset.data_type, dataset.dataset.nb_patient)
# monitoring.dump_error_by_gene(train_trace, dataset.dataset.data, outfname_g, exp_dir)
#print ("Saving the model...")
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print("Getting the dataset...")
dataset = datasets.get_dataset(opt, exp_dir)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir,
opt,
dataset.dataset.input_size(),
)
criterion = torch.nn.MSELoss()
# Training optimizer and stuff
if opt.loss == 'NLL':
criterion = torch.nn.NLLLoss()
os.mkdir(f'{exp_dir}/kmer_embs/')
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
# The training.
print("Start training.")
#monitoring and predictions
for t in range(epoch, opt.epoch):
start_timer = time.time()
for no_b, mini in enumerate(dataset):
inputs_s, inputs_k, targets = mini[0], mini[1], mini[2]
inputs_s = Variable(inputs_s, requires_grad=False).float()
inputs_k = Variable(inputs_k, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
if not opt.cpu:
inputs_s = inputs_s.cuda(opt.gpu_selection)
inputs_k = inputs_k.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
# Forward pass: Compute predicted y by passing x to the model
inputs_k = inputs_k.view(-1, 31, 4)
y_pred = my_model(inputs_k, inputs_s).float()
#import pdb; pdb.set_trace()
#targets = torch.reshape(targets,(targets.shape[0],1))
# Compute and print loss
loss = criterion(y_pred, targets)
if no_b % 5 == 0:
print(
f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss.data.cpu().numpy().reshape(1,)[0]}"
)
# Saving the emb
np.save(os.path.join(exp_dir, 'pixel_epoch_{}'.format(t)),
my_model.emb_1.weight.cpu().data.numpy())
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
loss.backward()
optimizer.step()
kmerembs_batch = my_model.get_embeddings(inputs_k, inputs_s)[0]
kmerembs = kmerembs_batch[:int(kmerembs_batch.shape[0] /
opt.nb_patient)]
np.save(f'{exp_dir}/kmer_embs/kmer_embs_batch_{no_b}',
kmerembs.cpu().data.numpy())
#print ("Saving the model...")
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
def main(argv=None):
opt = parse_args(argv)
# TODO: set the seed
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
if opt.cache == 0:
opt.cache = random.getrandbits(128)
exp_dir = opt.load_folder
if exp_dir is None: # we create a new folder if we don't load.
exp_dir = monitoring.create_experiment_folder(opt)
# creating the dataset
print("Getting the dataset...")
if not 'cached_dataset' in os.listdir('.'):
os.mkdir('cached_dataset')
dataset = datasets.get_dataset(opt, exp_dir)
# Creating a model
print("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(
exp_dir,
opt,
dataset.dataset.input_size(),
)
# Training optimizer and stuff
criterion = torch.nn.BCELoss()
if not opt.cpu:
print("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
loss_dict = {}
loss_dict['train_losses'] = []
valid_list = np.load('/u/trofimov/Emerson/processed_data/valid_list.npy')
loss_dict['valid_losses'] = []
# The training.
print("Start training.")
#monitoring and predictions
for t in range(epoch, opt.epoch):
loss_dict = monitoring.update_loss_dict(loss_dict, start=True)
for no_b, mini in enumerate(dataset):
inputs_x, targets = mini[0], mini[1]
inputs_x = Variable(inputs_x, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
if not opt.cpu:
inputs_x = inputs_x.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
y_pred = my_model(inputs_x)
loss = criterion(y_pred, targets)
loss_save = loss.data.cpu().numpy().reshape(1, )[0]
if no_b in valid_list:
loss_dict['valid_losses_epoch'].append(loss_save)
print(
f"Validation error {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}"
)
else:
loss_dict['train_losses_epoch'].append(loss_save)
print(
f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss_save}"
)
optimizer.zero_grad()
loss.backward()
optimizer.step()
validation_scores = loss_dict['valid_losses_epoch']
monitoring.save_checkpoint(my_model, optimizer, t, opt, exp_dir)
monitoring.update_loss_dict(loss_dict, start=False)
monitoring.save_loss(loss_dict, exp_dir)
if t % opt.plot_frequency == 0:
monitoring.plot_training_curve(exp_dir, loss_dict)
def impute(argv=None):
opt = parse_args(argv)
seed = opt.seed
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.manual_seed(seed)
exp_dir = opt.load_folder
new_save_dir = monitoring.create_experiment_folder(opt)
if exp_dir is None:
print ("Experiment doesn't exist!")
else:
# creating the dataset
print ("Getting the dataset...")
dataset = datasets.get_dataset(opt,exp_dir)
old_nb_patients = dataset.dataset.nb_patient
old_data_file = opt.data_file
new_data_file = opt.new_data_file
imputation_list = opt.imputation_list
nb_shuffles = opt.nb_shuffles
# Creating a model
print ("Getting the model...")
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(exp_dir,opt,dataset.dataset.input_size(),impute=True)
old_model_size = dataset.dataset.input_size()
### Making sure updates are only on the patient embedding layer
#my_model.freeze_all()
#optimizer = torch.optim.RMSprop(my_model.emb_2.parameters(), lr=opt.lr, weight_decay=opt.weight_decay)
### Replacing the first embeddings as the new number of patients to predict
my_model.emb_2.weight[:dataset.dataset.nb_patient,:] = Variable(torch.FloatTensor(np.zeros((dataset.dataset.nb_patient,2))), requires_grad=False).float()
# Training optimizer and stuff
criterion = torch.nn.MSELoss()
if not opt.cpu:
print ("Putting the model on gpu...")
my_model.cuda(opt.gpu_selection)
# The training.
print ("Start training.")
for nb_genes in imputation_list:
print (f'Imputation with {nb_genes} genes given...')
for shuffle in range(nb_shuffles):
opt.data_file = new_data_file
print ("Re-getting the dataset...")
dataset = datasets.get_dataset(opt,exp_dir, masked = nb_genes)
new_embs = np.zeros((dataset.dataset.nb_patient,2))
#monitoring and predictions
predictions=np.zeros((dataset.dataset.nb_patient,my_model.emb_1.weight.shape[0]))
indices_patients = np.arange(predictions.shape[0])
indices_genes = np.arange(predictions.shape[1])
xdata = np.transpose([np.tile(indices_genes, len(indices_patients)),
np.repeat(indices_patients, len(indices_genes))])
progress_bar_modulo = len(dataset)/100
for t in range(25):
my_model, optimizer, epoch, opt = monitoring.load_checkpoint(exp_dir,opt,old_model_size,impute=True)
temp_embs = my_model.emb_2.weight.cpu().data.numpy()
temp_embs[:dataset.dataset.nb_patient,:] = new_embs
my_model.emb_2.weight.data = my_model.emb_2.weight.data.copy_(torch.from_numpy(temp_embs))
#my_model.emb_2.weight[:dataset.dataset.nb_patient,:] = Variable(torch.FloatTensor(new_embs),requires_grad=False).float()
start_timer = time.time()
#if opt.save_error:
#outfname_g = f'shuffle_{shuffle}_{nb_genes}_genes_epoch_{t}_prediction_genes.npy'
#outfname_g = os.path.join(new_exp_dir,outfname_g)
#outfname_t = f'shuffle_{shuffle}_{nb_genes}_genes_epoch_{t}_prediction_tissue.npy'
#outfname_t = os.path.join(new_exp_dir,outfname_t)
#train_trace = np.zeros((dataset.dataset.nb_gene, dataset.dataset.nb_patient))
for no_b, mini in enumerate(dataset):
inputs, targets = mini[0], mini[1]
inputs = Variable(inputs, requires_grad=False).float()
targets = Variable(targets, requires_grad=False).float()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
targets = targets.cuda(opt.gpu_selection)
# Forward pass: Compute predicted y by passing x to the model
y_pred = my_model(inputs).float()
#if opt.save_error:
## Log the predicted values per sample and per gene (S.L. validation)
#batch_inputs = mini[0].numpy()
#predicted_values = y_pred.data.cpu().numpy()
#train_trace[batch_inputs[:,0],batch_inputs[:,1]] = predicted_values[:,0]
#import pdb; pdb.set_trace()
targets = torch.reshape(targets,(targets.shape[0],1))
# Compute and print loss
loss = criterion(y_pred, targets)
if no_b % 5 == 0:
print (f"Doing epoch {t},examples{no_b}/{len(dataset)}.Loss:{loss.data.cpu().numpy().reshape(1,)[0]}")
# Saving the emb
np.save(os.path.join(new_save_dir, 'pixel_epoch_{}'.format(t)),my_model.emb_1.weight.cpu().data.numpy() )
np.save(os.path.join(new_save_dir,'digit_epoch_{}'.format(t)),my_model.emb_2.weight.cpu().data.numpy())
# Zero gradients, perform a backward pass, and update the weights.
optimizer.zero_grad()
#import pdb; pdb.set_trace()
loss.backward()
optimizer.step()
new_embs = my_model.emb_2.weight.cpu().data.numpy()[:dataset.dataset.nb_patient,:]
#my_model.generate_datapoint([0,0], opt.gpu_selection)
for i in range(0,xdata.shape[0],1000):
#import pdb; pdb.set_trace()
inputs = torch.FloatTensor(xdata[i:i+1000,:])
inputs = Variable(inputs, requires_grad=False).float()
if not opt.cpu:
inputs = inputs.cuda(opt.gpu_selection)
y_pred = my_model(inputs).float()
predictions[inputs.data.cpu().numpy()[:,1].astype('int32'),inputs.data.cpu().numpy()[:,0].astype('int32')] = y_pred.data.cpu().numpy()[:,0]
outfname_pred = f'shuffle_{shuffle}_{nb_genes}_genes_prediction.npy'
outfname_pred = os.path.join(new_save_dir,outfname_pred)
np.save(outfname_pred, predictions)