def main(argv=None):
opt = parse_args(argv)
tasks = TCGAMeta(download=True, preload=True)
task = tasks[113]
# Setup the results dictionary
filename = "experiments/results/clinical-tasks.pkl"
try:
results = pickle.load(open(filename, "rb"), encoding='latin1')
print("Loaded Checkpointed Results")
except Exception as e:
print(e)
results = pd.DataFrame(columns=[
'task', 'acc_metric', 'model', 'graph', 'trial', 'train_size',
'time_elapsed'
])
print("Created a New Results Dictionary")
train_size = 50
trials = 3
cuda = True
exp = []
for trial in range(trials):
model = GCN(cuda=cuda,
dropout=opt.dropout,
num_layer=opt.num_layer,
channels=opt.channels,
embedding=opt.embedding,
aggregation=opt.aggregation,
lr=opt.lr,
agg_reduce=opt.agg_reduce,
seed=trial)
task._samples = task._samples - task._samples.mean(axis=0)
task._samples = task._samples / task._samples.var()
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
task._samples,
task._labels,
stratify=task._labels,
train_size=train_size,
test_size=len(task._labels) - train_size)
adj = sparse.csr_matrix(nx.to_numpy_matrix(GeneManiaGraph().nx_graph))
model.fit(X_train, y_train, adj=adj)
y_hat = []
for chunk in get_every_n(X_test, 10):
y_hat.extend(np.argmax(model.predict(chunk), axis=1).numpy())
exp.append(model.metric(y_test, y_hat))
print(exp)
report_results([{
"name": "acc_metric",
"type": "objective",
"value": np.array(exp).mean()
}])
def main(argv=None):
opt = parse_args(argv)
dataset = datasets.TCGADataset()
dataset.df = dataset.df - dataset.df.mean(axis=0)
gene_graph = GeneManiaGraph()
search_num_genes = [50, 100, 200, 300, 500, 1000, 2000, 4000, 8000, 16300]
test_size = 300
cuda = torch.cuda.is_available()
exp = []
for num_genes in search_num_genes:
start_time = time.time()
gene = "RPL4"
model = GCN(cuda=cuda,
dropout=opt.dropout,
num_layer=opt.num_layer,
channels=opt.channels,
embedding=opt.embedding,
aggregation=opt.aggregation,
lr=opt.lr,
agg_reduce=opt.agg_reduce)
dataset.labels = dataset.df[gene].where(
dataset.df[gene] > 0).notnull().astype("int")
dataset.labels = dataset.labels.values if type(
dataset.labels) == pd.Series else dataset.labels
X_train, X_test, y_train, y_test = sklearn.model_selection.train_test_split(
dataset.df,
dataset.labels,
stratify=dataset.labels,
train_size=opt.train_size,
test_size=opt.test_size,
random_state=opt.seed)
if num_genes == 16300:
neighbors = gene_graph.nx_graph
else:
neighbors = gene_graph.bfs_sample_neighbors(gene, num_genes)
X_train = X_train[list(neighbors.nodes)].copy()
X_test = X_test[list(neighbors.nodes)].copy()
X_train[gene] = 1
X_test[gene] = 1
adj = sparse.csr_matrix(nx.to_numpy_matrix(neighbors))
model.fit(X_train, y_train, adj=adj)
y_hat = model.predict(X_test)
y_hat = np.argmax(y_hat, axis=1)
auc = sklearn.metrics.roc_auc_score(y_test,
np.asarray(y_hat).flatten())
del model
exp.append(auc)
report_results([{
"name": "auc",
"type": "objective",
"value": np.array(exp).mean()
}])
lr=0.001,
num_epochs=100,
patience=30,
verbose=True,
seed=seed,
train_valid_split=0.8
)
elif model_name == 'MLP64':
model = MLP(name="MLP_lay2_chan64", cuda=cuda, dropout=True, num_layer=2, channels=64, train_valid_split=0.8, patience=30, lr=0.001)
elif model_name == 'MLP64_lr4':
model = MLP(name="MLP_lay2_chan64_lr.0.0001_nodropout", cuda=cuda, dropout=False, num_layer=2, channels=64, train_valid_split=0.8, patience=30, lr=0.0001)
try:
# print(x_train.shape, y_train.shape, adj.shape)
model.fit(x_train, y_train, adj=adj)
with torch.no_grad():
model.eval()
y_hat = model.predict(x_test)
y_hat = np.argmax(y_hat, axis=1)
# auc = sklearn.metrics.roc_auc_score(y_test, np.asarray(y_hat).flatten(), multi_class='ovo')
acc = sklearn.metrics.accuracy_score(y_test, np.asarray(y_hat).flatten())
f1 = sklearn.metrics.f1_score(y_test, np.asarray(y_hat).flatten(), average='macro')
experiment["model"] = model.name
experiment["auc"] = 0
experiment["acc"] = acc
experiment["f1"] = f1
experiment["num_genes"] = len(x_train.columns)
neighbors = list(gene_graph.first_degree(gene)[0])
neighbors = [n for n in neighbors if n in X_train.columns.values]
X_train = X_train.loc[:, neighbors].copy()
X_test = X_test.loc[:, neighbors].copy()
else:
X_train = X_train.copy()
X_test = X_test.copy()
try:
# Don't include expression of enquired gene?
# X_train[gene] = 1
# X_test[gene] = 1
with warnings.catch_warnings():
warnings.simplefilter("ignore")
model.fit(X_train, y_train, adj)
model.eval()
with torch.no_grad():
y_hat = model.predict(X_test)
auc = sklearn.metrics.roc_auc_score(y_test, np.argmax(y_hat,
axis=1))
acc = sklearn.metrics.accuracy_score(y_test,
np.argmax(y_hat, axis=1))
print("auc:", auc, " acc: ", acc)
experiment["auc"] = auc
experiment["acc"] = acc
results.append(experiment)
if auc > best_auc:
best_auc = copy.deepcopy(auc)
best_auc_model = copy.deepcopy(model)
if acc > best_acc:
patience=30,
lr=0.001)
elif model_name == 'MLP64_lr4':
model = MLP(name="MLP_lay2_chan64_lr.0.0001_nodropout",
cuda=cuda,
dropout=False,
num_layer=2,
channels=64,
train_valid_split=0.8,
patience=30,
lr=0.0001)
try:
# print(x_train.shape, y_train.shape, adj.shape)
model.fit(x_train, y_train, adj=adj, ontology_vectors=emb_vectors)
with torch.no_grad():
model.eval()
y_hat = model.predict(x_test)
y_hat = np.argmax(y_hat, axis=1)
# auc = sklearn.metrics.roc_auc_score(y_test, np.asarray(y_hat).flatten(), multi_class='ovo')
acc = sklearn.metrics.accuracy_score(
y_test,
np.asarray(y_hat).flatten())
f1 = sklearn.metrics.f1_score(y_test,
np.asarray(y_hat).flatten(),
average='macro')
experiment["model"] = model.name
experiment["auc"] = 0