def run_cnn(config):
headlines_list, labels = create_data_list(config.input_path)
bert_embedder = BertEmbedder(headlines_list, labels, config)
embeddings, labels = bert_embedder.get_sentence_embeddings()
embeddings = embeddings.unsqueeze(0).permute(1, 0, 2)
x_train, x_test, y_train, y_test = train_test_split(
embeddings, labels, test_size=config.test_size)
train_dataset = TensorDataset(x_train, y_train)
test_dataset = TensorDataset(x_test, y_test)
train_dataloader = DataLoader(train_dataset,
sampler=RandomSampler(train_dataset),
batch_size=config.batch_size)
test_dataloader = DataLoader(test_dataset,
sampler=SequentialSampler(test_dataset),
batch_size=config.batch_size)
# dataset = TensorDataset(embeddings, labels)
model = CNNModel(config, checkpoint_file=CNN_CLASSIFIER)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model.to(device)
# train_dataloader, test_dataloader = get_dataloader(dataset, config.test_size, config.batch_size)
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=config.lr)
fit_result = model.fit(train_dataloader, test_dataloader, loss_fn,
optimizer)
def main():
cuda = torch.cuda.is_available()
if cuda:
print('Device: {}'.format(torch.cuda.get_device_name(0)))
json_params_file = sys.argv[1]
with open(json_params_file, "r") as json_file:
params = json.load(json_file)
print(params)
if HYPERPARAM_MODE == params['mode']:
hyperparam_log_file = open('cnn_hyperparam_optimization.txt', 'w')
hyperparam_log_file.write(
'batch_size,dims,lr,best_epoch,best_f1,last_f1\n')
best_model_params = None
best_model_f1 = -1
base_params = params.copy()
del base_params['mode']
param_grid = ParameterGrid(base_params)
for search_params in param_grid:
print('Fitting model with params:', search_params)
siamese_model = CNNModel(search_params, cuda)
siamese_model.fit()
hyperparam_log_file.write('{},{},{},{},{},{}\n'.format(
search_params['batch_size'], search_params['dims'],
search_params['lr'], siamese_model.best_epoch,
siamese_model.best_f1, siamese_model.current_f1))
if siamese_model.best_f1 > best_model_f1:
best_model_f1 = siamese_model.best_f1
best_model_params = search_params
print('Best model params! F1:{}'.format(best_model_f1),
best_model_params)
hyperparam_log_file.close()
elif params['mode'] == 'EVAL_MODE':
siamese_model = CNNModel(params, cuda)
siamese_model.eval()
else:
siamese_model = CNNModel(params, cuda)
siamese_model.fit()
def cnn_cross_validation_model(config, n_splits):
scores = {}
kf = KFold(n_splits=n_splits)
params_grid = {'l_r': [3e-5, 5e-5, 1e-4, 1e-3], 'batch_size': [32, 64]}
headlines_list, labels = create_data_list(config.input_path)
bert_embedder = BertEmbedder(headlines_list, labels, config)
print("started embedding")
embeddings, labels = bert_embedder.get_sentence_embeddings()
print("finished embedding")
embeddings = embeddings.unsqueeze(0).permute(1, 0, 2)
x_train, x_test, y_train, y_test = train_test_split(
embeddings, labels, test_size=config.test_size)
max_acc = -numpy.inf
for index, params in enumerate(list(
ParameterGrid(param_grid=params_grid))):
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(
"-----------------------------------------------------------------"
)
print(f"current params are {params}")
print(f" index is: {index}")
curr_acc = 0
for train_idx, valid_idx in kf.split(x_train):
cnn = CNNModel(config, checkpoint_file=CNN_CLASSIFIER)
cnn.to(device)
cnn.batch_size = params['batch_size']
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(cnn.parameters(), lr=params["l_r"])
train_x, train_y = x_train[train_idx], y_train[train_idx]
train_set = TensorDataset(train_x, train_y)
valid_x, valid_y = x_train[valid_idx], y_train[valid_idx]
valid_set = TensorDataset(valid_x, valid_y)
train_dataloader = DataLoader(train_set,
sampler=RandomSampler(train_set),
batch_size=cnn.batch_size)
valid_dataloader = DataLoader(valid_set,
sampler=SequentialSampler(valid_set),
batch_size=cnn.batch_size)
fit_result = cnn.fit(train_dataloader, valid_dataloader, loss_fn,
optimizer)
curr_acc += max(fit_result.test_acc)
mean = curr_acc / n_splits
scores[f"{params}"] = mean
if mean > max_acc:
max_acc = mean
best_params = params
scores[index] = {"params": params, "acc": mean}
print(f"all scores: {scores}")
print(f"max accuracy achieved is {max_acc}")
print(f"best params are {best_params}")
train_dataset = TensorDataset(x_train, y_train)
test_dataset = TensorDataset(x_test, y_test)
train_dataloader = DataLoader(train_dataset,
sampler=RandomSampler(train_dataset),
batch_size=best_params['batch_size'])
test_dataloader = DataLoader(test_dataset,
sampler=SequentialSampler(test_dataset),
batch_size=best_params['batch_size'])
best_model = CNNModel(config, checkpoint_file=CNN_CLASSIFIER)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
best_model.to(device)
best_model.batch_size = best_params['batch_size']
loss_fn = nn.CrossEntropyLoss()
optimizer = optim.Adam(best_model.parameters(), lr=best_params["l_r"])
fit_result = best_model.fit(train_dataloader, test_dataloader, loss_fn,
optimizer)