def ensemble(state, X_test, y_test, g):
mod1 = Model().to(state['device'])
mod2 = Model().to(state['device'])
mod3 = Model().to(state['device'])
mod4 = Model().to(state['device'])
mod = Model().to(state['device'])
mod1.load_state_dict(torch.load(state['path1'])['model_state_dict'])
mod2.load_state_dict(torch.load(state['path2'])['model_state_dict'])
mod3.load_state_dict(torch.load(state['path3'])['model_state_dict'])
mod4.load_state_dict(torch.load(state['path4'])['model_state_dict'])
for p, p1, p2, p3, p4 in zip(mod.parameters(), mod1.parameters(),
mod2.parameters(), mod3.parameters(),
mod4.parameters()):
p.data.copy_(
p1.data.mul(0.25).add(p2.data.mul(0.25)).add(
p3.data.mul(0.25)).add(p4.data.mul(0.25)))
mod.state_dict()
acc = test_with_dropout(X_test, y_test, mod, state['device'],
state['cuda'])
path = g + str(state['itr']) + 'epoch.' + str(state['acq']) + 'acq.pth.tar'
state['rep'] = path
torch.save({'model_state_dict': mod.state_dict()}, state['rep'])
return mod, acc
def ini_model_train(opt):
X_ini, y_ini, X_test, y_test, X_train_All, y_train_All = ini_model(opt)
mod = Model().to(device)
optimizer = optim.SGD(mod.parameters(), lr=opt.ini_lr)
criterion = nn.CrossEntropyLoss()
num_batches_train = X_ini.shape[0] // opt.ini_batch_size
mod.train()
for i in range(opt.ini_epoch):
loss = 0
for j in range(num_batches_train):
slce = get_slice(j, opt.ini_batch_size)
X_tra = torch.from_numpy(X_ini[slce]).float().to(device)
Y_tra = torch.from_numpy(y_ini[slce]).long().to(device)
optimizer.zero_grad()
out = mod(X_tra)
batch_loss = criterion(out, Y_tra)
batch_loss.backward()
optimizer.step()
loss += batch_loss
mod.eval()
acc = test_without_dropout(X_test, y_test, mod, device)
print('\n[{}/{} epoch], training loss:{:.4f}, test accuracy is:{} \n'.
format(i, opt.ini_epoch,
loss.item() / num_batches_train, acc))
if i + 1 == opt.ini_epoch:
for d in range(opt.num_dev):
torch.save(
{
'epoch': i,
'model_state_dict': mod.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss.item()
},
os.path.join(opt.ini_model_path, 'device' + str(d),
"ini.model.pth.tar"))
torch.save(
{
'epoch': i,
'model_state_dict': mod.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss.item()
}, opt.ini_model_path)
return X_test, y_test, X_train_All, y_train_All
def en_ave(mod1, mod2, mod3, mod4, X_test, y_test, state):
print('=> load average ensemble')
mod = Model().to(device)
for p, p1, p2, p3, p4 in zip(mod.parameters(), mod1.parameters(),
mod2.parameters(), mod3.parameters(),
mod4.parameters()):
p.data.copy_(
p1.data.mul(0.25).add(p2.data.mul(0.25)).add(
p3.data.mul(0.25)).add(p4.data.mul(0.25)))
acc = test_without_dropout(X_test, y_test, mod, device)
path = os.path.join('exp', 'ensemble.') + str(
state['itr']) + 'epoch.' + str(state['acq']) + 'acq.pth.tar'
state['rep'] = path
torch.save({'model_state_dict': mod.state_dict()}, state['rep'])
return mod, acc
def ini_train(X_ini, y_ini, X_te, y_te, epochs, paths, device, batch_size, lr,
momentum, arr_drop):
mod = Model(arr_drop).to(device)
optimizer = optim.SGD(mod.parameters(), lr=lr, momentum=momentum)
criterion = nn.CrossEntropyLoss()
#batch_size = 200
num_batches_train = X_ini.shape[0] // batch_size
print("number of batch ", num_batches_train)
mod.train()
for i in range(epochs):
loss = 0
for j in range(num_batches_train):
slce = get_slice(j, batch_size)
X_tra = torch.from_numpy(X_ini[slce]).float().to(device)
Y_tra = torch.from_numpy(y_ini[slce]).long().to(device)
optimizer.zero_grad()
out = mod(X_tra)
batch_loss = criterion(out, Y_tra)
batch_loss.backward()
optimizer.step()
loss += batch_loss
mod.eval()
with torch.no_grad():
X_va = torch.from_numpy(X_te).float().to(device)
Y_va = torch.from_numpy(y_te).long().to(device)
output = mod(X_va)
preds = torch.max(output, 1)[1]
acc = accuracy_score(Y_va, preds)
print('\n[{}/{} epoch], training loss:{:.4f}, test accuracy is:{} \n'.
format(i, epochs,
loss.item() / num_batches_train, acc))
if i + 1 == epochs:
for path in paths:
torch.save(
{
'epoch': i,
'model_state_dict': mod.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'loss': loss.item()
}, os.path.join(path, "ini.model.pth.tar"))
return mod
def train():
model = Model()
model.to("cuda:0")
Opt = torch.optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3)
# checkpoint = torch.load("./model.pth")
# model.load_state_dict(checkpoint["model"])
# Opt.load_state_dict(checkpoint["Opt"])
for i in range(10000):
Opt.zero_grad()
imgs, targets = read_batch()
imgs = torch.tensor(imgs, dtype=torch.float32).to("cuda:0")
targets = torch.tensor(targets, dtype=torch.float32).to("cuda:0")
preds = model(imgs)
loss = make_loss(preds, targets)
loss.backward()
Opt.step()
if i % 10 == 0:
print("Iteration: %d, Loss: %f"%(i, loss))
if i % 10 == 0:
state = {'model':model.state_dict(), 'Opt':Opt.state_dict(), 'itr':i}
torch.save(state, "./model.pth")
def model(dataset, model_name=None, device=None, train=True):
"""加载模型"""
device = device or torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
net = Model(vocab_size=dataset.vocab_size,
embedding_dim=config.embedding_dim,
output_size=dataset.target_vocab_size,
encoder_hidden_size=config.encoder_hidden_size,
decoder_hidden_size=config.decoder_hidden_size,
encoder_layers=config.encoder_layers,
decoder_layers=config.decoder_layers,
dropout=config.dropout,
embedding_weights=dataset.vector_weights,
device=device)
if model_name: # 如果指定了模型名称, 就加载对应的模型
pre_trained_state_dict = torch.load(FILE_PATH + config.model_path +
model_name,
map_location=device)
state_dict = net.state_dict()
state_dict.update(pre_trained_state_dict)
net.load_state_dict(state_dict)
net.train() if train else net.eval()
return net
position_target = data['point_map']
img = img.to(device=device)
position_target = position_target.to(device=device,
dtype=torch.float32)
pred = model(img)
logits = pred['logits']
# loss calculation
loss_pos_val = loss_position(logits, position_target)
epoch_loss_val += loss_pos_val.item() * img.size(0)
loss_val.append(epoch_loss_val / len(val_indices))
# print statistics
print(f"epoch:[%.d] Validation loss: %.5f" % (epoch + 1, loss_val[-1]))
# Save the model
torch.save(model.state_dict(), 'stats/model_saved.pth')
# Save latent space feature maps along with some log statistics
latent = latent[1:, ...] # removes first, which was an torch.empty
loss_train = np.array(loss_train)
loss_val = np.array(loss_val)
mdic = {
'latent': latent,
'loss_train': loss_train,
'loss_val': loss_val,
'batch_size': batch_size,
'validation_split': validation_split,
'dataset_size': dataset_size,
'random_seed': random_seed
}
savemat("stats/log.mat", mdic)
def random_run(acquisition_iterations, X_Pool, y_Pool, pool_subset,
dropout_iterations, nb_classes, Queries, X_test, y_test, rep,
X_old, y_old, device, itr, cuda, g):
mod = Model().to(device)
if cuda:
cp = torch.load(rep)
print("\n ********load gpu version******* \n")
else:
cp = torch.load(rep, map_location='cpu')
mod.load_state_dict(cp['model_state_dict'])
optimizer = optim.Adam(mod.parameters(), lr=0.001,
weight_decay=0.5) #,weight_decay=0.5
#optimizer = optim.SGD(mod.parameters(), lr=0.001,weight_decay=0.5)
optimizer.load_state_dict(cp['optimizer_state_dict'])
criterion = nn.CrossEntropyLoss()
X_train = np.empty([0, 1, 28, 28])
y_train = np.empty([
0,
])
AA = []
losses_train = []
#acc = test(test_loader,mod,device,cuda)
acc = test(X_test, y_test, mod, device, cuda)
AA.append(acc)
print('initial test accuracy: ', acc)
for i in range(acquisition_iterations):
pool_subset_dropout = np.asarray(
random.sample(range(0, X_Pool.shape[0]), pool_subset))
X_Pool_Dropout = X_Pool[pool_subset_dropout, :, :, :]
y_Pool_Dropout = y_Pool[pool_subset_dropout]
x_pool_index = np.random.choice(X_Pool_Dropout.shape[0],
Queries,
replace=False)
Pooled_X = X_Pool_Dropout[x_pool_index, :, :, :]
Pooled_Y = y_Pool_Dropout[x_pool_index]
delete_Pool_X = np.delete(X_Pool, (pool_subset_dropout), axis=0)
delete_Pool_Y = np.delete(y_Pool, (pool_subset_dropout), axis=0)
delete_Pool_X_Dropout = np.delete(X_Pool_Dropout, (x_pool_index),
axis=0)
delete_Pool_Y_Dropout = np.delete(y_Pool_Dropout, (x_pool_index),
axis=0)
X_Pool = np.concatenate((delete_Pool_X, delete_Pool_X_Dropout), axis=0)
y_Pool = np.concatenate((delete_Pool_Y, delete_Pool_Y_Dropout), axis=0)
print('updated pool size is ', X_Pool.shape[0])
X_train = np.concatenate((X_train, Pooled_X), axis=0)
y_train = np.concatenate((y_train, Pooled_Y), axis=0)
print('number of data points from pool', X_train.shape[0])
batch_size = 100
X = np.vstack((X_old, Pooled_X))
y = np.hstack((y_old, Pooled_Y))
X, y = shuffle(X, y)
num_batch = X.shape[0] // batch_size
print("number of batch: ", num_batch)
mod.train()
for h in range(itr):
losses = 0
for j in range(num_batch):
slce = get_slice(j, batch_size)
X_fog_ = torch.from_numpy(X[slce]).float().to(device)
y_fog_ = torch.from_numpy(y[slce]).long().to(device)
optimizer.zero_grad()
out = mod(X_fog_)
train_loss = criterion(out, y_fog_)
losses += train_loss
train_loss.backward()
optimizer.step()
losses_train.append(losses.item() / num_batch)
acc = test(X_test, y_test, mod, device, cuda)
print('test accuracy: ', acc)
AA.append(acc)
torch.save(
{
'model_state_dict': mod.state_dict(),
'optimizer_state_dict': optimizer.state_dict()
}, g)
return AA, mod, X_train, y_train, losses_train, optimizer
def main(args: argparse.Namespace):
# Load input data
with open(args.train_metadata, 'r') as f:
train_posts = json.load(f)
with open(args.val_metadata, 'r') as f:
val_posts = json.load(f)
# Load labels
labels = {}
with open(args.label_intent, 'r') as f:
intent_labels = json.load(f)
labels['intent'] = {}
for label in intent_labels:
labels['intent'][label] = len(labels['intent'])
with open(args.label_semiotic, 'r') as f:
semiotic_labels = json.load(f)
labels['semiotic'] = {}
for label in semiotic_labels:
labels['semiotic'][label] = len(labels['semiotic'])
with open(args.label_contextual, 'r') as f:
contextual_labels = json.load(f)
labels['contextual'] = {}
for label in contextual_labels:
labels['contextual'][label] = len(labels['contextual'])
# Build dictionary from training set
train_captions = []
for post in train_posts:
train_captions.append(post['orig_caption'])
dictionary = Dictionary(tokenizer_method="TreebankWordTokenizer")
dictionary.build_dictionary_from_captions(train_captions)
# Set up torch device
if 'cuda' in args.device and torch.cuda.is_available():
device = torch.device(args.device)
kwargs = {'pin_memory': True}
else:
device = torch.device('cpu')
kwargs = {}
# Set up number of workers
num_workers = min(multiprocessing.cpu_count(), args.num_workers)
# Set up data loaders differently based on the task
# TODO: Extend to ELMo + word2vec etc.
if args.type == 'image_only':
train_dataset = ImageOnlyDataset(train_posts, labels)
val_dataset = ImageOnlyDataset(val_posts, labels)
train_data_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=args.shuffle,
num_workers=num_workers,
collate_fn=collate_fn_pad_image_only,
**kwargs)
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
num_workers=num_workers,
collate_fn=collate_fn_pad_image_only,
**kwargs)
elif args.type == 'image_text':
train_dataset = ImageTextDataset(train_posts, labels, dictionary)
val_dataset = ImageTextDataset(val_posts, labels, dictionary)
train_data_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=args.shuffle,
num_workers=num_workers,
collate_fn=collate_fn_pad_image_text,
**kwargs)
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
num_workers=num_workers,
collate_fn=collate_fn_pad_image_text,
**kwargs)
elif args.type == 'text_only':
train_dataset = TextOnlyDataset(train_posts, labels, dictionary)
val_dataset = TextOnlyDataset(val_posts, labels, dictionary)
train_data_loader = torch.utils.data.DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=args.shuffle,
num_workers=num_workers,
collate_fn=collate_fn_pad_text_only,
**kwargs)
val_data_loader = torch.utils.data.DataLoader(val_dataset,
batch_size=1,
num_workers=num_workers,
collate_fn=collate_fn_pad_text_only,
**kwargs)
# Set up the model
model = Model(vocab_size=dictionary.size()).to(device)
# Set up an optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=args.lr_scheduler_step_size, gamma=args.lr_scheduler_gamma) # decay by 0.1 every 15 epochs
# Set up loss function
loss_fn = torch.nn.CrossEntropyLoss()
# Setup tensorboard
if args.tensorboard:
writer = tensorboard.SummaryWriter(log_dir=args.log_dir + "/" + args.name, flush_secs=1)
else:
writer = None
# Training loop
if args.classification == 'intent':
keys = ['intent']
elif args.classification == 'semiotic':
keys = ['semiotic']
elif args.classification == 'contextual':
keys = ['contextual']
elif args.classification == 'all':
keys = ['intent', 'semiotic', 'contextual']
else:
raise ValueError("args.classification doesn't exist.")
best_auc_ovr = 0.0
best_auc_ovo = 0.0
best_acc = 0.0
best_model = None
best_optimizer = None
best_scheduler = None
for epoch in range(args.epochs):
for mode in ["train", "eval"]:
# Set up a progress bar
if mode == "train":
pbar = tqdm.tqdm(enumerate(train_data_loader), total=len(train_data_loader))
model.train()
else:
pbar = tqdm.tqdm(enumerate(val_data_loader), total=len(val_data_loader))
model.eval()
total_loss = 0
label = dict.fromkeys(keys, np.array([], dtype=np.int))
pred = dict.fromkeys(keys, None)
for _, batch in pbar:
if 'caption' not in batch:
caption_data = None
else:
caption_data = batch['caption'].to(device)
if 'image' not in batch:
image_data = None
else:
image_data = batch['image'].to(device)
label_batch = {}
for key in keys:
label_batch[key] = batch['label'][key].to(device)
if mode == "train":
model.zero_grad()
pred_batch = model(image_data, caption_data)
for key in keys:
label[key] = np.concatenate((label[key], batch['label'][key].cpu().numpy()))
x = pred_batch[key].detach().cpu().numpy()
x_max = np.max(x, axis=1).reshape(-1, 1)
z = np.exp(x - x_max)
prediction_scores = z / np.sum(z, axis=1).reshape(-1, 1)
if pred[key] is not None:
pred[key] = np.vstack((pred[key], prediction_scores))
else:
pred[key] = prediction_scores
loss_batch = {}
loss = None
for key in keys:
loss_batch[key] = loss_fn(pred_batch[key], label_batch[key])
if loss is None:
loss = loss_batch[key]
else:
loss += loss_bath[key]
total_loss += loss.item()
if mode == "train":
loss.backward()
optimizer.step()
# Terminate the progress bar
pbar.close()
# Update lr scheduler
if mode == "train":
scheduler.step()
for key in keys:
auc_score_ovr = roc_auc_score(label[key], pred[key], multi_class='ovr') # pylint: disable-all
auc_score_ovo = roc_auc_score(label[key], pred[key], multi_class='ovo') # pylint: disable-all
accuracy = accuracy_score(label[key], np.argmax(pred[key], axis=1))
print("[{} - {}] [AUC-OVR={:.3f}, AUC-OVO={:.3f}, ACC={:.3f}]".format(mode, key, auc_score_ovr, auc_score_ovo, accuracy))
if mode == "eval":
best_auc_ovr = max(best_auc_ovr, auc_score_ovr)
best_auc_ovo = max(best_auc_ovo, auc_score_ovo)
best_acc = max(best_acc, accuracy)
best_model = model
best_optimizer = optimizer
best_scheduler = scheduler
if writer:
writer.add_scalar('AUC-OVR/{}-{}'.format(mode, key), auc_score_ovr, epoch)
writer.add_scalar('AUC-OVO/{}-{}'.format(mode, key), auc_score_ovo, epoch)
writer.add_scalar('ACC/{}-{}'.format(mode, key), accuracy, epoch)
writer.flush()
if writer:
writer.add_scalar('Loss/{}'.format(mode), total_loss, epoch)
writer.flush()
print("[{}] Epoch {}: Loss = {}".format(mode, epoch, total_loss))
hparam_dict = {
'train_split': args.train_metadata,
'val_split': args.val_metadata,
'lr': args.lr,
'epochs': args.epochs,
'batch_size': args.batch_size,
'num_workers': args.num_workers,
'shuffle': args.shuffle,
'lr_scheduler_gamma': args.lr_scheduler_gamma,
'lr_scheduler_step_size': args.lr_scheduler_step_size,
}
metric_dict = {
'AUC-OVR': best_auc_ovr,
'AUC-OVO': best_auc_ovo,
'ACC': best_acc
}
if writer:
writer.add_hparams(hparam_dict=hparam_dict, metric_dict=metric_dict)
writer.flush()
Path(args.output_dir).mkdir(exist_ok=True)
torch.save({
'hparam_dict': hparam_dict,
'metric_dict': metric_dict,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict(),
}, Path(args.output_dir) / '{}.pt'.format(args.name))
