def test_running():
epochs = 100000
seq_batch_size = 100
print_yes = 100
loss_func = torch.nn.functional.nll_loss
# create network and optimizer
net = RNN(100, 120, 150, 2)
net.to(device) # add cuda to device
optim = torch.optim.Adam(net.parameters(), lr=3e-5)
# main training loop:
for epoch in range(epochs):
dat = get_batch(train_data, seq_batch_size)
dat = torch.LongTensor([vocab.find(item) for item in dat])
# pull x and y
x_t = dat[:-1]
y_t = dat[1:]
hidden = net.init_hidden()
# turn all into cuda
x_t, y_t, hidden = x_t.to(device), y_t.to(device), hidden.to(device)
# initialize hidden state and forward pass
logprob, hidden = net.forward(x_t, hidden)
loss = loss_func(logprob, y_t)
# update
optim.zero_grad()
loss.backward()
optim.step()
# print the loss for every kth iteration
if epoch % print_yes == 0:
print('*' * 100)
print('\n epoch {}, loss:{} \n'.format(epoch, loss))
# make sure to pass True flag for running on cuda
print('sample speech:\n', run_words(net, vocab, 500, True))
def train():
iterators, dataset = data_loaders(batch_size)
model = RNN(input_size, hidden_size, num_layers, dropout, n_classes,
dataset.get_class_weights())
optimizer = optim.SGD(model.parameters(), lr)
log = train_procedure(model, iterators, n_epochs, optimizer)
save_log(log)
def load_model(args):
if args.model == 'rnn':
model = RNN(25, args.dim_hidden, 25)
elif args.model == 'ffnn':
model = FFNN(25,
args.dim_hidden,
memory=args.memory,
num_hidden_layers=args.num_hidden_layers)
elif args.model == 'cnn3d':
model = CNN3D((10, 10), ((4, 4, 2), (7, 12, 2)), 10, 7, 10, 7, 2,
100) #/TODO un hard code
else:
raise Exception(
"Only rnn, ffnn, cnn3d model types currently supported")
if args.loss_fn == 'mse':
loss_fn = torch.nn.MSELoss()
else:
raise Exception("Only mse loss function currently supported")
if args.optimizer == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
else:
raise Exception("Only adam optimizer currently supported")
return model, loss_fn, optimizer
def attention_test(args):
chosen_params = dict(params)
results = []
for attention in [False, True]:
chosen_params['attention'] = attention
runs = 5
acc_d = {}
f1_d = {}
for i in tqdm(range(runs)):
train_dataset, valid_dataset, test_dataset = data.load_dataset(
args.train_batch_size,
args.test_batch_size,
min_freq=chosen_params['min_freq'])
embedding = data.generate_embedding_matrix(
train_dataset.dataset.text_vocab,
rand=chosen_params['rand_emb'],
freeze=chosen_params['freeze'])
model = RNN(embedding, chosen_params)
criterion = nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=chosen_params['lr'])
for epoch in range(args.epochs):
print(f'******* epoch: {epoch+1} *******')
train(model, train_dataset, optimizer, criterion, args)
evaluate(model, valid_dataset, criterion, 'Validation')
acc, f1 = evaluate(model, test_dataset, criterion, 'Test')
acc_d['acc_' + 'run' + str(i)] = acc
f1_d['f1_' + 'run' + str(i)] = f1
mean = np.mean(list(acc_d.values()))
std = np.std(list(acc_d.values()))
acc_d['mean'] = mean
acc_d['std'] = std
mean = np.mean(list(f1_d.values()))
std = np.std(list(f1_d.values()))
f1_d['mean'] = mean
f1_d['std'] = std
results.append((acc_d, f1_d))
with open(os.path.join(SAVE_DIR, 'attention.txt'), 'a') as f:
print(f'', file=f)
for idx, (acc, f1) in enumerate(results):
print('[no attention]' if idx == 0 else '[attention]', file=f)
print(acc, file=f)
print(f1, file=f)
def hyperparam_optim_test(args):
var_params = {
'cell_name': ['lstm'],
'hidden_size': [50, 150, 300],
'num_layers': [2, 4, 5],
'min_freq': [0, 100, 500],
'lr': [1e-3, 1e-4, 1e-7],
'dropout': [0, 0.4, 0.7],
'freeze': [False, True],
'rand_emb': [False, True],
'attention': [False]
}
results = []
for i in tqdm(range(10)):
chosen_params = {k: random.choice(v) for (k, v) in var_params.items()}
train_dataset, valid_dataset, test_dataset = data.load_dataset(
args.train_batch_size,
args.test_batch_size,
min_freq=chosen_params['min_freq'])
embedding = data.generate_embedding_matrix(
train_dataset.dataset.text_vocab,
rand=chosen_params['rand_emb'],
freeze=chosen_params['freeze'])
model = RNN(embedding, chosen_params)
criterion = nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=chosen_params['lr'])
for epoch in range(args.epochs):
print(f'******* epoch: {epoch+1} *******')
train(model, train_dataset, optimizer, criterion, args)
evaluate(model, valid_dataset, criterion, 'Validation')
acc, f1 = evaluate(model, test_dataset, criterion, 'Test')
result = dict(chosen_params)
result['acc'] = acc
result['f1'] = f1
results.append(result)
with open(os.path.join(SAVE_DIR, 'params_search.txt'), 'a') as f:
for result in results:
print(result, file=f)
def cell_comparison_test(args):
train_dataset, valid_dataset, test_dataset = data.load_dataset(
args.train_batch_size, args.test_batch_size)
embedding = data.generate_embedding_matrix(
train_dataset.dataset.text_vocab)
var_params = {
'hidden_size': [50, 150, 300],
'num_layers': [1, 2, 4],
'dropout': [0.1, 0.4, 0.7],
'bidirectional': [True, False],
'attention': [False]
}
for idx, (key, values) in enumerate(var_params.items()):
fig, ax = plt.subplots(nrows=1, ncols=1)
ax.set_title('Variable ' + key)
for cell_name in tqdm(['rnn', 'lstm', 'gru']):
results = []
for i in range(len(values)):
current_params = {
k: v[i] if k == key else v[1]
for (k, v) in var_params.items()
}
current_params['cell_name'] = cell_name
model = RNN(embedding, current_params)
model
criterion = nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for epoch in range(args.epochs):
print(f'******* epoch: {epoch+1} *******')
train(model, train_dataset, optimizer, criterion, args)
evaluate(model, valid_dataset, criterion, 'Validation')
result, _ = evaluate(model, test_dataset, criterion, 'Test')
results.append(result)
ax.plot(values, results, marker='o', label=cell_name)
plt.legend(loc='best')
plt.xlabel(key)
plt.ylabel('accuracy')
fig.savefig(os.path.join(SAVE_DIR, key + '.png'))
plt.close(fig)
def main():
epochs = 301
seq_batch_size = 200
print_yes = 100
iscuda = False
# create our network, optimizer and loss function
net = RNN(len(chars), 100, 150, 2) #instanciate a RNN object
optim = torch.optim.Adam(net.parameters(), lr=6e-4)
loss_func = torch.nn.functional.nll_loss
if iscuda:
net = net.cuda()
# main training loop:
for epoch in range(epochs):
dat = getSequence(book, seq_batch_size)
dat = torch.LongTensor(
[chars.find(item) for item in dat]
) #find corresponding char index for each character and store this in tensor
# pull x, y and initialize hidden state
if iscuda:
x_t = dat[:-1].cuda()
y_t = dat[1:].cuda()
hidden = net.init_hidden().cuda()
else:
x_t = dat[:-1]
y_t = dat[1:]
hidden = net.init_hidden()
# forward pass
logprob, hidden = net.forward(x_t, hidden)
loss = loss_func(logprob, y_t)
# update
optim.zero_grad()
loss.backward()
optim.step()
# print the loss for every kth iteration
if epoch % print_yes == 0:
print('*' * 60)
print('\n epoch {}, loss:{} \n'.format(epoch, loss))
print('sample speech:\n', test_words(net, chars, seq_batch_size))
torch.save(net.state_dict(), 'trainedBook_v2.pt')
def main(args):
chosen_params = dict(params)
chosen_params['attention'] = True
train_dataset, valid_dataset, test_dataset = data.load_dataset(
args.train_batch_size, args.test_batch_size)
embedding = data.generate_embedding_matrix(
train_dataset.dataset.text_vocab, freeze=False)
model = RNN(embedding, chosen_params)
criterion = nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for epoch in range(args.epochs):
print(f'******* epoch: {epoch} *******')
train(model, train_dataset, optimizer, criterion, args)
evaluate(model, valid_dataset, criterion, 'Validation')
evaluate(model, test_dataset, criterion, 'Test')
def no_test_forward():
loss_func = torch.nn.functional.nll_loss
net = RNN(100, 100, 100)
net.to(device) # add cuda to device
optim = torch.optim.Adam(net.parameters(), lr=1e-4)
# step 2: create a training batch of data, size 101, format this data and convert it to pytorch long tensors
dat = get_batch(train_data, 100)
dat = torch.LongTensor([vocab.find(item) for item in dat])
# step 3: convert our dat into input/output
x_t = dat[:-1]
y_t = dat[1:]
ho = net.init_hidden()
# remember to load all variables used by the model to the device, this means the i/o as well as the hidden state
x_t, y_t, ho = x_t.to(device), y_t.to(device), ho.to(device)
# test forward pass
log_prob, hidden = net.forward(x_t, ho)
# let's see if the forward pass of the next hidden state is already cuda
#log_prob2, hidden2 = net.forward(x_t, hidden)
loss = loss_func(log_prob, y_t)
optim.zero_grad()
loss.backward()
optim.step()
def embedding_baseline_test(args):
chosen_params = dict(params)
results = []
for rand_emb in [True, False]:
chosen_params['rand_emb'] = rand_emb
train_dataset, valid_dataset, test_dataset = data.load_dataset(
args.train_batch_size,
args.test_batch_size,
min_freq=chosen_params['min_freq'])
embedding = data.generate_embedding_matrix(
train_dataset.dataset.text_vocab,
rand=chosen_params['rand_emb'],
freeze=chosen_params['freeze'])
result = {}
for m in ['baseline', 'rnn']:
if m == 'rnn':
model = RNN(embedding, chosen_params)
else:
model = Baseline(embedding)
criterion = nn.BCEWithLogitsLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
for epoch in range(args.epochs):
print(f'******* epoch: {epoch} *******')
train(model, train_dataset, optimizer, criterion, args)
evaluate(model, valid_dataset, criterion, 'Validation')
acc, f1 = evaluate(model, test_dataset, criterion, 'Test')
result[m + '_acc_rand_emb' + str(rand_emb)] = acc
result[m + '_f1_rand_emb' + str(rand_emb)] = f1
results.append(result)
with open(os.path.join(SAVE_DIR, 'embedding_baseline.txt'), 'a') as f:
for res in results:
print(res, file=f)
'pretrained_word_embeddings_file': pretrained_word_embeddings_file, 'transform_train': transform_train,
'transform_val': transform_val, 'WEIGHT_DECAY': WEIGHT_DECAY, 'ADAM_FLAG': ADAM_FLAG, 'RNN_DROPOUT':RNN_DROPOUT,
'CNN_DROPOUT': CNN_DROPOUT, 'GRAD_CLIP': GRAD_CLIP}
print('Initializing models...')
encoder = CNN(NO_WORD_EMBEDDINGS, pretrained_cnn_dir, freeze=True, dropout_prob=CNN_DROPOUT, model_name='resnet152')
decoder = RNN(VOCAB_SIZE, NO_WORD_EMBEDDINGS, hidden_size=HIDDEN_SIZE, num_layers=NUM_LAYERS,
pre_trained_file=pretrained_word_embeddings_file, freeze=False, dropout_prob=RNN_DROPOUT)
params['encoder'] = encoder
params['decoder'] = decoder
encoder.cuda()
decoder.cuda()
print('Initializing optimizer...')
model_paras = list(encoder.parameters()) + list(decoder.parameters())
optimizer = optim.Adam(model_paras, lr=LR, weight_decay=WEIGHT_DECAY)
params['optimizer'] = optimizer
pickle.dump(params, open(init_params_file, 'wb'))
# initialize accumulators.
current_epoch = 1
batch_step_count = 1
time_used_global = 0.0
checkpoint = 1
# load lastest model to resume training
class Trainer:
"""
训练
"""
def __init__(self, _hparams):
utils.set_seed(_hparams.fixed_seed)
self.train_loader = get_train_loader(_hparams)
self.val_loader = get_val_loader(_hparams)
self.encoder = CNN().to(DEVICE)
self.decoder = RNN(fea_dim=_hparams.fea_dim,
embed_dim=_hparams.embed_dim,
hid_dim=_hparams.hid_dim,
max_sen_len=_hparams.max_sen_len,
vocab_pkl=_hparams.vocab_pkl).to(DEVICE)
self.loss_fn = nn.CrossEntropyLoss()
self.optimizer = torch.optim.Adam(self.get_params(), lr=_hparams.lr)
self.writer = SummaryWriter()
self.max_sen_len = _hparams.max_sen_len
self.val_cap = _hparams.val_cap
self.ft_encoder_lr = _hparams.ft_encoder_lr
self.ft_decoder_lr = _hparams.ft_decoder_lr
self.best_CIDEr = 0
def fine_tune_encoder(self, fine_tune_epochs, val_interval, save_path,
val_path):
print('*' * 20, 'fine tune encoder for', fine_tune_epochs, 'epochs',
'*' * 20)
self.encoder.fine_tune()
self.optimizer = torch.optim.Adam([
{
'params': self.encoder.parameters(),
'lr': self.ft_encoder_lr
},
{
'params': self.decoder.parameters(),
'lr': self.ft_decoder_lr
},
])
self.training(fine_tune_epochs, val_interval, save_path, val_path)
self.encoder.froze()
print('*' * 20, 'fine tune encoder complete', '*' * 20)
def get_params(self):
"""
模型需要优化的全部参数，此处encoder暂时设计不用训练，故不加参数
:return:
"""
return list(self.decoder.parameters())
def training(self, max_epochs, val_interval, save_path, val_path):
"""
训练
:param val_path: 保存验证过程生成句子的路径
:param save_path: 保存模型的地址
:param val_interval: 验证的间隔
:param max_epochs: 最大训练的轮次
:return:
"""
print('*' * 20, 'train', '*' * 20)
for epoch in range(max_epochs):
self.set_train()
epoch_loss = 0
epoch_steps = len(self.train_loader)
for step, (img, cap,
cap_len) in tqdm(enumerate(self.train_loader)):
# batch_size * 3 * 224 * 224
img = img.to(DEVICE)
cap = cap.to(DEVICE)
self.optimizer.zero_grad()
features = self.encoder.forward(img)
outputs = self.decoder.forward(features, cap)
outputs = pack_padded_sequence(outputs,
cap_len - 1,
batch_first=True)[0]
targets = pack_padded_sequence(cap[:, 1:],
cap_len - 1,
batch_first=True)[0]
train_loss = self.loss_fn(outputs, targets)
epoch_loss += train_loss.item()
train_loss.backward()
self.optimizer.step()
epoch_loss /= epoch_steps
self.writer.add_scalar('epoch_loss', epoch_loss, epoch)
print('epoch_loss: {}, epoch: {}'.format(epoch_loss, epoch))
if (epoch + 1) % val_interval == 0:
CIDEr = self.validating(epoch, val_path)
if self.best_CIDEr <= CIDEr:
self.best_CIDEr = CIDEr
self.save_model(save_path, epoch)
def save_model(self, save_path, train_epoch):
"""
保存最好的模型
:param save_path: 保存模型文件的地址
:param train_epoch: 当前训练的轮次
:return:
"""
model_state_dict = {
'encoder_state_dict': self.encoder.state_dict(),
'decoder_state_dict': self.decoder.state_dict(),
'tran_epoch': train_epoch,
}
print('*' * 20, 'save model to: ', save_path, '*' * 20)
torch.save(model_state_dict, save_path)
def validating(self, train_epoch, val_path):
"""
验证
:param val_path: 保存验证过程生成句子的路径
:param train_epoch: 当前训练的epoch
:return:
"""
print('*' * 20, 'validate', '*' * 20)
self.set_eval()
sen_json = []
with torch.no_grad():
for val_step, (img, img_id) in tqdm(enumerate(self.val_loader)):
img = img.to(DEVICE)
features = self.encoder.forward(img)
sens, _ = self.decoder.sample(features)
sen_json.append({'image_id': int(img_id), 'caption': sens[0]})
with open(val_path, 'w') as f:
json.dump(sen_json, f)
result = coco_eval(self.val_cap, val_path)
scores = {}
for metric, score in result:
scores[metric] = score
self.writer.add_scalar(metric, score, train_epoch)
return scores['CIDEr']
def set_train(self):
self.encoder.train()
self.decoder.train()
def set_eval(self):
self.encoder.eval()
self.decoder.eval()
train_X, train_y = util.create_dataset(training_set_scaled, input_size=INPUT_SIZE)
# print(train_X.shape)
''' (672, 60) '''
# Reshape为 (batch, time_step, input_size), 这是放入LSTM的shape
train_X = train_X.reshape(train_X.shape[0], 1, train_X.shape[1])
# print(train_X.shape)
''' (672, 1, 60) '''
# Part 2 - Building the RNN
rnn = RNN(INPUT_SIZE, HIDDEN_SIZE, NUM_LAYERS, OUTPUT_SIZE)
optimiser = torch.optim.Adam(rnn.parameters(), lr=args.lr)
loss_func = nn.MSELoss()
if CUDA:
loss_func.cuda()
plt.figure(1, figsize=(12, 5))
plt.ion()
hidden_state = None
for epoch in range(args.epochs):
inputs = Variable(torch.from_numpy(train_X).float())
labels = Variable(torch.from_numpy(train_y).float())
if CUDA:
inputs, labels = inputs.cuda(), labels.cuda()
class TextClassifier:
def __init__(self, batch_size, iterations, initial_lr, hidden_size,
dropout, kernel_sz, num_layers):
self.use_cuda = torch.cuda.is_available()
self.device = torch.device('cuda:0' if self.use_cuda else 'cpu')
self.data = DataReader()
train_iter, val_iter, test_iter = self.data.init_dataset(
batch_size, ('cuda:0' if self.use_cuda else 'cpu'))
self.train_batch_loader = BatchGenerator(train_iter, 'text', 'label')
self.val_batch_loader = BatchGenerator(val_iter, 'text', 'label')
self.test_batch_loader = BatchGenerator(test_iter, 'text', 'label')
# Store hyperparameters
self.batch_size = batch_size
self.iterations = iterations
self.initial_lr = initial_lr
# Create Model
emb_size, emb_dim = self.data.TEXT.vocab.vectors.size()
# padding = (math.floor(kernel_sz / 2), 0)
# self.model = CNN(emb_size=emb_size, emb_dimension=emb_dim,
# output_size=len(self.data.LABEL.vocab),
# dropout=dropout, kernel_sz=kernel_sz, stride=1, padding=padding,
# out_filters=hidden_size, pretrained_emb=self.data.TEXT.vocab.vectors)
self.model = RNN(emb_size=emb_size,
emb_dimension=emb_dim,
pretrained_emb=self.data.TEXT.vocab.vectors,
output_size=len(self.data.LABEL.vocab),
num_layers=num_layers,
hidden_size=hidden_size,
dropout=dropout)
if self.use_cuda:
self.model.cuda()
def train(self, min_stride=3):
train_loss_hist = []
val_loss_hist = []
train_acc_hist = []
val_acc_hist = []
test_acc_hist = []
best_score = 0.0
loss = 0.0
for itr in range(self.iterations):
print("\nIteration: " + str(itr + 1))
optimizer = optim.SGD(self.model.parameters(), lr=self.initial_lr)
self.model.train()
total_loss = 0.0
total_acc = 0.0
steps = 0
data_iter = iter(self.train_batch_loader)
for i in range(len(self.train_batch_loader)):
((x_batch, x_len_batch), y_batch) = next(data_iter)
# if torch.min(x_len_batch) > min_stride:
optimizer.zero_grad()
loss, logits = self.model.forward(x_batch, y_batch)
acc = torch.sum(torch.argmax(logits, dim=1) == y_batch)
total_loss += loss.item()
total_acc += acc.item()
steps += 1
loss.backward()
optimizer.step()
train_loss_hist.append(total_loss / steps)
train_acc_hist.append(total_acc / len(self.data.train_data))
val_loss, val_acc = self.eval_model(self.val_batch_loader,
len(self.data.val_data))
val_loss_hist.append(val_loss)
val_acc_hist.append(val_acc)
if val_acc > best_score:
best_score = val_acc
test_loss, test_acc = self.eval_model(self.test_batch_loader,
len(self.data.test_data))
print("Train: {Loss: " + str(total_loss / steps) + ", Acc: " +
str(total_acc / len(self.data.train_data)) + " }")
print("Val: {Loss: " + str(val_loss) + ", Acc: " + str(val_acc) +
" }")
test_acc_hist.append(test_acc)
return train_loss_hist, train_acc_hist, val_loss_hist, val_acc_hist, test_acc_hist
def eval_model(self, batch_loader, N, min_stride=3):
self.model.eval()
total_loss = 0.0
total_acc = 0.0
steps = 0
batch_iterator = iter(batch_loader)
with torch.no_grad():
for i in range(len(batch_loader)):
((x_batch, x_len_batch), y_batch) = next(batch_iterator)
# if torch.min(x_len_batch) > min_stride:
loss, logits = self.model(x_batch, y_batch)
acc = torch.sum(torch.argmax(logits, dim=1) == y_batch)
total_loss += loss.item()
total_acc += acc.item()
return (total_loss / N), (total_acc / N)
# different things here than in the RNNs.
# Also, the Transformer also has other hyperparameters
# (such as the number of attention heads) which can change it's behavior.
model = TRANSFORMER(vocab_size=vocab_size, n_units=args.hidden_size,
n_blocks=args.num_layers, dropout=1.-args.dp_keep_prob)
# these 3 attributes don't affect the Transformer's computations;
# they are only used in run_epoch
model.batch_size=args.batch_size
model.seq_len=args.seq_len
model.vocab_size=vocab_size
else:
print("Model type not recognized.")
model = model.to(device)
total_params = sum(p.numel() for p in model.parameters())
print("===> Total Model Parameters: ", total_params)
with open (os.path.join(experiment_path,'exp_config.txt'), 'a') as f:
f.write('Total Model Parameters:'+' '+str(total_params)+'\n')
# LOSS FUNCTION
loss_fn = nn.CrossEntropyLoss()
if args.optimizer == 'ADAM':
optimizer = torch.optim.Adam(model.parameters(), lr=args.initial_lr)
# LEARNING RATE SCHEDULE
lr = args.initial_lr
lr_decay_base = 1 / 1.15
m_flat_lr = 14.0 # we will not touch lr for the first m_flat_lr epochs
args = get_args()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
if not os.path.exists(args.prepro_root):
os.makedirs(args.prepro_root)
train_dataset = torch.load(args.prepro_root + args.train_dataset_path)
valid_dataset = torch.load(args.prepro_root + args.valid_dataset_path)
train_loader = data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True)
valid_loader = data.DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False)
valid_tensor = torch.load(args.prepro_root+'valid_writer_keywd.pkl').to(device)
model = RNN(args.num_readers, args.num_writers, args.num_keywords,
args.num_items, args.num_magazines, args.hid_dim, valid_tensor).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
criterion = torch.nn.CrossEntropyLoss(ignore_index=0)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=5, gamma=0.5)
model_parameters = filter(lambda p: p.requires_grad, model.parameters())
params = sum([np.prod(p.size()) for p in model_parameters])
print(model)
print('# of params : ', params)
if args.start_epoch:
model.load_state_dict(torch.load(args.save_path+'%d_rnn_attention.pkl' % args.start_epoch))
best_loss = 9999999
for epoch in range(args.num_epochs):
model.train()
for i, data in enumerate(tqdm.tqdm(train_loader, desc='Train')):
# chunk into sequences of length seq_length + 1
batches = list(chunks(batches, seq_length + 1))
# chunk sequences into batches
batches = list(chunks(batches, minibatch_size))
# convert batches to tensors and transpose
batches = [torch.LongTensor(batch).transpose_(0, 1) for batch in batches]
loss_function = nn.CrossEntropyLoss()
model = RNN(minibatch_size, chars_len, hidden_size, chars_len, n_layers,
minibatch_size).to(DEVICE)
hidden = Variable(model.create_hidden()).to(DEVICE)
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
def loop(itr, extras, stateful):
batch_tensor = next(itr)
if use_cuda:
batch_tensor = batch_tensor.cuda()
# reset the model
model.zero_grad()
# everything except the last
input_variable = Variable(batch_tensor[:-1]).to(DEVICE)
# everything except the first, flattened
target_variable = Variable(
def main():
args = get_args()
log.info(f'Parsed arguments: \n{pformat(args.__dict__)}')
assert args.cond_type.lower() in ['none', 'platanios', 'oestling']
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
log.info('Using device {}.'.format(device))
use_apex = False
if torch.cuda.is_available() and args.fp16:
log.info('Loading Nvidia Apex and using AMP')
from apex import amp, optimizers
use_apex = True
else:
log.info('Using FP32')
amp = None
log.info(f'Using time stamp {timestamp} to save models and logs.')
if not args.no_seed:
log.info(f'Setting random seed to {args.seed} for reproducibility.')
torch.manual_seed(args.seed)
random.seed(args.seed)
data = Corpus(args.datadir)
data_splits = [
{
'split': 'train',
'languages': args.dev_langs + args.target_langs,
'invert_include': True,
},
{
'split': 'valid',
'languages': args.dev_langs,
},
{
'split': 'test',
'languages': args.target_langs,
},
]
if args.refine:
data_splits.append({
'split': 'train_100',
'languages': args.target_langs,
'ignore_missing': True,
})
data_splits = data.make_datasets(data_splits, force_rebuild=args.rebuild)
train_set, val_set, test_set = data_splits['train'], data_splits[
'valid'], data_splits['test']
dictionary = data_splits['dictionary']
train_language_distr = get_sampling_probabilities(train_set, 1.0)
train_set = Dataset(train_set,
batchsize=args.batchsize,
bptt=args.bptt,
reset_on_iter=True,
language_probabilities=train_language_distr)
val_set = Dataset(val_set,
make_config=True,
batchsize=args.valid_batchsize,
bptt=args.bptt,
eval=True)
test_set = Dataset(test_set,
make_config=True,
batchsize=args.test_batchsize,
bptt=args.bptt,
eval=True)
train_loader = DataLoader(train_set, num_workers=args.workers)
val_loader = DataLoader(val_set, num_workers=args.workers)
test_loader = DataLoader(test_set, num_workers=args.workers)
if args.refine:
refine_set = dict()
for lang, lang_d in data_splits['train_100'].items():
refine_set[lang] = Dataset({lang: lang_d},
batchsize=args.valid_batchsize,
bptt=args.bptt,
make_config=True)
n_token = len(dictionary.idx2tkn)
# Load and preprocess matrix of typological features
# TODO: implement this, the OEST
# prior_matrix = load_prior(args.prior, corpus.dictionary.lang2idx)
# n_components = min(50, *prior_matrix.shape)
# pca = PCA(n_components=n_components, whiten=True)
# prior_matrix = pca.fit_transform(prior_matrix)
prior = None
model = RNN(args.cond_type,
prior,
n_token,
n_input=args.emsize,
n_hidden=args.nhidden,
n_layers=args.nlayers,
dropout=args.dropouto,
dropoute=args.dropoute,
dropouth=args.dropouth,
dropouti=args.dropouti,
wdrop=args.wdrop,
wdrop_layers=[0, 1, 2],
tie_weights=True).to(device)
if args.opt_level != 'O2':
loss_function = SplitCrossEntropyLoss(args.emsize,
splits=[]).to(device)
else:
loss_function = CrossEntropyLoss().to(
device) # Should be ok to use with a vocabulary of this small size
if use_apex:
optimizer = optimizers.FusedAdam(model.parameters(),
lr=args.lr,
weight_decay=args.wdecay)
else:
params = list(filter(lambda p: p.requires_grad,
model.parameters())) + list(
loss_function.parameters())
optimizer = Adam(params, lr=args.lr, weight_decay=args.wdecay)
if use_apex:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.opt_level)
parameters = {
'model': model,
'optimizer': optimizer,
'loss_function': loss_function,
'use_apex': use_apex,
'amp': amp if use_apex else None,
'clip': args.clip,
'alpha': args.alpha,
'beta': args.beta,
'bptt': args.bptt,
'device': device,
'prior': args.prior,
}
# Add backward hook for gradient clipping
if args.clip:
if use_apex:
for p in amp.master_params(optimizer):
p.register_hook(
lambda grad: torch.clamp(grad, -args.clip, args.clip))
else:
for p in model.parameters():
p.register_hook(
lambda grad: torch.clamp(grad, -args.clip, args.clip))
if args.prior == 'vi':
prior = VIPrior(model, device=device)
parameters['prior'] = prior
def sample_weights(module: torch.nn.Module, input: torch.Tensor):
prior.sample_weights(module)
sample_weights_hook = model.register_forward_pre_hook(sample_weights)
# Load model checkpoint if available
start_epoch = 1
if args.resume:
if args.checkpoint is None:
log.error(
'No checkpoint passed. Specify it using the --checkpoint flag')
checkpoint = None
else:
log.info('Loading the checkpoint at {}'.format(args.checkpoint))
checkpoint = load_model(args.checkpoint, **parameters)
start_epoch = checkpoint['epoch']
if args.wdrop:
for rnn in model.rnns:
if isinstance(rnn, WeightDrop):
rnn.dropout = args.wdrop
elif rnn.zoneout > 0:
rnn.zoneout = args.wdrop
saved_models = list()
result_str = '| Language {} | test loss {:5.2f} | test ppl {:8.2f} | test bpc {:8.3f}'
def test():
log.info('=' * 89)
log.info('Running test set (zero-shot results)...')
test_loss, avg_loss = evaluate(test_loader, **parameters)
log.info('Test set finished | test loss {} | test bpc {}'.format(
test_loss, test_loss / math.log(2)))
for lang, avg_l_loss in avg_loss.items():
langstr = dictionary.idx2lang[lang]
log.info(
result_str.format(langstr, avg_l_loss, math.exp(avg_l_loss),
avg_l_loss / math.log(2)))
log.info('=' * 89)
if args.train:
f = 1.
stored_loss = 1e32
epochs_no_improve = 0
val_losses = list()
# calculate specific language lr
data_spec_count = sum([len(ds) for l, ds in train_set.data.items()])
data_spec_avg = data_spec_count / len(train_set.data.items())
data_spec_lrweights = dict([(l, data_spec_avg / len(ds))
for l, ds in train_set.data.items()])
# estimate total number of steps
total_steps = sum(
[len(ds) // args.bptt
for l, ds in train_set.data.items()]) * args.no_epochs
steps = 0
try:
pbar = tqdm.trange(start_epoch,
args.no_epochs + 1,
position=1,
dynamic_ncols=True)
for epoch in pbar:
steps = train(train_loader,
lr_weights=data_spec_lrweights,
**parameters,
total_steps=total_steps,
steps=steps,
scaling=args.scaling,
n_samples=args.n_samples,
tb_writer=tb_writer)
val_loss, _ = evaluate(val_loader, **parameters)
pbar.set_description('Epoch {} | Val loss {}'.format(
epoch, val_loss))
# Save model
if args.prior == 'vi':
sample_weights_hook.remove()
filename = path.join(
args.checkpoint_dir, '{}_epoch{}{}_{}.pth'.format(
timestamp, epoch, '_with_apex' if use_apex else '',
args.prior))
torch.save(make_checkpoint(epoch + 1, **parameters), filename)
saved_models.append(filename)
if args.prior == 'vi':
sample_weights_hook = model.register_forward_pre_hook(
sample_weights)
# Early stopping
if val_loss < stored_loss:
epochs_no_improve = 0
stored_loss = val_loss
else:
epochs_no_improve += 1
if epochs_no_improve == args.patience:
log.info('Early stopping at epoch {}'.format(epoch))
break
val_losses.append(val_loss)
# Reduce lr every 1/3 total epochs
if epoch - 1 > f / 3 * args.no_epochs:
log.info('Epoch {}/{}. Dividing LR by 10'.format(
epoch, args.no_epochs))
for g in optimizer.param_groups:
g['lr'] = g['lr'] / 10
f += 1.
test()
except KeyboardInterrupt:
log.info('Registered KeyboardInterrupt. Stopping training.')
log.info('Saving last model to disk')
if args.prior == 'vi':
sample_weights_hook.remove()
torch.save(
make_checkpoint(epoch, **parameters),
path.join(
args.checkpoint_dir, '{}_epoch{}{}_{}.pth'.format(
timestamp, epoch, '_with_apex' if use_apex else '',
args.prior)))
return
elif args.test:
test()
# Only test on existing languages if there are no held out languages
if not args.target_langs:
exit(0)
importance = 1e-5
# If use UNIV, calculate informed prior, else use boring prior
if args.prior == 'laplace':
if not isinstance(
prior,
LaplacePrior): # only calculate matrix if it is not supplied.
log.info('Creating laplace approximation dataset')
laplace_set = Dataset(data_splits['train'],
batchsize=args.batchsize,
bptt=100,
reset_on_iter=True)
laplace_loader = DataLoader(laplace_set, num_workers=args.workers)
log.info('Creating Laplacian prior')
prior = LaplacePrior(model,
loss_function,
laplace_loader,
use_apex=use_apex,
amp=amp,
device=device)
parameters['prior'] = prior
torch.save(
make_checkpoint('fisher_matrix', **parameters),
path.join(
args.checkpoint_dir, '{}_fishers_matrix{}_{}.pth'.format(
timestamp, '_with_apex' if use_apex else '',
args.prior)))
importance = 1e5
elif args.prior == 'ninf':
log.info('Creating non-informative Gaussian prior')
parameters['prior'] = GaussianPrior()
elif args.prior == 'vi':
importance = 1e-5
elif args.prior == 'hmc':
raise NotImplementedError
else:
raise ValueError(
f'Passed prior {args.prior} is not an implemented inference technique.'
)
best_model = saved_models[-1] if not len(
saved_models) == 0 else args.checkpoint
# Remove sampling hook from model
if args.prior == 'vi':
sample_weights_hook.remove()
# Refine on 100 samples on each target
if args.refine:
# reset learning rate
optimizer.param_groups[0]['lr'] = args.lr
loss = 0
results = dict()
# Create individual tests sets
test_sets = dict()
for lang, lang_d in data_splits['test'].items():
test_sets[lang] = DataLoader(Dataset({lang: lang_d},
make_config=True,
batchsize=args.test_batchsize,
bptt=args.bptt,
eval=True),
num_workers=args.workers)
for lang, lang_data in tqdm.tqdm(refine_set.items()):
final_loss = False
refine_dataloader = DataLoader(lang_data, num_workers=args.workers)
load_model(best_model, **parameters)
log.info(f'Refining for language {dictionary.idx2lang[lang]}')
for epoch in range(1, args.refine_epochs + 1):
refine(refine_dataloader, **parameters, importance=importance)
if epoch % 5 == 0:
final_loss = True
loss, avg_loss = evaluate(test_sets[lang],
model,
loss_function,
only_l=lang,
report_all=True,
device=device)
for lang, avg_l_loss in avg_loss.items():
langstr = dictionary.idx2lang[lang]
log.debug(
result_str.format(langstr, avg_l_loss,
math.exp(avg_l_loss),
avg_l_loss / math.log(2)))
if not final_loss:
loss, avg_loss = evaluate(test_sets[lang],
model,
loss_function,
only_l=lang,
report_all=True,
device=device)
for lang, avg_l_loss in avg_loss.items():
langstr = dictionary.idx2lang[lang]
log.info(
result_str.format(langstr, avg_l_loss,
math.exp(avg_l_loss),
avg_l_loss / math.log(2)))
results[lang] = avg_l_loss
log.info('=' * 89)
log.info('FINAL FEW SHOT RESULTS: ')
log.info('=' * 89)
for lang, avg_l_loss in results.items():
langstr = dictionary.idx2lang[lang]
log.info(
result_str.format(langstr, avg_l_loss, math.exp(avg_l_loss),
avg_l_loss / math.log(2)))
log.info('=' * 89)
def main(args):
this_dir = osp.join(osp.dirname(__file__), '.')
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
datasets = {
phase: DataLayer(
data_root=osp.join(args.data_root, phase),
phase=phase,
)
for phase in args.phases
}
data_loaders = {
phase: data.DataLoader(
datasets[phase],
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
)
for phase in args.phases
}
model = Model(
input_size=args.input_size,
hidden_size=args.hidden_size,
bidirectional=args.bidirectional,
num_classes=args.num_classes,
).apply(utl.weights_init).to(device)
criterion = nn.CrossEntropyLoss().to(device)
softmax = nn.Softmax(dim=1).to(device)
optimizer = optim.RMSprop(model.parameters(), lr=args.lr)
for epoch in range(args.start_epoch, args.start_epoch + args.epochs):
losses = {phase: 0.0 for phase in args.phases}
corrects = {phase: 0.0 for phase in args.phases}
start = time.time()
for phase in args.phases:
training = 'Test' not in phase
if training:
model.train(True)
else:
if epoch in args.test_intervals:
model.train(False)
else:
continue
with torch.set_grad_enabled(training):
for batch_idx, (spatial, temporal, length,
target) in enumerate(data_loaders[phase]):
spatial_input = torch.zeros(*spatial.shape)
temporal_input = torch.zeros(*temporal.shape)
target_input = []
length_input = []
index = utl.argsort(length)[::-1]
for i, idx in enumerate(index):
spatial_input[i] = spatial[idx]
temporal_input[i] = temporal[idx]
target_input.append(target[idx])
length_input.append(length[idx])
spatial_input = spatial_input.to(device)
temporal_input = temporal_input.to(device)
target_input = torch.LongTensor(target_input).to(device)
pack1 = pack_padded_sequence(spatial_input,
length_input,
batch_first=True)
pack2 = pack_padded_sequence(temporal_input,
length_input,
batch_first=True)
score = model(pack1, pack2)
loss = criterion(score, target_input)
losses[phase] += loss.item() * target_input.shape[0]
if args.debug:
print(loss.item())
if training:
optimizer.zero_grad()
loss.backward()
optimizer.step()
else:
pred = torch.max(softmax(score), 1)[1].cpu()
corrects[phase] += torch.sum(
pred == target_input.cpu()).item()
end = time.time()
print('Epoch {:2} | '
'Train loss: {:.5f} Val loss: {:.5f} | '
'Test loss: {:.5f} accuracy: {:.5f} | '
'running time: {:.2f} sec'.format(
epoch,
losses['Train'] / len(data_loaders['Train'].dataset),
losses['Validation'] /
len(data_loaders['Validation'].dataset),
losses['Test'] / len(data_loaders['Test'].dataset),
corrects['Test'] / len(data_loaders['Test'].dataset),
end - start,
))
if epoch in args.test_intervals:
torch.save(
model.state_dict(),
osp.join(this_dir,
'./state_dict-epoch-' + str(epoch) + '.pth'))
# init model
if arch == "rnn":
model = RNN(**params_model, output_size=num_classes).to(device=device)
elif arch == "gru":
model = GRU(**params_model, output_size=num_classes).to(device=device)
elif arch == "lstm":
model = LSTM(**params_model, output_size=num_classes).to(device=device)
else:
raise Exception(
"Only 'rnn', 'gru', and 'lstm' are available model options.")
print("Model size: {0}".format(count_parameters(model)))
criterion = nn.NLLLoss()
op = torch.optim.SGD(model.parameters(), **params_op)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(op,
patience=4,
factor=0.5,
verbose=True)
# resume from checkpoint
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
best_acc1 = checkpoint['best_acc1']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
print("=> loaded checkpoint '{}' (epoch {})".format(
def main(argv):
global args
args = parser.parse_args(argv)
if args.threads == -1:
args.threads = torch.multiprocessing.cpu_count() - 1 or 1
print('===> Configuration')
print(args)
cuda = args.cuda
if cuda:
if torch.cuda.is_available():
print('===> {} GPUs are available'.format(
torch.cuda.device_count()))
else:
raise Exception("No GPU found, please run with --no-cuda")
# Fix the random seed for reproducibility
# random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if cuda:
torch.cuda.manual_seed(args.seed)
# Data loading
print('===> Loading entire datasets')
with open(args.data_path + 'train.seqs', 'rb') as f:
train_seqs = pickle.load(f)
with open(args.data_path + 'train.labels', 'rb') as f:
train_labels = pickle.load(f)
with open(args.data_path + 'valid.seqs', 'rb') as f:
valid_seqs = pickle.load(f)
with open(args.data_path + 'valid.labels', 'rb') as f:
valid_labels = pickle.load(f)
with open(args.data_path + 'test.seqs', 'rb') as f:
test_seqs = pickle.load(f)
with open(args.data_path + 'test.labels', 'rb') as f:
test_labels = pickle.load(f)
max_code = max(
map(lambda p: max(map(lambda v: max(v), p)),
train_seqs + valid_seqs + test_seqs))
num_features = max_code + 1
print(" ===> Construct train set")
train_set = VisitSequenceWithLabelDataset(train_seqs,
train_labels,
num_features,
reverse=False)
print(" ===> Construct validation set")
valid_set = VisitSequenceWithLabelDataset(valid_seqs,
valid_labels,
num_features,
reverse=False)
print(" ===> Construct test set")
test_set = VisitSequenceWithLabelDataset(test_seqs,
test_labels,
num_features,
reverse=False)
train_loader = DataLoader(dataset=train_set,
batch_size=args.batch_size,
shuffle=True,
collate_fn=visit_collate_fn,
num_workers=args.threads)
valid_loader = DataLoader(dataset=valid_set,
batch_size=args.eval_batch_size,
shuffle=False,
collate_fn=visit_collate_fn,
num_workers=args.threads)
test_loader = DataLoader(dataset=test_set,
batch_size=args.eval_batch_size,
shuffle=False,
collate_fn=visit_collate_fn,
num_workers=args.threads)
print('===> Dataset loaded!')
# Create model
print('===> Building a Model')
model = RNN(dim_input=num_features, dim_emb=128, dim_hidden=128)
if cuda:
model = model.cuda()
print(model)
print('===> Model built!')
weight_class0 = torch.mean(torch.FloatTensor(train_set.labels))
weight_class1 = 1.0 - weight_class0
weight = torch.FloatTensor([weight_class0, weight_class1])
criterion = nn.CrossEntropyLoss(weight=weight)
if args.cuda:
criterion = criterion.cuda()
optimizer = torch.optim.SGD(model.parameters(),
lr=args.lr,
momentum=args.momentum,
nesterov=False,
weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer, 'min')
best_valid_epoch = 0
best_valid_loss = sys.float_info.max
train_losses = []
valid_losses = []
if not os.path.exists(args.save):
os.makedirs(args.save)
for ei in trange(args.epochs, desc="Epochs"):
# Train
_, _, train_loss = rnn_epoch(train_loader,
model,
criterion=criterion,
optimizer=optimizer,
train=True)
train_losses.append(train_loss)
# Eval
_, _, valid_loss = rnn_epoch(valid_loader, model, criterion=criterion)
valid_losses.append(valid_loss)
scheduler.step(valid_loss)
is_best = valid_loss < best_valid_loss
if is_best:
best_valid_epoch = ei
best_valid_loss = valid_loss
# evaluate on the test set
test_y_true, test_y_pred, test_loss = rnn_epoch(
test_loader, model, criterion=criterion)
if args.cuda:
test_y_true = test_y_true.cpu()
test_y_pred = test_y_pred.cpu()
test_auc = roc_auc_score(test_y_true.numpy(),
test_y_pred.numpy()[:, 1],
average="weighted")
test_aupr = average_precision_score(test_y_true.numpy(),
test_y_pred.numpy()[:, 1],
average="weighted")
with open(args.save + 'train_result.txt', 'w') as f:
f.write('Best Validation Epoch: {}\n'.format(ei))
f.write('Best Validation Loss: {}\n'.format(valid_loss))
f.write('Train Loss: {}\n'.format(train_loss))
f.write('Test Loss: {}\n'.format(test_loss))
f.write('Test AUROC: {}\n'.format(test_auc))
f.write('Test AUPR: {}\n'.format(test_aupr))
torch.save(model, args.save + 'best_model.pth')
torch.save(model.state_dict(), args.save + 'best_model_params.pth')
# plot
if args.plot:
plt.figure(figsize=(12, 9))
plt.plot(np.arange(len(train_losses)),
np.array(train_losses),
label='Training Loss')
plt.plot(np.arange(len(valid_losses)),
np.array(valid_losses),
label='Validation Loss')
plt.xlabel('epoch')
plt.ylabel('Loss')
plt.legend(loc="best")
plt.tight_layout()
plt.savefig(args.save + 'loss_plot.eps', format='eps')
plt.close()
print('Best Validation Epoch: {}\n'.format(best_valid_epoch))
print('Best Validation Loss: {}\n'.format(best_valid_loss))
print('Train Loss: {}\n'.format(train_loss))
print('Test Loss: {}\n'.format(test_loss))
print('Test AUROC: {}\n'.format(test_auc))
print('Test AUPR: {}\n'.format(test_aupr))
if __name__ == "__main__":
train_iter, dev_iter, test_iter, TEXT, LABEL = load_iters(
batch_size, device, data_path, vectors)
vocab_size = len(TEXT.vocab.itos)
# build model
if use_rnn:
model = RNN(vocab_size, embed_size, hidden_size, num_layers,
num_classes, bidirectional, dropout_rate)
else:
model = CNN(vocab_size, embed_size, num_classes, num_filters,
kernel_sizes, dropout_rate)
if vectors is not None:
model.embed.from_pretrained(TEXT.vocab.vectors, freeze=freeze)
model.to(device)
optimizer = Adam(model.parameters(), lr=learning_rate)
loss_func = nn.CrossEntropyLoss()
writer = SummaryWriter('logs', comment="rnn")
for epoch in trange(train_epochs, desc="Epoch"):
model.train()
ep_loss = 0
for step, batch in enumerate(tqdm(train_iter, desc="Iteration")):
(inputs, lens), labels = batch.text, batch.label
outputs = model(inputs, lens)
loss = loss_func(outputs, labels)
ep_loss += loss.item()
model.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(model.parameters(), clip)
optimizer.step()
def train():
global_step = 0
# Loaded pretrained VAE
vae = VAE(hp.vsize).to(DEVICE)
ckpt = sorted(glob.glob(os.path.join(hp.ckpt_dir, 'vae',
'*k.pth.tar')))[-1]
vae_state = torch.load(ckpt)
vae.load_state_dict(vae_state['model'])
vae.eval()
print('Loaded vae ckpt {}'.format(ckpt))
rnn = RNN(hp.vsize, hp.asize, hp.rnn_hunits).to(DEVICE)
ckpts = sorted(glob.glob(os.path.join(hp.ckpt_dir, 'rnn', '*k.pth.tar')))
if ckpts:
ckpt = ckpts[-1]
rnn_state = torch.load(ckpt)
rnn.load_state_dict(rnn_state['model'])
global_step = int(os.path.basename(ckpt).split('.')[0][:-1]) * 1000
print('Loaded rnn ckpt {}'.format(ckpt))
data_path = hp.data_dir if not hp.extra else hp.extra_dir
# optimizer = torch.optim.RMSprop(rnn.parameters(), lr=1e-3)
optimizer = torch.optim.Adam(rnn.parameters(), lr=1e-4)
dataset = GameEpisodeDataset(data_path, seq_len=hp.seq_len)
loader = DataLoader(dataset,
batch_size=1,
shuffle=True,
drop_last=True,
num_workers=hp.n_workers,
collate_fn=collate_fn)
testset = GameEpisodeDataset(data_path, seq_len=hp.seq_len, training=False)
test_loader = DataLoader(testset,
batch_size=1,
shuffle=False,
drop_last=False,
collate_fn=collate_fn)
ckpt_dir = os.path.join(hp.ckpt_dir, 'rnn')
sample_dir = os.path.join(ckpt_dir, 'samples')
os.makedirs(sample_dir, exist_ok=True)
l1 = nn.L1Loss()
while global_step < hp.max_step:
# GO_states = torch.zeros([hp.batch_size, 1, hp.vsize+hp.asize]).to(DEVICE)
with tqdm(enumerate(loader), total=len(loader), ncols=70,
leave=False) as t:
t.set_description('Step {}'.format(global_step))
for idx, (obs, actions) in t:
obs, actions = obs.to(DEVICE), actions.to(DEVICE)
with torch.no_grad():
latent_mu, latent_var = vae.encoder(obs) # (B*T, vsize)
z = latent_mu
# z = vae.reparam(latent_mu, latent_var) # (B*T, vsize)
z = z.view(-1, hp.seq_len, hp.vsize) # (B*n_seq, T, vsize)
# import pdb; pdb.set_trace()
next_z = z[:, 1:, :]
z, actions = z[:, :-1, :], actions[:, :-1, :]
states = torch.cat([z, actions], dim=-1) # (B, T, vsize+asize)
# states = torch.cat([GO_states, next_states[:,:-1,:]], dim=1)
x, _, _ = rnn(states)
loss = l1(x, next_z)
optimizer.zero_grad()
loss.backward()
optimizer.step()
global_step += 1
if global_step % hp.log_interval == 0:
eval_loss = evaluate(test_loader, vae, rnn, global_step)
now = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
with open(os.path.join(ckpt_dir, 'train.log'), 'a') as f:
log = '{} || Step: {}, train_loss: {:.4f}, loss: {:.4f}\n'.format(
now, global_step, loss.item(), eval_loss)
f.write(log)
S = 2
y = vae.decoder(x[S, :, :])
v = vae.decoder(next_z[S, :, :])
save_image(
y,
os.path.join(sample_dir,
'{:04d}-rnn.png'.format(global_step)))
save_image(
v,
os.path.join(sample_dir,
'{:04d}-vae.png'.format(global_step)))
save_image(
obs[S:S + hp.seq_len - 1],
os.path.join(sample_dir,
'{:04d}-obs.png'.format(global_step)))
if global_step % hp.save_interval == 0:
d = {
'model': rnn.state_dict(),
'optimizer': optimizer.state_dict(),
}
torch.save(
d,
os.path.join(
ckpt_dir,
'{:03d}k.pth.tar'.format(global_step // 1000)))
# All 14 years of data
X_test_dep_std = X
X_train_dep_std = np.expand_dims(X_train_dep_std, axis=0)
y_train_dep_std = np.expand_dims(y_train_dep_std, axis=0)
X_test_dep_std = np.expand_dims(X_test_dep_std, axis=0)
# Transfer to Pytorch Variable
X_train_dep_std = Variable(torch.from_numpy(X_train_dep_std).float())
y_train_dep_std = Variable(torch.from_numpy(y_train_dep_std).float())
X_test_dep_std = Variable(torch.from_numpy(X_test_dep_std).float())
# Define rnn model
model = RNN(input_size=5, hidden_size=40, num_layers=2, class_size=1, dropout=0.5, rnn_type='lstm',dropout_bool=True)
# Define optimization function
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4) # optimize all rnn parameters
# Define loss function
loss_func = nn.MSELoss()
# Start training
for iter in range(10000):
model.train()
prediction = model(X_train_dep_std)
loss = loss_func(prediction, y_train_dep_std)
optimizer.zero_grad() # clear gradients for this training step
loss.backward() # back propagation, compute gradients
optimizer.step()
if iter % 100 == 0:
print("iteration: %s, loss: %s" % (iter, loss.item()))
# Save model
if args.freeze_embedding:
model.embedding.weight.requires_grad = False
model = model.to(device)
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)
myprint(f'The model has {count_parameters(model):,} trainable parameters')
if args.optimizer == "SGD":
print("Using SGD")
optimizer = optim.SGD(model.parameters(),
weight_decay=args.weight_decay,
lr=args.lr,
momentum=args.momentum,
nesterov=args.nesterov)
else:
print("Using Adam")
optimizer = optim.Adam(model.parameters(),
weight_decay=args.weight_decay,
lr=args.lr,
amsgrad=args.amsgrad)
criterion = nn.CrossEntropyLoss().to(device)
accuracy = categorical_accuracy
if args.task == "MIMIC-D":
accuracy = f1_score
train_set = PCDataset(DATA_ROOT, exclude_patterns=val_signer)
val_set = PCDataset(DATA_ROOT, exclude_patterns=signerlist)
print('-' * 80)
print(
f'[INFO] Training on {len(train_set)} samples from {len(signerlist)} signers'
)
print(
f'[INFO] Validating on {len(val_set)} samples from {len(val_signer)} signers'
)
train_loader = PCDataLoader(train_set)
val_loader = PCDataLoader(val_set)
model = RNN(60, 100, 30)
criterion = torch.nn.NLLLoss()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.9)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=10,
gamma=0.1)
_, history = train(model, train_loader, val_loader, criterion, optimizer,
scheduler, NUM_EPOCH, None, SAVE_PATH)
print(
f'[INFO] Training finished with best validation accuracy: {max(history["val_acc"]):.4f}'
)
print('-' * 80)
plot_history(history, SAVE_PATH, 'svg')
n_blocks=args.num_layers,
dropout=1. - args.dp_keep_prob)
# these 3 attributes don't affect the Transformer's computations;
# they are only used in run_epoch
model.batch_size = args.batch_size
model.seq_len = args.seq_len
model.vocab_size = vocab_size
else:
print("Model type not recognized.")
model = model.to(device)
# LOSS FUNCTION
loss_fn = nn.CrossEntropyLoss()
if args.optimizer == 'ADAM':
optimizer = torch.optim.Adam(model.parameters(), lr=args.initial_lr)
# LEARNING RATE SCHEDULE
lr = args.initial_lr
lr_decay_base = 1 / 1.15
m_flat_lr = 14.0 # we will not touch lr for the first m_flat_lr epochs
###############################################################################
#
# DEFINE COMPUTATIONS FOR PROCESSING ONE EPOCH
#
###############################################################################
def repackage_hidden(h):
"""
else:
# no embedding layer (one-hot encoding)
model = OneHotRNN(vocabulary=dataset.vocabulary,
rnn_type=args.rnn_type,
hidden_size=args.hidden_size,
n_layers=args.n_layers,
dropout=args.dropout,
bidirectional=args.bidirectional,
nonlinearity=args.nonlinearity)
# optionally, load model parameters from file
if args.pretrain_model is not None:
model.load_state_dict(torch.load(args.pretrain_model))
# set up optimizer
optimizer = optim.Adam(model.parameters(),
betas=(0.9, 0.999), ## default
eps=1e-08, ## default
lr=args.learning_rate)
# set up early stopping
early_stop = EarlyStopping(patience=args.patience)
# set up training schedule file
sched_filename = "training_schedule-" + str(args.sample_idx + 1) + ".csv"
sched_file = os.path.join(args.output_dir, sched_filename)
# iterate over epochs
counter = 0
for epoch in range(args.max_epochs):
# iterate over batches
print("Train Size: ", len(loader_train.sampler.indices))
print("Validation Size: ", len(loader_val.sampler.indices))
num_classes = len(train_set.label_encoder.classes_)
model = RNN(embeddings, num_classes=num_classes, **_hparams)
weights = class_weigths(train_set.labels)
if torch.cuda.is_available():
model.cuda()
weights = weights.cuda()
criterion = torch.nn.CrossEntropyLoss(weight=weights)
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = torch.optim.Adam(parameters, lr=lr)
#############################################################
# Train
#############################################################
best_val_loss = None
colms_l = ['Train_Loss', 'Val_Loss']
colms_acc = ['Train_Acc', 'Val_Acc']
colms_f1 = ['Train_F1', 'Val_F1']
df_l = pd.DataFrame(columns=colms_l, index=range(1, EPOCHS + 1))
df_acc = pd.DataFrame(columns=colms_acc, index=range(1, EPOCHS + 1))
df_f1 = pd.DataFrame(columns=colms_f1, index=range(1, EPOCHS + 1))
def main(args):
# hyperparameters
batch_size = args.batch_size
num_workers = 1
# Image Preprocessing
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
])
# load COCOs dataset
IMAGES_PATH = 'data/train2014'
CAPTION_FILE_PATH = 'data/annotations/captions_train2014.json'
vocab = load_vocab()
train_loader = get_coco_data_loader(path=IMAGES_PATH,
json=CAPTION_FILE_PATH,
vocab=vocab,
transform=transform,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
IMAGES_PATH = 'data/val2014'
CAPTION_FILE_PATH = 'data/annotations/captions_val2014.json'
val_loader = get_coco_data_loader(path=IMAGES_PATH,
json=CAPTION_FILE_PATH,
vocab=vocab,
transform=transform,
batch_size=batch_size,
shuffle=True,
num_workers=num_workers)
losses_val = []
losses_train = []
# Build the models
ngpu = 1
initial_step = initial_epoch = 0
embed_size = args.embed_size
num_hiddens = args.num_hidden
learning_rate = 1e-3
num_epochs = 3
log_step = args.log_step
save_step = 500
checkpoint_dir = args.checkpoint_dir
encoder = CNN(embed_size)
decoder = RNN(embed_size,
num_hiddens,
len(vocab),
1,
rec_unit=args.rec_unit)
# Loss
criterion = nn.CrossEntropyLoss()
if args.checkpoint_file:
encoder_state_dict, decoder_state_dict, optimizer, *meta = utils.load_models(
args.checkpoint_file, args.sample)
initial_step, initial_epoch, losses_train, losses_val = meta
encoder.load_state_dict(encoder_state_dict)
decoder.load_state_dict(decoder_state_dict)
else:
params = list(decoder.parameters()) + list(
encoder.linear.parameters()) + list(encoder.batchnorm.parameters())
optimizer = torch.optim.Adam(params, lr=learning_rate)
if torch.cuda.is_available():
encoder.cuda()
decoder.cuda()
if args.sample:
return utils.sample(encoder, decoder, vocab, val_loader)
# Train the Models
total_step = len(train_loader)
try:
for epoch in range(initial_epoch, num_epochs):
for step, (images, captions,
lengths) in enumerate(train_loader, start=initial_step):
# Set mini-batch dataset
images = utils.to_var(images, volatile=True)
captions = utils.to_var(captions)
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
# Forward, Backward and Optimize
decoder.zero_grad()
encoder.zero_grad()
if ngpu > 1:
# run on multiple GPU
features = nn.parallel.data_parallel(
encoder, images, range(ngpu))
outputs = nn.parallel.data_parallel(
decoder, features, range(ngpu))
else:
# run on single GPU
features = encoder(images)
outputs = decoder(features, captions, lengths)
train_loss = criterion(outputs, targets)
losses_train.append(train_loss.data[0])
train_loss.backward()
optimizer.step()
# Run validation set and predict
if step % log_step == 0:
encoder.batchnorm.eval()
# run validation set
batch_loss_val = []
for val_step, (images, captions,
lengths) in enumerate(val_loader):
images = utils.to_var(images, volatile=True)
captions = utils.to_var(captions, volatile=True)
targets = pack_padded_sequence(captions,
lengths,
batch_first=True)[0]
features = encoder(images)
outputs = decoder(features, captions, lengths)
val_loss = criterion(outputs, targets)
batch_loss_val.append(val_loss.data[0])
losses_val.append(np.mean(batch_loss_val))
# predict
sampled_ids = decoder.sample(features)
sampled_ids = sampled_ids.cpu().data.numpy()[0]
sentence = utils.convert_back_to_text(sampled_ids, vocab)
print('Sample:', sentence)
true_ids = captions.cpu().data.numpy()[0]
sentence = utils.convert_back_to_text(true_ids, vocab)
print('Target:', sentence)
print(
'Epoch: {} - Step: {} - Train Loss: {} - Eval Loss: {}'
.format(epoch, step, losses_train[-1], losses_val[-1]))
encoder.batchnorm.train()
# Save the models
if (step + 1) % save_step == 0:
utils.save_models(encoder, decoder, optimizer, step, epoch,
losses_train, losses_val, checkpoint_dir)
utils.dump_losses(
losses_train, losses_val,
os.path.join(checkpoint_dir, 'losses.pkl'))
except KeyboardInterrupt:
pass
finally:
# Do final save
utils.save_models(encoder, decoder, optimizer, step, epoch,
losses_train, losses_val, checkpoint_dir)
utils.dump_losses(losses_train, losses_val,
os.path.join(checkpoint_dir, 'losses.pkl'))
