def eval(tag_path, corpus_path):
correct = 0
total = 0
acc_list = []
model_name = MODEL_NAME
embedding_dim = EMBEDDING_DIM
hidden_dim = HIDDEN_DIM
word_to_ix = WORD_TO_IX
model = BiLSTM(len(word_to_ix), 5, embedding_dim, hidden_dim)
checkpoint = torch.load(model_name)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()
tag_to_ix = {'1': 0, '2': 1, '3': 2, '4': 3, '5': 4}
sentences, tags = load_train_data(tag_path, corpus_path)
labels = torch.tensor([[tag_to_ix[tag]] for tag in tags[:]])
with torch.no_grad():
for i, sen in enumerate(tqdm(sentences[:])):
input = prepare_sequence(sen, word_to_ix)
output = model(input)
_, predicted = torch.max(output.data, 1)
label = labels[i]
total += label.size(0)
correct += (predicted == label).sum().item()
acc = round(100 * correct / total, 2)
acc_list.append(acc)
assert len(acc_list) == len(sentences)
final_acc = acc
plt.plot(list(range(len(tags))), acc_list)
plt.xlabel('pred_num')
plt.ylabel('accuracy / %')
plt.show()
return final_acc
def get_time_to_score(tsv_path, thing, model_path):
time_to_count = {}
time_to_scoresum = {}
if thing == 'hair_dryer':
id = '732252283'
elif thing == 'microwave':
id = '423421857'
else:
id = '246038397'
with open('train_' + thing + '_word_to_ix.json', 'r') as j:
word_to_ix = json.load(j)
embedding_dim = EMBEDDING_DIM
hidden_dim = HIDDEN_DIM
model = BiLSTM(len(word_to_ix), 5, embedding_dim, hidden_dim)
checkpoints = torch.load(model_path)
model.load_state_dict(checkpoints['model_state_dict'])
model.eval()
with open(tsv_path, 'r') as f:
reader = csv.reader(f, delimiter='\t')
for i, r in enumerate(reader):
if i == 0 or r[4] != id:
continue
month, _, year = r[14].split('/')
if year not in {'2014', '2015'}:
continue
time = get_idx_by_year_month(int(year), int(month))
if time < 8:
continue
sen = (r[12] + ' ' + r[13]).lower()
sen = re.sub(r'[^A-Za-z0-9,.!]+', ' ', sen)
input = prepare_sequence(sen.split(), word_to_ix)
with torch.no_grad():
output = model(input)
_, predicted = torch.max(output.data, 1)
pred_score = predicted.item()
if time not in time_to_count:
time_to_count[time] = 0
time_to_scoresum[time] = 0.
time_to_count[time] += 1
time_to_scoresum[time] += pred_score
time_to_scoremean = {}
for time in time_to_count.keys():
time_to_scoremean[time] = time_to_scoresum[time] / time_to_count[time]
print(time_to_count)
return time_to_scoremean
class Trainer:
def __init__(self, config):
self.config = config
self.load_data() # 加载数据集
self.model = BiLSTM(self.config, self.vocab_size,
self.word_vectors) # 初始化模型
def load_data(self):
self.train_dataloader = TrainData(self.config)
self.eval_dataloader = TestData(self.config)
train_data_path = os.path.join(self.config.BASE_DIR,
self.config.train_data_path)
self.train_inputs, self.train_labels, self.t2ix = self.train_dataloader.gen_train_data(
train_data_path)
eval_data_path = os.path.join(self.config.BASE_DIR,
self.config.eval_data_path)
self.eval_inputs, self.eval_labels, _ = self.eval_dataloader.gen_test_data(
eval_data_path)
self.vocab_size = self.train_dataloader.vocab_size
self.word_vectors = self.train_dataloader.word_vectors
def train(self):
gpu_options = tf.GPUOptions(per_process_gpu_memory_fraction=0.9,
allow_growth=True)
sess_config = tf.ConfigProto(log_device_placement=False,
allow_soft_placement=True,
gpu_options=gpu_options)
with tf.Session(config=sess_config) as sess:
sess.run(tf.global_variables_initializer()) # 初始化计算图变量
current_step = 0
# 创建Train/Eval的summar路径和写入对象
train_summary_path = os.path.join(
self.config.BASE_DIR, self.config.summary_path + "/train")
eval_summary_path = os.path.join(
self.config.BASE_DIR, self.config.summary_path + "/eval")
self._check_directory(train_summary_path)
self._check_directory(eval_summary_path)
train_summary_writer = tf.summary.FileWriter(
train_summary_path, sess.graph)
eval_summary_writer = tf.summary.FileWriter(
eval_summary_path, sess.graph)
# Train & Eval Process
for epoch in range(self.config.epochs):
print(f"----- Epoch {epoch + 1}/{self.config.epochs} -----")
for batch in self.train_dataloader.next_batch(
self.train_inputs, self.train_labels,
self.config.batch_size):
summary, loss, predictions = self.model.train(
sess, batch, self.config.keep_prob)
accuracy = self.model.get_metrics(sess, batch)
train_summary_writer.add_summary(summary, current_step)
print(
f"! Train epoch: {epoch}, step: {current_step}, train loss: {loss}, accuracy: {accuracy}"
)
current_step += 1
if self.eval_dataloader and current_step % self.config.eval_every == 0:
losses = []
acces = []
for eval_batch in self.eval_dataloader.next_batch(
self.eval_inputs, self.eval_labels,
self.config.batch_size):
eval_summary, eval_loss, eval_predictions = self.model.eval(
sess, eval_batch)
eval_accuracy = self.model.get_metrics(sess, batch)
eval_summary_writer.add_summary(
eval_summary, current_step)
losses.append(eval_loss)
acces.append(eval_accuracy)
print(
f"! Eval epoch: {epoch}, step: {current_step}, eval loss: {sum(losses) / len(losses)}, accuracy: {sum(acces) / len(acces)}"
)
if self.config.ckpt_model_path:
save_path = os.path.join(
self.config.BASE_DIR,
self.config.ckpt_model_path)
self._check_directory(save_path)
model_save_path = os.path.join(
save_path, self.config.model_name)
self.model.saver.save(sess,
model_save_path,
global_step=current_step)
def _check_directory(self, path):
if not os.path.exists(path):
os.makedirs(path)
# create a iterator of the correct shape and type
iter = tf.data.Iterator.from_structure(train_batches.output_types, train_batches.output_shapes)
xs, ys = iter.get_next()
train_init_opt = iter.make_initializer(train_batches)
dev_init_opt = iter.make_initializer(dev_batches)
# index+=1
model = BiLSTM(param)
# print('xs')
# print(xs)
# print('ys')
# print(ys)
loss,train_opt,pred_train,train_summaries,global_step,lstm_cell_fw,x_check = model.train(xs,ys)
logits_eval,probs_eval,pred_eval,ys = model.eval(xs,ys)
#Variables for early stop
dev_history = []
dev_best = 0
stop_times = 0
logging.info('# Session')
saver = tf.train.Saver(max_to_keep=model.epoch)
with tf.Session() as sess:
ckpt = tf.train.latest_checkpoint(ckpt_path)
if ckpt is None:
logging.info("Initializing from scratch")
sess.run(tf.global_variables_initializer())
class Seq_MNIST_Trainer():
def __init__(self, trainer_params, args):
self.args = args
self.trainer_params = trainer_params
random.seed(trainer_params.random_seed)
torch.manual_seed(trainer_params.random_seed)
if args.cuda:
torch.cuda.manual_seed_all(trainer_params.random_seed)
kwargs = {'num_workers': 1, 'pin_memory': True} if args.cuda else {}
self.train_data = seq_mnist_train(trainer_params)
self.val_data = seq_mnist_val(trainer_params)
self.train_loader = DataLoader(self.train_data, batch_size=trainer_params.batch_size, shuffle=True, **kwargs)
self.val_loader = DataLoader(self.val_data, batch_size=trainer_params.test_batch_size, shuffle=True, **kwargs)
self.starting_epoch = 1
self.prev_loss = 10000
self.model = BiLSTM(trainer_params)
self.criterion = wp.CTCLoss(size_average=True)
self.labels = [i for i in range(trainer_params.num_classes-1)]
self.decoder = seq_mnist_decoder(labels=self.labels)
if args.resume or args.eval or args.export:
print("Loading model from {}".format(args.save_path))
package = torch.load(args.save_path, map_location=lambda storage, loc: storage)
self.model.load_state_dict(package['state_dict'])
if args.cuda:
torch.cuda.set_device(args.gpus)
self.model = self.model.cuda()
self.optimizer = optim.Adam(self.model.parameters(), lr=trainer_params.lr)
if args.resume:
self.optimizer.load_state_dict(package['optim_dict'])
self.starting_epoch = package['starting_epoch']
self.prev_loss = package['prev_loss']
if args.cuda:
for state in self.optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if args.init_bn_fc_fusion:
if not trainer_params.prefused_bn_fc:
self.model.batch_norm_fc.init_fusion()
self.trainer_params.prefused_bn_fc = True
else:
raise Exception("BN and FC are already fused.")
def serialize(self, model, trainer_params, optimizer, starting_epoch, prev_loss):
package = {'state_dict': model.state_dict(),
'trainer_params': trainer_params,
'optim_dict' : optimizer.state_dict(),
'starting_epoch' : starting_epoch,
'prev_loss': prev_loss
}
return package
def save_model(self, epoch, loss_value):
print("Model saved at: {}\n".format(self.args.save_path))
self.prev_loss = loss_value
torch.save(self.serialize(model=self.model, trainer_params=self.trainer_params,
optimizer=self.optimizer, starting_epoch=epoch + 1, prev_loss=self.prev_loss), self.args.save_path)
def train(self, epoch):
self.model.train()
for i, (item) in enumerate(self.train_loader):
data, labels, output_len, lab_len = item
data = Variable(data.transpose(1,0), requires_grad=False)
labels = Variable(labels.view(-1), requires_grad=False)
output_len = Variable(output_len.view(-1), requires_grad=False)
lab_len = Variable(lab_len.view(-1), requires_grad=False)
if self.args.cuda:
data = data.cuda()
output = self.model(data)
# print("Input = ", data.shape)
# print("model output (x) = ", output)
# print("GTs (y) = ", labels.type())
# print("model output len (xs) = ", output_len.type())
# print("GTs len (ys) = ", lab_len.type())
# exit(0)
loss = self.criterion(output, labels, output_len, lab_len)
loss_value = loss.data[0]
print("Loss value for epoch = {}/{} and batch {}/{} is = {:.4f}".format(epoch,
self.trainer_params.epochs, (i+1)*self.trainer_params.batch_size, len(self.train_data) , loss_value))
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
if self.args.cuda:
torch.cuda.synchronize()
def test(self, epoch=0, save_model_flag=False):
self.model.eval()
loss_value = 0
for i, (item) in enumerate(self.val_loader):
data, labels, output_len, lab_len = item
data = Variable(data.transpose(1,0), requires_grad=False)
labels = Variable(labels.view(-1), requires_grad=False)
output_len = Variable(output_len.view(-1), requires_grad=False)
lab_len = Variable(lab_len.view(-1), requires_grad=False)
if self.args.cuda:
data = data.cuda()
output = self.model(data)
# print("Input = ", data)
# print("model output (x) = ", output.shape)
# print("model output (x) = ", output)
# print("Label = ", labels)
# print("model output len (xs) = ", output_len)
# print("GTs len (ys) = ", lab_len)
index = random.randint(0,self.trainer_params.test_batch_size-1)
label = labels[index*self.trainer_params.word_size:(index+1)*self.trainer_params.word_size].data.numpy()
label = label-1
prediction = self.decoder.decode(output[:,index,:], output_len[index], lab_len[index])
accuracy = self.decoder.hit(prediction, label)
print("Sample Label = {}".format(self.decoder.to_string(label)))
print("Sample Prediction = {}".format(self.decoder.to_string(prediction)))
print("Accuracy on Sample = {:.2f}%\n\n".format(accuracy))
loss = self.criterion(output, labels, output_len, lab_len)
loss_value += loss.data.numpy()
loss_value /= (len(self.val_data)//self.trainer_params.test_batch_size)
print("Average Loss Value for Val Data is = {:.4f}\n".format(float(loss_value)))
if loss_value < self.prev_loss and save_model_flag:
self.save_model(epoch, loss_value)
def eval_model(self):
self.test()
def train_model(self):
for epoch in range(self.starting_epoch, self.trainer_params.epochs + 1):
self.train(epoch)
self.test(epoch=epoch, save_model_flag=True)
if epoch%20==0:
self.optimizer.param_groups[0]['lr'] = self.optimizer.param_groups[0]['lr']*0.98
def export_model(self, simd_factor, pe):
self.model.eval()
self.model.export('r_model_fw_bw.hpp', simd_factor, pe)
def export_image(self, idx=100):
img, label = self.val_data.images[:,idx,:], self.val_data.labels[0][idx]
img = img.transpose(1, 0)
label -= 1
label = self.decoder.to_string(label)
from PIL import Image
from matplotlib import cm
im = Image.fromarray(np.uint8(cm.gist_earth(img)*255))
im.save('test_image.png')
img = img.transpose(1, 0)
img = np.reshape(img, (-1, 1))
np.savetxt("test_image.txt", img, fmt='%.10f')
f = open('test_image_gt.txt','w')
f.write(label)
f.close()
print("Exported image with label = {}".format(label))
def main(options):
use_cuda = (len(options.gpuid) >= 1)
if options.gpuid:
cuda.set_device(options.gpuid[0])
train, dev, test, vocab = torch.load(open(options.data_file, 'rb'),
pickle_module=dill)
batched_train, batched_train_mask, _ = utils.tensor.advanced_batchize(
train, options.batch_size, vocab.stoi["<pad>"])
batched_dev, batched_dev_mask, _ = utils.tensor.advanced_batchize(
dev, options.batch_size, vocab.stoi["<pad>"])
vocab_size = len(vocab)
if options.load_file:
rnnlm = torch.load(options.load_file)
else:
rnnlm = BiLSTM(vocab_size)
if use_cuda > 0:
rnnlm.cuda()
else:
rnnlm.cpu()
criterion = torch.nn.NLLLoss()
optimizer = eval("torch.optim." + options.optimizer)(rnnlm.parameters(),
options.learning_rate)
# main training loop
last_dev_avg_loss = float("inf")
rnnlm.train()
for epoch_i in range(options.epochs):
logging.info("At {0}-th epoch.".format(epoch_i))
# srange generates a lazy sequence of shuffled range
for i, batch_i in enumerate(utils.rand.srange(len(batched_train))):
train_batch = Variable(
batched_train[batch_i]) # of size (seq_len, batch_size)
train_mask = Variable(batched_train_mask[batch_i])
if use_cuda:
train_batch = train_batch.cuda()
train_mask = train_mask.cuda()
sys_out_batch = rnnlm(
train_batch
) # (seq_len, batch_size, vocab_size) # TODO: substitute this with your module
train_in_mask = train_mask.view(-1)
train_in_mask = train_in_mask.unsqueeze(1).expand(
len(train_in_mask), vocab_size)
train_out_mask = train_mask.view(-1)
sys_out_batch = sys_out_batch.view(-1, vocab_size)
train_out_batch = train_batch.view(-1)
sys_out_batch = sys_out_batch.masked_select(train_in_mask).view(
-1, vocab_size)
train_out_batch = train_out_batch.masked_select(train_out_mask)
loss = criterion(sys_out_batch, train_out_batch)
logging.debug("loss at batch {0}: {1}".format(i, loss.data[0]))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# validation -- this is a crude esitmation because there might be some paddings at the end
dev_loss = 0.0
rnnlm.eval()
for batch_i in range(len(batched_dev)):
dev_batch = Variable(batched_dev[batch_i], volatile=True)
dev_mask = Variable(batched_dev_mask[batch_i], volatile=True)
if use_cuda:
dev_batch = dev_batch.cuda()
dev_mask = dev_mask.cuda()
sys_out_batch = rnnlm(dev_batch)
dev_in_mask = dev_mask.view(-1)
dev_in_mask = dev_in_mask.unsqueeze(1).expand(
len(dev_in_mask), vocab_size)
dev_out_mask = dev_mask.view(-1)
sys_out_batch = sys_out_batch.view(-1, vocab_size)
dev_out_batch = dev_batch.view(-1)
sys_out_batch = sys_out_batch.masked_select(dev_in_mask).view(
-1, vocab_size)
dev_out_batch = dev_out_batch.masked_select(dev_out_mask)
loss = criterion(sys_out_batch, dev_out_batch)
dev_loss += loss
dev_avg_loss = dev_loss / len(batched_dev)
logging.info(
"Average loss value per instance is {0} at the end of epoch {1}".
format(dev_avg_loss.data[0], epoch_i))
#if (last_dev_avg_loss - dev_avg_loss).data[0] < options.estop:
# logging.info("Early stopping triggered with threshold {0} (previous dev loss: {1}, current: {2})".format(epoch_i, last_dev_avg_loss.data[0], dev_avg_loss.data[0]))
# break
torch.save(
rnnlm,
open(
options.model_file +
".nll_{0:.2f}.epoch_{1}".format(dev_avg_loss.data[0], epoch_i),
'wb'),
pickle_module=dill)
last_dev_avg_loss = dev_avg_loss
import torchvision
from model import BiLSTM
from data import load_dataset
from config import model_name, device
if __name__ == "__main__":
# the string to test!
test_string = "<s> john can"
# ########################
# LOAD DATASET
# ########################
corpus, word_to_idx, idx_to_word, train_dataset = load_dataset()
# ########################
# TEST VARIABLES
# ########################
model = BiLSTM(len(corpus))
model.load_state_dict(torch.load(model_name))
model.eval()
sentence = test_string.split()
sentence = torch.tensor([[word_to_idx[w] for w in sentence]])
s = model.sample(sentence)
print(test_string.split() + s)
dev_batch_sizes[fold],
cfg.num_classes,
index=str(fold),
shuffle=False)
# create a iterator of the correct shape and type
iter = tf.data.Iterator.from_structure(test_batches.output_types,
test_batches.output_shapes)
xs, ys = iter.get_next()
test_init_opt = iter.make_initializer(test_batches)
dev_init_opt = iter.make_initializer(dev_batches)
logging.info('# load Model')
model = BiLSTM(param)
logits_test, probs_test, pred_test, ys = model.eval(xs, ys)
logits_dev, probs_dev, pred_dev, ys = model.eval(xs, ys)
logging.info('# Session')
with tf.Session() as sess:
model_path = os.path.join(cfg.result_path, cfg.version,
'index' + str(index) + '_models',
parameter_string, str(fold))
ckpt = tf.train.latest_checkpoint(model_path)
saver = tf.train.Saver()
saver.restore(sess, ckpt)
sess.run(dev_init_opt) # begin iterate dev data
all_probs_dev, dev_labels, accuracy_dev = evaluate_data(
'Dev', dev_labels, parameter_string, fold, all_probs_dev,