def main():
Config = config.get_args()
set_seed(Config.seed)
word2ix, ix2word, max_len, avg_len = build_word_dict(Config.train_path)
test_data = CommentDataSet(Config.test_path, word2ix, ix2word)
test_loader = DataLoader(
test_data,
batch_size=16,
shuffle=False,
num_workers=0,
collate_fn=mycollate_fn,
)
weight = torch.zeros(len(word2ix), Config.embedding_dim)
model = SentimentModel(embedding_dim=Config.embedding_dim,
hidden_dim=Config.hidden_dim,
LSTM_layers=Config.LSTM_layers,
drop_prob=Config.drop_prob,
pre_weight=weight)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# device = torch.device("cpu")
criterion = nn.CrossEntropyLoss()
model.load_state_dict(torch.load(Config.model_save_path),
strict=True) # 模型加载
confuse_meter = ConfuseMeter()
confuse_meter = test(test_loader, device, model, criterion)
def main():
Config = config.get_args()
set_seed(Config.seed)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
word2ix, ix2word, max_len, avg_len = build_word_dict(Config.train_path)
weight = torch.zeros(len(word2ix), Config.embedding_dim)
model = SentimentModel(embedding_dim=Config.embedding_dim,
hidden_dim=Config.hidden_dim,
LSTM_layers=Config.LSTM_layers,
drop_prob=Config.drop_prob,
pre_weight=weight)
model.load_state_dict(torch.load(Config.model_save_path),
strict=True) # 模型加载
# comment_str = "忘不掉的一句台词，是杜邦公司笑着对男主说：“Sue me”。我记得前段时间某件事，也是同样的说辞，“欢迎来起诉中华有为”。也是同样的跋扈。若干年后，会看到改编的电影吗。"
result = predict(Config.comment_str, model, device, word2ix)
print(Config.comment_str, result)
def create_model(log_dir=None, val_tuple=None, vocab=None):
return SentimentModel(
config={
'vocab_size': MAX_WORDS,
'log_dir': log_dir,
'tensorboard': {
'vocab': vocab,
'val_tuple': val_tuple,
},
})
def test(args):
if args.device == "gpu":
cfg_proto = tf.ConfigProto(intra_op_parallelism_threads=2)
else:
cfg_proto = None
with tf.Session(config=cfg_proto) as sess:
# Loading the vocabulary files
vocab, rev_vocab = load_vocab(args)
args.vocab_size = len(rev_vocab)
# Creating test model
with tf.variable_scope("model", reuse=None):
model_test = SentimentModel(args, None, mode='eval')
# Reload model from checkpoints, if any
steps_done = initialize_weights(sess, model_test, args, mode='test')
logger.info("loaded %d completed steps", steps_done)
test_set = load_eval_data(args, split='test')
correct, incorrect, losses = evaluate(sess, model_test, test_set, args)
with open(os.path.join(args.train_dir, 'incorrect.txt'), 'w') as f:
f.write(str(incorrect))
percent_correct = float(correct) * 100.0 / len(test_set)
logger.info("Correct Predictions - %.4f. Eval Losses - %.4f",
percent_correct, losses)
from model import SentimentModel
from transformers import AutoTokenizer, AutoModel
import pandas as pd
import csv
import torch
s = ["Negative", "Random", "Positive"]
bert_model = AutoModel.from_pretrained('google/bert_uncased_L-4_H-256_A-4')
bert_tokenizer = AutoTokenizer.from_pretrained(
'google/bert_uncased_L-4_H-256_A-4')
model = SentimentModel(bert_model)
model.freeze_weights()
model.lstm.load_state_dict(torch.load('rnn.pth'))
model.layers.load_state_dict(torch.load('checkpoint.pth'))
model.eval()
f = open('result.csv', 'w', newline='')
out_file = csv.writer(f)
data = pd.read_csv('results_ocr.csv')
out_file.writerow(['Filename', 'Category'])
for idx, row in data.iterrows():
if type(row['Text']) == float:
out_file.writerow([row['Filename'], "Random"])
else:
scores = model([row['Text']], bert_tokenizer)
sentiment = torch.argmax(torch.exp(scores), dim=1)
out_file.writerow([row['Filename'], s[sentiment]])
import nltk, re
from nltk.corpus import stopwords
from nltk.stem import SnowballStemmer
from math import ceil
import config
#from tqdm import tqdm
# create and configure the app
app = Flask(__name__, instance_relative_config=True)
app.config.from_object("config.Config")
filename = app.config['MODEL']
max_length = app.config['SENTENCE_MAX_LENGTH']
embedding_dim = app.config['EMBEDDING_DIM']
vacab_size = app.config['VOCABULARY_SIZE']
model = SentimentModel(embedding_dim, vacab_size, max_length)
model.load_weights(filename)
def preprocess(text, stem=False):
stemmer = SnowballStemmer('english')
text_cleaning_re = "@\S+|https?:\S+|http?:\S|[^A-Za-z0-9]+"
text = re.sub(text_cleaning_re, ' ', str(text).lower()).strip()
tokens = []
stop_words = stopwords.words('english')
for token in text.split():
if token not in stop_words:
if stem:
tokens.append(stemmer.stem(token))
import torch.nn as nn
from data import load_file
from model import SentimentModel
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
TEXT, LABEL, train, valid, test, train_iter, valid_iter, test_iter = load_file(filepath='data/',
device=device)
INPUT_DIM = len(TEXT.vocab)
EMBEDDING_DIM = 100
HIDDEN_DIM = 256
OUTPUT_DIM = 1
model = SentimentModel(INPUT_DIM, EMBEDDING_DIM, HIDDEN_DIM, OUTPUT_DIM)
optimizer = torch.optim.SGD(model.parameters(), lr=3e-3)
criterion = nn.BCEWithLogitsLoss()
model = model.to(device)
criterion = criterion.to(device)
def binary_accuracy(preds, y):
rounded_preds = torch.round(torch.sigmoid(preds))
correct = (rounded_preds == y).float()
acc = correct.sum() / len(correct)
return acc
def run(fold=0):
# kfold type of data input
data = pd.read_csv(config.TRAIN_FOLDS_FILE)
df_train = data[data['kfold'] != fold].reset_index(drop=True)
df_valid = data[data['kfold'] == fold].reset_index(drop=True)
train_data = CommentData(comments=df_train['Comment'],
labels=df_train['Label_encoded'],
sentiments=df_train['Sentiment_encoded'])
train_dataloader = torch.utils.data.DataLoader(
train_data,
batch_size=config.TRAIN_BATCH_SIZE,
# num_workers = 4
)
valid_data = CommentData(comments=df_valid['Comment'],
labels=df_valid['Label_encoded'],
sentiments=df_valid['Sentiment_encoded'])
valid_dataloader = torch.utils.data.DataLoader(
valid_data,
batch_size=config.VALID_BATCH_SIZE,
# num_workers = 4
)
device = torch.device('cuda')
model_config = RobertaConfig.from_pretrained(config.ROBERTA_PATH)
model_config.output_hidden_states = True
model = SentimentModel(model_config, config.OUTPUT_SIZE)
model.to(device)
param_optimizer = list(model.named_parameters())
no_decay = ['bias', 'LayerNorm.bias', 'LayerNorm.weight']
optimizer_parameters = [{
'params':
[p for n, p in param_optimizer if not any(nd in n for nd in no_decay)],
'weight_decay':
0.001
}, {
'params':
[p for n, p in param_optimizer if any(nd in n for nd in no_decay)],
'weight_decay':
0.01
}]
num_train_steps = int(
len(df_train) / config.TRAIN_BATCH_SIZE * config.EPOCHS)
optimizer = AdamW(optimizer_parameters, lr=3e-5)
scheduler = get_linear_schedule_with_warmup(
optimizer, num_warmup_steps=0, num_training_steps=num_train_steps)
# train_fn(data_loader, model, device, optimizer, scheduler=None)
train_loss_rec = []
eval_loss_rec = []
early_stopping = utils.EarlyStopping(patience=5, mode='min')
for epoch in range(config.EPOCHS):
print(f'########### fold = {fold} epoch = {epoch} ############')
loss_train = engine.train_fn(data_loader=train_dataloader,
model=model,
device=device,
optimizer=optimizer,
scheduler=scheduler)
train_loss_rec.append(loss_train)
losses_eval = engine.eval_fn(valid_dataloader, model, device)
eval_loss_rec.append(losses_eval)
print(f'train_loss = {loss_train} eval_loss = {losses_eval}')
# print(f'save model_{fold}.bin')
# torch.save(model.state_dict(), config.OUTPUT_PATH + f'/model_{fold}.bin')
early_stopping(losses_eval,
model,
model_path=config.OUTPUT_PATH +
f'/model_label_{fold}.bin')
if early_stopping.early_stop:
print('Early stopping')
break
def train(args):
max_epochs = args.config.max_epochs
batch_size = args.config.batch_size
if args.device == "gpu":
cfg_proto = tf.ConfigProto(intra_op_parallelism_threads=2)
cfg_proto.gpu_options.allow_growth = True
else:
cfg_proto = None
with tf.Session(config=cfg_proto) as sess:
# Loading the vocabulary files
vocab, rev_vocab = load_vocab(args)
args.vocab_size = len(rev_vocab)
# Loading all the training data
train_files, training_size = load_train_data(args)
queue = tf.train.string_input_producer(train_files,
num_epochs=max_epochs,
shuffle=True)
# Creating training model
with tf.variable_scope("model", reuse=None):
model = SentimentModel(args, queue, mode='train')
# Reload model from checkpoints, if any
steps_done = initialize_weights(sess, model, args, mode='train')
logger.info("loaded %d completed steps", steps_done)
# Load the w2v embeddings
if steps_done == 0 and args.config.cnn_mode != 'rand':
w2v_array = load_w2v(args, rev_vocab)
initialize_w2v(sess, model, w2v_array)
# Reusing weights for evaluation model
with tf.variable_scope("model", reuse=True):
model_eval = SentimentModel(args, None, mode='eval')
dev_set = load_eval_data(args, split='dev')
# This need not be zero due to incomplete runs
epoch = model.epoch.eval()
remaining_examples = training_size * max_epochs - (
model.global_step.eval() * batch_size)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
percent_best = 0.0
while epoch < max_epochs:
logger.info("Epochs done - %d", epoch)
frac_num_batches = float(remaining_examples) / (max_epochs -
epoch) / batch_size
if epoch == max_epochs - 1:
# Last batch may have some extra elements
num_batches = math.ceil(frac_num_batches)
else:
num_batches = round(frac_num_batches)
num_batches = int(num_batches)
logger.info(
"%d remaining examples, %d epochs left, %.4f fractional number of batches, %d chosen",
remaining_examples, max_epochs - epoch, frac_num_batches,
num_batches)
remaining_examples -= num_batches * batch_size
epoch_start = time.time()
if coord.should_stop():
break
for i in range(1, num_batches + 1):
output_feed = [model.updates, model.clip, model.losses]
_, _, losses = sess.run(output_feed)
if i % 100 == 0:
logger.info(
"minibatches done %d. Training Loss %.4f. Time elapsed in epoch %.4f.",
i, losses, (time.time() - epoch_start) / 3600.0)
if i % args.config.eval_frequency == 0 or i == num_batches:
logger.info("Evaluating model after %d minibatches", i)
correct, _, losses = evaluate(sess, model_eval, dev_set,
args)
percent_correct = float(correct) * 100.0 / len(dev_set)
logger.info("Correct Predictions - %.4f. Eval Loss - %.4f",
percent_correct, losses)
if percent_correct > percent_best:
percent_best = percent_correct
logger.info("Saving Best Model")
checkpoint_path = os.path.join(args.best_dir,
"sentence.ckpt")
model.best_saver.save(sess,
checkpoint_path,
global_step=model.global_step,
write_meta_graph=False)
# Also save the model for continuing in future
checkpoint_path = os.path.join(args.train_dir,
"sentence.ckpt")
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step,
write_meta_graph=False)
# Update epoch counter
sess.run(model.epoch_incr)
epoch += 1
checkpoint_path = os.path.join(args.train_dir, "sentence.ckpt")
model.saver.save(sess,
checkpoint_path,
global_step=model.global_step,
write_meta_graph=False)
coord.request_stop()
coord.join(threads)
# preprocess and save word encodings
preprocessor = Preprocessor(max_vocab=args.max_vocab)
data = preprocessor.fit_transform(dataset=data)
preprocessor.save(args.prepro_save_path)
# validation split
data.split_data(validation_count=args.validation_count)
train_ds, val_ds = data.to_dataset()
# to dataLoaders
train_set = DataLoader(train_ds, batch_size=args.batch_size, shuffle=True)
val_set = DataLoader(val_ds, batch_size=args.batch_size, shuffle=False)
print('Initializing model...')
mod = SentimentModel(
len(preprocessor.vocab2enc) + 3, args.embedding_dim, args.hidden_dim)
opt = Adam(mod.parameters(), lr=args.lr)
print('Training...')
fit(training=train_set,
model=mod,
validation=val_set,
optimizer=opt,
loss=torch.nn.BCELoss(),
epochs=args.epochs)
# Saving model
print('Saving model...')
torch.save(mod, args.model_save_path)
print('Done!')
def predict():
# kfold type of data input
data = pd.read_csv(config.TEST_FILE)
data['Label_encoded'] = 0
data['Sentiment_encoded'] = 0
df_test = data
test_data = CommentData(comments=df_test['Comment'],
labels=df_test['Label_encoded'],
sentiments=df_test['Sentiment_encoded'])
test_dataloader = torch.utils.data.DataLoader(
test_data,
batch_size=config.TEST_BATCH_SIZE,
# num_workers = 4
)
# model
device = torch.device('cuda')
model_config = BertConfig.from_pretrained(config.BERT_PATH)
model_config.output_hidden_states = True
model0 = SentimentModel(model_config, config.OUTPUT_SIZE_SENTIMENT)
model0.to(device)
# model0 = nn.DataParallel(model0)
model0.load_state_dict(torch.load(config.SAVED_MODEL_PATH +
'/model_0.bin'))
model0.eval()
model1 = SentimentModel(model_config, config.OUTPUT_SIZE_SENTIMENT)
model1.to(device)
# model1 = nn.DataParallel(model1)
model1.load_state_dict(torch.load(config.SAVED_MODEL_PATH +
'/model_1.bin'))
model1.eval()
model2 = SentimentModel(model_config, config.OUTPUT_SIZE_SENTIMENT)
model2.to(device)
# model2 = nn.DataParallel(model2)
model2.load_state_dict(torch.load(config.SAVED_MODEL_PATH +
'/model_2.bin'))
model2.eval()
model3 = SentimentModel(model_config, config.OUTPUT_SIZE_SENTIMENT)
model3.to(device)
# model3 = nn.DataParallel(model3)
model3.load_state_dict(torch.load(config.SAVED_MODEL_PATH +
'/model_3.bin'))
model3.eval()
model4 = SentimentModel(model_config, config.OUTPUT_SIZE_SENTIMENT)
model4.to(device)
# model4 = nn.DataParallel(model4)
model4.load_state_dict(torch.load(config.SAVED_MODEL_PATH +
'/model_4.bin'))
model4.eval()
# process raw output
model_prediction = []
with torch.no_grad():
tq0 = tqdm(test_dataloader, total=len(test_dataloader))
for bi, data in tqdm(enumerate(tq0)):
# load data / ready to input
input_ids = data['input_ids']
token_type_ids = data['token_type_ids']
attention_mask = data['attention_mask']
label = data['label']
sentiment = data['sentiment']
# prepare input data
input_ids = input_ids.to(device, dtype=torch.long)
token_type_ids = token_type_ids.to(device, dtype=torch.long)
attention_mask = attention_mask.to(device, dtype=torch.long)
label = label.to(device, dtype=torch.long)
sentiment = sentiment.to(device, dtype=torch.long)
# forward(self, ids, mask, type_ids)
out0 = model0(ids=input_ids,
mask=attention_mask,
type_ids=token_type_ids)
out1 = model1(ids=input_ids,
mask=attention_mask,
type_ids=token_type_ids)
out2 = model2(ids=input_ids,
mask=attention_mask,
type_ids=token_type_ids)
out3 = model3(ids=input_ids,
mask=attention_mask,
type_ids=token_type_ids)
out4 = model4(ids=input_ids,
mask=attention_mask,
type_ids=token_type_ids)
out = (out0 + out1 + out2 + out3 + out4) / 5
out = torch.softmax(out, dim=1).cpu().detach().numpy()
for ix, result in enumerate(out):
pred = np.argmax(result)
model_prediction.append(pred)
sample = pd.read_csv(config.TEST_FILE)
sample['sentiment_pred'] = model_prediction
sample.to_csv(config.OUTPUT_PATH + '/pred_sentiment.csv', index=False)
def main():
Config = config.get_args()
set_seed(Config.seed)
word2ix, ix2word, max_len, avg_len = build_word_dict(Config.train_path)
train_data = CommentDataSet(Config.train_path, word2ix, ix2word)
train_loader = DataLoader(
train_data,
batch_size=16,
shuffle=True,
num_workers=0,
collate_fn=mycollate_fn,
)
validation_data = CommentDataSet(Config.validation_path, word2ix, ix2word)
validation_loader = DataLoader(
validation_data,
batch_size=16,
shuffle=True,
num_workers=0,
collate_fn=mycollate_fn,
)
test_data = CommentDataSet(Config.test_path, word2ix, ix2word)
test_loader = DataLoader(
test_data,
batch_size=16,
shuffle=False,
num_workers=0,
collate_fn=mycollate_fn,
)
weight = pre_weight(len(word2ix), Config.pred_word2vec_path,
Config.embedding_dim, word2ix, ix2word)
model = SentimentModel(embedding_dim=Config.embedding_dim,
hidden_dim=Config.hidden_dim,
LSTM_layers=Config.LSTM_layers,
drop_prob=Config.drop_prob,
pre_weight=weight)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
optimizer = optim.Adam(model.parameters(), lr=Config.lr)
scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
step_size=10,
gamma=0.1) # 学习率调整
criterion = nn.CrossEntropyLoss()
# 因为使用tensorboard画图会产生很多日志文件，这里进行清空操作
if os.path.exists(Config.tensorboard_path):
shutil.rmtree(Config.tensorboard_path)
os.mkdir(Config.tensorboard_path)
for epoch in range(Config.epochs):
train_loader = tqdm(train_loader)
train_loader.set_description(
'[%s%04d/%04d %s%f]' %
('Epoch:', epoch + 1, Config.epochs, 'lr:', scheduler.get_lr()[0]))
train(epoch, Config.epochs, train_loader, device, model, criterion,
optimizer, scheduler, Config.tensorboard_path)
validate(epoch, validation_loader, device, model, criterion,
Config.tensorboard_path)
# 模型保存
if os.path.exists(Config.model_save_path) == False:
os.mkdir('./modelDict/')
torch.save(model.state_dict(), Config.model_save_path)
confuse_meter = ConfuseMeter()
confuse_meter = test(test_loader, device, model, criterion)
test_data = batched_dataset.take(TEST_BATCHES)
train_data = batched_dataset.skip(TEST_BATCHES)
VOCAB_LENGTH = len(tokenizer.vocab)
EMB_DIM = 200
CNN_FILTERS = 100
DNN_UNITS = 256
OUTPUT_CLASSES = 6
DROPOUT_RATE = 0.2
NB_EPOCHS = 10
model = SentimentModel(vocabulary_size=VOCAB_LENGTH,
embedding_dimensions=EMB_DIM,
cnn_filters=CNN_FILTERS,
dnn_units=DNN_UNITS,
model_output_classes=OUTPUT_CLASSES,
dropout_rate=DROPOUT_RATE)
if OUTPUT_CLASSES == 2:
model.compile(loss="binary_crossentropy",
optimizer="adam",
metrics=["accuracy"])
else:
model.compile(loss="sparse_categorical_crossentropy",
optimizer="adam",
metrics=["sparse_categorical_accuracy"])
model.fit(train_data, epochs=NB_EPOCHS)
# -------------------------------------------------------------------------------
def preprocess_and_train():
# read dataset
data = pd.read_csv('./training.1600000.processed.noemoticon.csv', encoding='latin-1', header=None)
data.columns = ('target','uid', 'time', 'query', 'user', 'text')
# create new dataframe
sent_df = pd.DataFrame(None, columns=('target', 'text'))
sent_df['target'] = data['target']
sent_df['text'] = data['text'].apply(preprocess_text)
sent_df['tweet_size'] = data['text'].apply(lambda x:len(x.split()))
# select random sample of 400,000 tweets from total dataset (training on a smaller dataset)
sent_df_sample = sent_df[(sent_df['tweet_size']>10) & (sent_df['target']==0)].sample(n=200000, random_state=SentConfig.SEED)
sent_df_sample = sent_df_sample.append(sent_df[(sent_df['tweet_size']>10) & (sent_df['target']==4)].sample(n=200000, random_state=SentConfig.SEED))
# split dataset into train, test, validation set
train, test = train_test_split(sent_df_sample, test_size=0.1)
train, val = train_test_split(train, test_size=0.05)
# create necessary dataloaders, for advantage of batching by pytorch
train_dl = SentimentDL(train)
val_dl = SentimentDL(val)
test_dl = SentimentDL(test)
train_loader = DataLoader(train_dl, batch_size=SentConfig.TRAIN_BATCH_SIZE, shuffle=True)
validation_loader = DataLoader(val_dl, batch_size=SentConfig.VALID_BATCH_SIZE, shuffle=True)
test_loader = DataLoader(test_dl, batch_size=SentConfig.VALID_BATCH_SIZE, shuffle=True)
# select the cuda device if available
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
# create model object
model = SentimentModel()
model.to(device)
# ready with optimizer and scheduler objects
# do not apply weight decay in AdamW to, bias layer and normalization terms
no_decay = ['bias', 'LayerNorm.weight', 'LayerNorm.bias'] # taken from https://huggingface.co/transformers/training.html
# more named parameteres in model.named_parameters()
optimizer_grouped_parameters = [
{'params': [p for n, p in model.named_parameters() if not any(nd in n for nd in no_decay)], 'weight_decay': 0.01},
{'params': [p for n, p in model.named_parameters() if any(nd in n for nd in no_decay)], 'weight_decay': 0.0}
]
# optim = AdamW(model.parameters(), lr=5e-5)
optim = AdamW(optimizer_grouped_parameters, lr=5e-5)
# learning rate scheduling
num_train_steps = int((train_dl.__len__()/SentConfig.TRAIN_BATCH_SIZE)*SentConfig.EPOCHS)
num_warmup_steps = int(0.05*num_train_steps)
scheduler = get_cosine_schedule_with_warmup(optim, num_warmup_steps, num_train_steps)
# Training : done on the basis of attaining better F1 score on the validation dataset
scores = []
min_f1 = 0
for epoch in range(SentConfig.EPOCHS):
_ = train_function(train_loader, model, optim, scheduler, device)
_, results = evaluation_function(validation_loader, model, device)
validation_f1 = round(f1_score(results[:,1], results[:,0]),4)
accuracy = round(accuracy_score(results[:,1], results[:,0]), 4)
scores.append((validation_f1, accuracy))
print('epoch num: ', epoch, 'f1 score: ',validation_f1 , 'accuracy: ', accuracy)
if validation_f1 > min_f1:
# save model if validation f1 score is
torch.save(model.state_dict(), "SentimentModel.bin")
# update max loss
min_f1 = validation_f1
# plotting scores
scores = np.array(scores)
fig, ax = plt.subplots(1, 2, figsize=(14,6))
ax[0].plot(range(SentConfig.EPOCHS), scores[:,0], 'r')
ax[1].plot(range(SentConfig.EPOCHS), scores[:,1])
ax[0].set(xlabel='Epoch num', ylabel='F1 Score')
ax[1].set(xlabel='Epoch num', ylabel='Accuracy')
ax[0].set_title('validation set f1 score at each epoch')
ax[1].set_title('validation set accuracy at each apoch')
# F1 score calculation on test predictions
state_dict_ = torch.load('SentimentModel.bin')
model = SentimentModel()
model.load_state_dict(state_dict_)
model.to(device)
_, results = evaluation_function(test_loader, model, device, inference=True)
print(classification_report(results[:,1], results[:,0]))
print(round(accuracy_score(results[:,1], results[:,0]),4))
to_lower_case = bert_layer.resolved_object.do_lower_case.numpy()
tokenizer = BertTokenizer(vocabulary_file, to_lower_case)
VOCAB_LENGTH = len(tokenizer.vocab)
EMB_DIM = 200
CNN_FILTERS = 100
DNN_UNITS = 256
OUTPUT_CLASSES = 6
DROPOUT_RATE = 0.2
NB_EPOCHS = 8
model = SentimentModel(vocabulary_size=VOCAB_LENGTH,
embedding_dimensions=EMB_DIM,
cnn_filters=CNN_FILTERS,
dnn_units=DNN_UNITS,
model_output_classes=OUTPUT_CLASSES,
dropout_rate=DROPOUT_RATE)
if OUTPUT_CLASSES == 2:
model.compile(loss="binary_crossentropy",
optimizer="adam",
metrics=["accuracy"])
else:
model.compile(loss="sparse_categorical_crossentropy",
optimizer="adam",
metrics=["sparse_categorical_accuracy"])
latest = tf.train.latest_checkpoint('./new_weights')
model.load_weights(latest)
test_data = TensorDataset(torch.LongTensor(x_test), torch.LongTensor(y_test)) batch_size = 50 train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=True) valid_loader = DataLoader(valid_data, batch_size=batch_size, shuffle=True) test_loader = DataLoader(test_data, batch_size=batch_size, shuffle=True) # Instantiate the model w/ hyperparams vocab_size = len(vocab_to_int)+1 # +1 for the 0 padding output_size = 1 embedding_dim = 400 hidden_dim = 256 n_layers = 2 model = SentimentModel(vocab_size, output_size, embedding_dim, hidden_dim, n_layers) # print(model) # loss and optimization functions lr=0.01 criterion = torch.nn.BCELoss() optimizer = torch.optim.Adam(model.parameters(), lr=lr) # training params epochs = 2 counter = 0 print_every = 100 clip = 5 model.train()
