def sum_clf_test(dataset,
config,
opts,
transfer=False,
output_dir=None,
checkpoint_name='scv2_aux_ft_gu_last'):
opts.name = config["name"]
X_test, y_test, posts_test, pids, human_summaries = dataset
vocab = None
if transfer:
opts.transfer = config["pretrained_lm"]
checkpoint = load_checkpoint(opts.transfer)
config["vocab"].update(checkpoint["config"]["vocab"])
dict_pattern_rename(checkpoint["config"]["model"],
{"rnn_": "bottom_rnn_"})
config["model"].update(checkpoint["config"]["model"])
vocab = checkpoint["vocab"]
####################################################################
# Load Preprocessed Datasets
####################################################################
if config["preprocessor"] == "twitter":
preprocessor = twitter_preprocessor()
else:
preprocessor = None
####################################################################
# Model
####################################################################
ntokens = 70004
model = SummarizationClassifier(ntokens, len(set([0, 1])),
**config["model"])
model.to(opts.device)
clf_criterion = nn.CrossEntropyLoss()
lm_criterion = nn.CrossEntropyLoss(ignore_index=0)
embed_parameters = filter(lambda p: p.requires_grad,
model.embed.parameters())
bottom_parameters = filter(
lambda p: p.requires_grad,
chain(model.bottom_rnn.parameters(), model.vocab.parameters()))
if config["model"]["has_att"]:
top_parameters = filter(
lambda p: p.requires_grad,
chain(model.top_rnn.parameters(), model.attention.parameters(),
model.classes.parameters()))
else:
top_parameters = filter(
lambda p: p.requires_grad,
chain(model.top_rnn.parameters(), model.classes.parameters()))
embed_optimizer = optim.ASGD(embed_parameters, lr=0.0001)
rnn_optimizer = optim.ASGD(bottom_parameters)
top_optimizer = Adam(top_parameters, lr=config["top_lr"])
####################################################################
# Training Pipeline
####################################################################
# Trainer: responsible for managing the training process
trainer = SumClfTrainer(model,
None,
None, (lm_criterion, clf_criterion),
[embed_optimizer, rnn_optimizer, top_optimizer],
config,
opts.device,
valid_loader_train_set=None,
unfreeze_embed=config["unfreeze_embed"],
unfreeze_rnn=config["unfreeze_rnn"],
test_loader=None)
####################################################################
# Resume Training from a previous checkpoint
####################################################################
if transfer:
print("Transferring Encoder weights ...")
dict_pattern_rename(checkpoint["model"], {
"encoder": "bottom_rnn",
"decoder": "vocab"
})
load_state_dict_subset(model, checkpoint["model"])
print(model)
_vocab = trainer.load_checkpoint(name=checkpoint_name, path=None)
test_set = SUMDataset(X_test,
posts_test,
y_test,
seq_len=config['data']['seq_len'],
post_len=config['data']['post_len'],
preprocess=preprocessor,
vocab=_vocab)
test_lengths = [len(x) for x in test_set.data]
test_sampler = SortedSampler(test_lengths)
# test_loader = DataLoader(test_set, sampler=test_sampler,
# batch_size=config["batch_size"],
# num_workers=opts.cores, collate_fn=SumCollate())
test_loader = DataLoader(test_set,
sampler=test_sampler,
batch_size=config["batch_size"],
num_workers=0,
collate_fn=SumCollate())
trainer.test_loader = test_loader
_, labels_array, predicted = trainer.test_epoch()
pids_dic = {}
if human_summaries is None:
for x, y, sent, z in zip(y_test, predicted, X_test, pids):
if z in pids_dic:
pids_dic[z].append([x, y, sent])
else:
pids_dic[z] = [[x, y, sent]]
else:
for x, y, sent, z, h_summary in zip(y_test, predicted, X_test, pids,
human_summaries):
if z in pids_dic:
pids_dic[z].append([x, y, sent, h_summary])
else:
pids_dic[z] = [[x, y, sent, h_summary]]
# import os
# if not os.path.exists('{}/ref_abs'.format(output_dir)):
# os.mkdir('{}/ref_abs'.format(output_dir))
# if not os.path.exists('{}/dec'.format(output_dir)):
# os.mkdir('{}/dec'.format(output_dir))
file_index = 0
all_summaries = []
for elem_key in pids_dic:
current_summary = ''
for pair in pids_dic[elem_key]:
if pair[1] == 1:
current_summary += pair[2] + '\n'
all_summaries.append(current_summary)
return all_summaries
def sent_clf(dataset, config, opts, transfer=False):
from logger.experiment import Experiment
opts.name = config["name"]
X_train, y_train, _, X_val, y_val, _ = dataset
vocab = None
if transfer:
opts.transfer = config["pretrained_lm"]
checkpoint = load_checkpoint(opts.transfer)
config["vocab"].update(checkpoint["config"]["vocab"])
dict_pattern_rename(checkpoint["config"]["model"],
{"rnn_": "bottom_rnn_"})
config["model"].update(checkpoint["config"]["model"])
vocab = checkpoint["vocab"]
####################################################################
# Load Preprocessed Datasets
####################################################################
if config["preprocessor"] == "twitter":
preprocessor = twitter_preprocessor()
else:
preprocessor = None
print("Building training dataset...")
train_set = ClfDataset(X_train,
y_train,
vocab=vocab,
preprocess=preprocessor,
vocab_size=config["vocab"]["size"],
seq_len=config["data"]["seq_len"])
print("Building validation dataset...")
val_set = ClfDataset(X_val,
y_val,
seq_len=train_set.seq_len,
preprocess=preprocessor,
vocab=train_set.vocab)
src_lengths = [len(x) for x in train_set.data]
val_lengths = [len(x) for x in val_set.data]
# select sampler & dataloader
train_sampler = BucketBatchSampler(src_lengths, config["batch_size"], True)
val_sampler = SortedSampler(val_lengths)
val_sampler_train = SortedSampler(src_lengths)
train_loader = DataLoader(train_set,
batch_sampler=train_sampler,
num_workers=opts.cores,
collate_fn=ClfCollate())
val_loader = DataLoader(val_set,
sampler=val_sampler,
batch_size=config["batch_size"],
num_workers=opts.cores,
collate_fn=ClfCollate())
val_loader_train_dataset = DataLoader(train_set,
sampler=val_sampler_train,
batch_size=config["batch_size"],
num_workers=opts.cores,
collate_fn=ClfCollate())
####################################################################
# Model
####################################################################
ntokens = len(train_set.vocab)
model = Classifier(ntokens, len(set(train_set.labels)), **config["model"])
model.to(opts.device)
clf_criterion = nn.CrossEntropyLoss()
lm_criterion = nn.CrossEntropyLoss(ignore_index=0)
embed_parameters = filter(lambda p: p.requires_grad,
model.embed.parameters())
bottom_parameters = filter(
lambda p: p.requires_grad,
chain(model.bottom_rnn.parameters(), model.vocab.parameters()))
if config["model"]["has_att"]:
top_parameters = filter(
lambda p: p.requires_grad,
chain(model.top_rnn.parameters(), model.attention.parameters(),
model.classes.parameters()))
else:
top_parameters = filter(
lambda p: p.requires_grad,
chain(model.top_rnn.parameters(), model.classes.parameters()))
embed_optimizer = optim.ASGD(embed_parameters, lr=0.0001)
rnn_optimizer = optim.ASGD(bottom_parameters)
top_optimizer = Adam(top_parameters, lr=config["top_lr"])
####################################################################
# Training Pipeline
####################################################################
# Trainer: responsible for managing the training process
trainer = SentClfTrainer(model,
train_loader,
val_loader, (lm_criterion, clf_criterion),
[embed_optimizer, rnn_optimizer, top_optimizer],
config,
opts.device,
valid_loader_train_set=val_loader_train_dataset,
unfreeze_embed=config["unfreeze_embed"],
unfreeze_rnn=config["unfreeze_rnn"])
####################################################################
# Experiment: logging and visualizing the training process
####################################################################
# exp = Experiment(opts.name, config, src_dirs=opts.source,
# output_dir=EXP_DIR)
# exp.add_metric("ep_loss_lm", "line", "epoch loss lm",
# ["TRAIN", "VAL"])
# exp.add_metric("ep_loss_cls", "line", "epoch loss class",
# ["TRAIN", "VAL"])
# exp.add_metric("ep_f1", "line", "epoch f1", ["TRAIN", "VAL"])
# exp.add_metric("ep_acc", "line", "epoch accuracy", ["TRAIN", "VAL"])
#
# exp.add_value("epoch", title="epoch summary")
# exp.add_value("progress", title="training progress")
ep_loss_lm = [10000, 10000]
ep_loss_cls = [10000, 10000]
ep_f1 = [0, 0]
ep_acc = [0, 0]
e_log = 0
progress = 0
####################################################################
# Resume Training from a previous checkpoint
####################################################################
if transfer:
print("Transferring Encoder weights ...")
dict_pattern_rename(checkpoint["model"], {
"encoder": "bottom_rnn",
"decoder": "vocab"
})
load_state_dict_subset(model, checkpoint["model"])
print(model)
####################################################################
# Training Loop
####################################################################
best_loss = None
early_stopping = EarlyStopping("min", config["patience"])
for epoch in range(0, config["epochs"]):
train_loss = trainer.train_epoch()
val_loss, y, y_pred = trainer.eval_epoch(val_set=True)
_, y_train, y_pred_train = trainer.eval_epoch(train_set=True)
# exp.update_metric("ep_loss_lm", train_loss[0], "TRAIN")
ep_loss_lm[0] = train_loss[0]
# exp.update_metric("ep_loss_lm", val_loss[0], "VAL")
ep_loss_lm[1] = val_loss[0]
# exp.update_metric("ep_loss_cls", train_loss[1], "TRAIN")
# exp.update_metric("ep_loss_cls", val_loss[1], "VAL")
ep_loss_cls[0] = train_loss[1]
ep_loss_cls[1] = val_loss[1]
# exp.update_metric("ep_f1", f1_macro(y_train, y_pred_train),
# "TRAIN")
ep_f1[0] = f1_macro(y_train, y_pred_train)
# exp.update_metric("ep_f1", f1_macro(y, y_pred), "VAL")
ep_f1[1] = f1_macro(y, y_pred)
# exp.update_metric("ep_acc", acc(y_train, y_pred_train), "TRAIN")
# exp.update_metric("ep_acc", acc(y, y_pred), "VAL")
ep_acc[0] = acc(y_train, y_pred_train)
ep_acc[1] = acc(y, y_pred)
# print('Train lm Loss : {}\nVal lm Loss : {}\nTrain cls Loss : {}\nVal cls Loss : {}\n Train f1 : {}\nVal f1 : {}\nTrain acc : {}\n Val acc : {}'.format(
# ep_loss_lm[0], ep_loss_lm[1], ep_loss_cls[0], ep_loss_cls[1], ep_f1[0], ep_f1[1], ep_acc[0], ep_acc[1]
# ))
# epoch_log = exp.log_metrics(["ep_loss_lm", "ep_loss_cls","ep_f1", "ep_acc"])
epoch_log = 'Train lm Loss : {}\nVal lm Loss : {}\nTrain cls Loss : {}\nVal cls Loss : {}\n Train f1 : {}\nVal f1 : {}\nTrain acc : {}\n Val acc : {}'.format(
ep_loss_lm[0], ep_loss_lm[1], ep_loss_cls[0], ep_loss_cls[1],
ep_f1[0], ep_f1[1], ep_acc[0], ep_acc[1])
print(epoch_log)
# exp.update_value("epoch", epoch_log)
e_log = epoch_log
# print('')
# Save the model if the val loss is the best we've seen so far.
# if not best_loss or val_loss[1] < best_loss:
# best_loss = val_loss[1]
# trainer.best_acc = acc(y, y_pred)
# trainer.best_f1 = f1_macro(y, y_pred)
# trainer.checkpoint(name=opts.name, timestamp=True)
best_loss = val_loss[1]
trainer.best_acc = acc(y, y_pred)
trainer.best_f1 = f1_macro(y, y_pred)
trainer.checkpoint(name=opts.name, tags=str(epoch))
# if early_stopping.stop(val_loss[1]):
# print("Early Stopping (according to classification loss)....")
# break
print("\n" * 2)
return best_loss, trainer.best_acc, trainer.best_f1
X_train, X_test, y_train, y_test = load_wassa()
# 3 - convert labels from strings to integers
label_encoder = LabelEncoder()
label_encoder = label_encoder.fit(y_train)
with open("../submissions/label_encoder.pkl", "wb") as r:
pickle.dump(label_encoder, r)
y_train = label_encoder.transform(y_train)
y_test = label_encoder.transform(y_test)
name = "wassa"
#####################################################################
# Define Dataloaders
#####################################################################
preprocessor = twitter_preprocessor()
# preprocessor = None
if preprocessor is None:
train_name = "train_simple_split_{}".format(name)
val_name = "valid_simple_split_{}".format(name)
else:
train_name = "train_ekphrasis_{}".format(name)
val_name = "valid_ekphrasis_{}".format(name)
train_set = WordDataset(X_train,
y_train,
word2idx,
name=train_name,
max_length=35,
preprocess=preprocessor)
test_set = WordDataset(X_test,
def submission(dataset, models=[], lm=[], gold=[]):
X = load_test_wassa(dataset)
with open("label_encoder.pkl", "rb") as f:
label_encoder = pickle.load(f)
# load embeddings
file = os.path.join(BASE_PATH, "embeddings", "ntua_twitter_300.txt")
word2idx, idx2word, weights = load_word_vectors(file, 300)
dummy_y = [[0] * 6] * len(X)
dummy_y = torch.tensor(dummy_y)
posteriors_list = []
predicted_list = []
for i in range(0, len(models)):
checkpoint_name = models[i]
if lm[i]:
model, optimizer, word2idx, idx2word, loss, acc, f1 = \
load_checkpoint_pre_lm(checkpoint_name)
else:
model, optimizer, vocab, loss, acc, f1 = \
load_checkpoint_with_f1(checkpoint_name)
#####################################################################
# Define Dataloaders
#####################################################################
preprocessor = twitter_preprocessor()
# for new experiments remember to empty _cache!
test_set = WordDataset(X,
dummy_y,
word2idx,
name="wassa_test_submit",
preprocess=preprocessor)
sampler = SequentialSampler(test_set)
test_loader = DataLoader(test_set, batch_size=32, sampler=sampler)
#####################################################################
# Load Trained Model
#####################################################################
model.eval()
model.to(config.DEVICE)
print(model)
#####################################################################
# Evaluate Trained Model on test set & Calculate predictions
#####################################################################
labels, predicted, posteriors = test_clf(model=model,
data_source=test_loader,
device=config.DEVICE)
# pprint(labels)
pprint(predicted)
predicted_list.append(predicted)
posteriors_list.append(posteriors)
# pred, accuracy, f1 = ensemble_voting(predicted_list, gold, dataset)
pred, accuracy, f1 = ensemble_posteriors(posteriors_list, gold, dataset)
#####################################################################
# Create submission file with the predictions3M_GU13__35_noconc_2att
#####################################################################
write_predictions(pred, label_encoder)
return
def sent_clf_no_aux(dataset, config, opts, transfer=False):
from logger.experiment import Experiment
opts.name = config["name"]
X_train, y_train, X_val, y_val = dataset
vocab = None
if transfer:
opts.transfer = config["pretrained_lm"]
checkpoint = load_checkpoint(opts.transfer)
config["vocab"].update(checkpoint["config"]["vocab"])
dict_pattern_rename(checkpoint["config"]["model"],
{"rnn_": "bottom_rnn_"})
config["model"].update(checkpoint["config"]["model"])
vocab = checkpoint["vocab"]
####################################################################
# Data Loading and Preprocessing
####################################################################
if config["preprocessor"] == "twitter":
preprocessor = twitter_preprocessor()
else:
preprocessor = None
print("Building training dataset...")
train_set = ClfDataset(X_train,
y_train,
vocab=vocab,
preprocess=preprocessor,
vocab_size=config["vocab"]["size"],
seq_len=config["data"]["seq_len"])
print("Building validation dataset...")
val_set = ClfDataset(X_val,
y_val,
seq_len=train_set.seq_len,
preprocess=preprocessor,
vocab=train_set.vocab)
src_lengths = [len(x) for x in train_set.data]
val_lengths = [len(x) for x in val_set.data]
# select sampler & dataloader
train_sampler = BucketBatchSampler(src_lengths, config["batch_size"], True)
val_sampler = SortedSampler(val_lengths)
val_sampler_train = SortedSampler(src_lengths)
train_loader = DataLoader(train_set,
batch_sampler=train_sampler,
num_workers=opts.cores,
collate_fn=ClfCollate())
val_loader = DataLoader(val_set,
sampler=val_sampler,
batch_size=config["batch_size"],
num_workers=opts.cores,
collate_fn=ClfCollate())
val_loader_train_dataset = DataLoader(train_set,
sampler=val_sampler_train,
batch_size=config["batch_size"],
num_workers=opts.cores,
collate_fn=ClfCollate())
####################################################################
# Model
####################################################################
ntokens = len(train_set.vocab)
model = NaiveClassifier(ntokens,
len(set(train_set.labels)),
attention=config["model"]["has_att"],
**config["model"])
model.to(opts.device)
criterion = nn.CrossEntropyLoss()
if config["gu"]:
embed_parameters = filter(lambda p: p.requires_grad,
model.embed.parameters())
bottom_parameters = filter(lambda p: p.requires_grad,
chain(model.bottom_rnn.parameters()))
if config["model"]["has_att"]:
top_parameters = filter(
lambda p: p.requires_grad,
chain(model.attention.parameters(),
model.classes.parameters()))
else:
top_parameters = filter(lambda p: p.requires_grad,
model.classes.parameters())
embed_optimizer = Adam(embed_parameters)
rnn_optimizer = Adam(bottom_parameters)
top_optimizer = Adam(top_parameters)
# Trainer: responsible for managing the training process
trainer = SentClfNoAuxTrainer(
model,
train_loader,
val_loader,
criterion, [embed_optimizer, rnn_optimizer, top_optimizer],
config,
opts.device,
valid_loader_train_set=val_loader_train_dataset,
unfreeze_embed=config["unfreeze_embed"],
unfreeze_rnn=config["unfreeze_rnn"])
else:
parameters = filter(lambda p: p.requires_grad, model.parameters())
optimizer = optim.Adam(parameters, lr=config["top_lr"])
# Trainer: responsible for managing the training process
trainer = SentClfNoAuxTrainer(
model,
train_loader,
val_loader,
criterion, [optimizer],
config,
opts.device,
valid_loader_train_set=val_loader_train_dataset)
####################################################################
# Experiment: logging and visualizing the training process
####################################################################
exp = Experiment(opts.name,
config,
src_dirs=opts.source,
output_dir=EXP_DIR)
exp.add_metric("ep_loss", "line", "epoch loss class", ["TRAIN", "VAL"])
exp.add_metric("ep_f1", "line", "epoch f1", ["TRAIN", "VAL"])
exp.add_metric("ep_acc", "line", "epoch accuracy", ["TRAIN", "VAL"])
exp.add_value("epoch", title="epoch summary")
exp.add_value("progress", title="training progress")
####################################################################
# Resume Training from a previous checkpoint
####################################################################
if transfer:
print("Transferring Encoder weights ...")
dict_pattern_rename(checkpoint["model"], {"encoder": "bottom_rnn"})
load_state_dict_subset(model, checkpoint["model"])
print(model)
####################################################################
# Training Loop
####################################################################
best_loss = None
early_stopping = EarlyStopping("min", config["patience"])
for epoch in range(1, config["epochs"] + 1):
train_loss = trainer.train_epoch()
val_loss, y, y_pred = trainer.eval_epoch(val_set=True)
_, y_train, y_pred_train = trainer.eval_epoch(train_set=True)
# Calculate accuracy and f1-macro on the evaluation set
exp.update_metric("ep_loss", train_loss.item(), "TRAIN")
exp.update_metric("ep_loss", val_loss.item(), "VAL")
exp.update_metric("ep_f1", f1_macro(y_train, y_pred_train), "TRAIN")
exp.update_metric("ep_f1", f1_macro(y, y_pred), "VAL")
exp.update_metric("ep_acc", acc(y_train, y_pred_train), "TRAIN")
exp.update_metric("ep_acc", acc(y, y_pred), "VAL")
print()
epoch_log = exp.log_metrics(["ep_loss", "ep_f1", "ep_acc"])
print(epoch_log)
exp.update_value("epoch", epoch_log)
###############################################################
# Unfreezing the model after X epochs
###############################################################
# Save the model if the val loss is the best we've seen so far.
if not best_loss or val_loss < best_loss:
best_loss = val_loss
trainer.best_acc = acc(y, y_pred)
trainer.best_f1 = f1_macro(y, y_pred)
trainer.checkpoint(name=opts.name)
if early_stopping.stop(val_loss):
print("Early Stopping (according to cls loss)....")
break
print("\n" * 2)
return best_loss, trainer.best_acc, trainer.best_f1