def run(n_init,
max_features,
umap_n_components,
dataset,
val_dataset,
labels,
result_dir,
random_state
):
# Set random states
np.random.seed(random_state)
target_names = fetch_20newsgroups().target_names
idx = [target_names.index(l) for l in labels]
# load data
train_df = pd.read_csv(dataset)
train_df = train_df.query("labels in @idx")
print(train_df.shape)
train_texts = train_df['texts'].to_numpy()
train_labels = train_df['labels'].to_numpy()
print(train_texts.shape)
val_df = pd.read_csv(val_dataset)
val_df = val_df.query("labels in @idx")
print(val_df.shape)
val_texts = val_df['texts'].to_numpy()
val_labels = val_df['labels'].to_numpy()
print(val_texts.shape)
tfidf = TfidfVectorizer(max_features=max_features, stop_words='english')
X_train = tfidf.fit_transform(train_texts)
X_test = tfidf.transform(val_texts)
#umap = UMAP(n_components=umap_n_components)
#X_train = umap.fit_transform(X_train.toarray())
#X_test = umap.transform(X_test.toarray())
print(len(np.unique(train_labels)))
kmeans = KMeans(n_init=n_init, n_clusters=len(np.unique(train_labels)))
kmeans.fit(X_train)
predicted_labels = kmeans.predict(X_test)
best_matching, accuracy = cluster_accuracy(val_labels, predicted_labels)
ari = adjusted_rand_score(val_labels, predicted_labels)
nmi = normalized_mutual_info_score(val_labels, predicted_labels)
purity = purity_score(y_true=val_labels, y_pred=predicted_labels)
run_results = {}
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
run_results['purity'] = purity # use purity to compare with microsoft paper
os.makedirs(result_dir, exist_ok=True)
result_df = pd.DataFrame.from_records([run_results])
result_df.to_csv(os.path.join(result_dir, f'20newsgroups_samples-kmeans.csv'), index=False)
def run(n_init, max_features, umap_n_components, dataset, result_dir,
random_state, do_umap):
# Set random states
np.random.seed(random_state)
# load data
train_df = pd.read_csv(dataset)
texts = train_df['texts'].to_numpy()
labels = train_df['labels'].to_numpy()
le = LabelEncoder()
labels = le.fit_transform(labels)
tfidf = TfidfVectorizer(max_features=max_features, stop_words='english')
X_train = tfidf.fit_transform(texts)
if do_umap:
umap = UMAP(n_components=umap_n_components)
X_train = umap.fit_transform(X_train.toarray())
kmeans = KMeans(n_init=n_init, n_clusters=len(np.unique(labels)))
predicted_labels = kmeans.fit_predict(X_train)
best_matching, accuracy = cluster_accuracy(labels, predicted_labels)
ari = adjusted_rand_score(labels, predicted_labels)
nmi = normalized_mutual_info_score(labels, predicted_labels)
purity = purity_score(y_true=labels, y_pred=predicted_labels)
run_results = {}
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
run_results[
'purity'] = purity # use purity to compare with microsoft paper
os.makedirs(result_dir, exist_ok=True)
result_df = pd.DataFrame.from_records([run_results])
result_df.to_csv(os.path.join(result_dir, f'trec6-kmeans.csv'),
index=False)
def run(n_init, models, embedding_extractor, batch_size, dataset,
train_idx_file, val_idx_file, result_dir, random_state, device):
# Set random states
np.random.seed(random_state)
torch.manual_seed(random_state)
torch.cuda.manual_seed_all(random_state)
os.makedirs(result_dir, exist_ok=True)
# load data
df = pd.read_csv(dataset)
with open(train_idx_file, 'r') as f:
train_idx = np.array(list(map(int, f.readlines())))
with open(val_idx_file, 'r') as f:
val_idx = np.array(list(map(int, f.readlines())))
all_idx = np.concatenate((train_idx, val_idx))
df_train = df.iloc[all_idx].copy()
train_texts = df_train['texts'].to_numpy()
train_labels = df_train['labels'].to_numpy()
train_data = TextDataset(train_texts, train_labels)
train_data_loader = DataLoader(dataset=train_data,
batch_size=batch_size,
shuffle=False)
df_val = df.iloc[val_idx].copy()
val_texts = df_val['texts'].to_numpy()
val_labels = df_val['labels'].to_numpy()
val_data = TextDataset(val_texts, val_labels)
val_data_loader = DataLoader(dataset=val_data,
batch_size=batch_size,
shuffle=False)
results = []
for model in models:
# init lm model & tokenizer
lm_model = AutoModel.from_pretrained(model,
return_dict=True,
output_hidden_states=True)
tokenizer = AutoTokenizer.from_pretrained(model,
return_dict=True,
output_hidden_states=True)
lm_model.to(device)
train_embeddings = []
train_labels = []
for batch_texts, batch_labels in tqdm(
train_data_loader, desc="Extracting train embeddings"):
inputs = tokenizer(list(batch_texts),
return_tensors='pt',
padding=True,
truncation=True)
inputs = inputs.to(device)
with torch.no_grad():
outputs = lm_model(**inputs)
extracted_embeddings = embedding_extractor(
outputs).cpu().detach().numpy()
train_embeddings.append(extracted_embeddings)
train_labels.extend(batch_labels.numpy().astype('int'))
X_train = np.vstack(train_embeddings)
train_labels = np.array(train_labels)
test_embeddings = []
val_labels = []
for batch_texts, batch_labels in tqdm(
val_data_loader, desc="Extracting val embeddings"):
inputs = tokenizer(list(batch_texts),
return_tensors='pt',
padding=True,
truncation=True)
inputs = inputs.to(device)
with torch.no_grad():
outputs = lm_model(**inputs)
extracted_embeddings = embedding_extractor(
outputs).cpu().detach().numpy()
test_embeddings.append(extracted_embeddings)
val_labels.extend(batch_labels.numpy().astype('int'))
X_test = np.vstack(test_embeddings)
val_labels = np.array(val_labels)
kmeans = KMeans(n_init=n_init, n_clusters=len(np.unique(train_labels)))
kmeans.fit(X_train)
predicted_labels = kmeans.predict(X_test)
best_matching, accuracy = cluster_accuracy(val_labels,
predicted_labels)
ari = adjusted_rand_score(val_labels, predicted_labels)
nmi = normalized_mutual_info_score(val_labels, predicted_labels)
purity = purity_score(y_true=val_labels, y_pred=predicted_labels)
run_results = {}
run_results['model'] = model
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
run_results[
'purity'] = purity # use purity to compare with microsoft paper
results.append(run_results)
with open(
os.path.join(result_dir,
f'{model.replace("/", "_")}_embeddings.h'),
'wb') as f:
pickle.dump([X_train, train_labels, X_test, val_labels], f)
result_df = pd.DataFrame.from_records(results)
result_df.to_csv(os.path.join(
result_dir, f'ag_news_subset5-sbert-embeddings-kmeans.csv'),
index=False)
def run(n_epochs, lr, train_batch_size, val_batch_size, base_model,
clustering_loss_weight, embedding_extractor, annealing_alphas, dataset,
val_dataset, result_dir, early_stopping, early_stopping_tol, device,
random_state):
# Set random states
np.random.seed(random_state)
torch.manual_seed(random_state)
torch.cuda.manual_seed_all(random_state)
# load data
train_df = pd.read_csv(dataset)
train_texts = train_df['texts'].to_numpy()
train_labels = train_df['labels'].to_numpy()
train_data = TextDataset(train_texts, train_labels)
train_data_loader = DataLoader(dataset=train_data,
batch_size=train_batch_size,
shuffle=False)
val_df = pd.read_csv(val_dataset)
val_texts = val_df['texts'].to_numpy()
val_labels = val_df['labels'].to_numpy()
val_data = TextDataset(val_texts, val_labels)
val_data_loader = DataLoader(dataset=val_data,
batch_size=val_batch_size,
shuffle=False)
# init lm model & tokenizer
lm_model = AutoModelForMaskedLM.from_pretrained(base_model,
return_dict=True,
output_hidden_states=True)
tokenizer = AutoTokenizer.from_pretrained(base_model,
return_dict=True,
output_hidden_states=True)
lm_model.to(device)
# init clustering model
model, initial_centroids, initial_embeddings = init_model(
lm_model=lm_model,
tokenizer=tokenizer,
data_loader=train_data_loader,
embedding_extractor=embedding_extractor,
n_clusters=np.unique(train_labels).shape[0],
device=device)
# init optimizer & scheduler
opt = torch.optim.RMSprop(
params=model.parameters(),
lr=lr, # 2e-5, 5e-7,
eps=1e-8)
total_steps = len(train_data_loader) * n_epochs
scheduler = get_linear_schedule_with_warmup(
optimizer=opt,
num_warmup_steps=int(len(train_data_loader) * 0.5),
num_training_steps=total_steps)
# train the model
hist = train(n_epochs=n_epochs,
model=model,
optimizer=opt,
scheduler=scheduler,
annealing_alphas=annealing_alphas,
train_data_loader=train_data_loader,
eval_data_loader=val_data_loader,
clustering_loss_weight=clustering_loss_weight,
early_stopping=early_stopping,
early_stopping_tol=early_stopping_tol,
verbose=True)
# do eval
run_results = {}
predicted_labels, true_labels = evaluate(model=model,
eval_data_loader=val_data_loader,
verbose=True)
best_matching, accuracy = cluster_accuracy(true_labels, predicted_labels)
ari = adjusted_rand_score(true_labels, predicted_labels)
nmi = normalized_mutual_info_score(true_labels, predicted_labels)
purity = purity_score(y_true=true_labels, y_pred=predicted_labels)
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
run_results[
'purity'] = purity # use purity to compare with microsoft paper
# save train hist
os.makedirs(result_dir, exist_ok=True)
result_df = pd.DataFrame.from_records([run_results])
result_df.to_csv(os.path.join(result_dir, '20_newsgroups-distilbert.csv'),
index=False)
# save results & model
os.makedirs(result_dir, exist_ok=True)
with open(os.path.join(result_dir, 'train_hist.h'), 'wb') as f:
pickle.dump(hist, file=f)
torch.save(model, os.path.join(result_dir, 'model.bin'))
def run(n_init,
max_features,
umap_n_components,
dataset,
train_idx_file,
val_idx_file,
result_dir,
random_state,
do_umap=True):
# Set random states
np.random.seed(random_state)
# load data
df = pd.read_csv(dataset)
with open(train_idx_file, 'r') as f:
train_idx = np.array(list(map(int, f.readlines())))
with open(val_idx_file, 'r') as f:
val_idx = np.array(list(map(int, f.readlines())))
all_idx = np.concatenate((train_idx, val_idx))
df_train = df.iloc[all_idx].copy()
train_texts = df_train['texts'].to_numpy()
train_labels = df_train['labels'].to_numpy()
df_val = df.iloc[val_idx].copy()
val_texts = df_val['texts'].to_numpy()
val_labels = df_val['labels'].to_numpy()
tfidf = TfidfVectorizer(max_features=max_features, stop_words='english')
X_train = tfidf.fit_transform(train_texts)
X_test = tfidf.transform(val_texts)
if do_umap:
umap = UMAP(n_components=umap_n_components)
X_train = umap.fit_transform(X_train.toarray())
X_test = umap.transform(X_test.toarray())
kmeans = KMeans(n_init=n_init, n_clusters=len(np.unique(train_labels)))
kmeans.fit(X_train)
predicted_labels = kmeans.predict(X_test)
best_matching, accuracy = cluster_accuracy(val_labels, predicted_labels)
ari = adjusted_rand_score(val_labels, predicted_labels)
nmi = normalized_mutual_info_score(val_labels, predicted_labels)
purity = purity_score(y_true=val_labels, y_pred=predicted_labels)
run_results = {}
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
run_results[
'purity'] = purity # use purity to compare with microsoft paper
os.makedirs(result_dir, exist_ok=True)
result_df = pd.DataFrame.from_records([run_results])
result_df.to_csv(os.path.join(result_dir, f'ag_news_subset5-kmeans.csv'),
index=False)
def run(n_epochs,
hyperparam_grid,
batch_size,
val_batch_size,
base_model,
dataset,
train_idx_file,
val_idx_file,
result_dir,
early_stopping,
early_stopping_tol,
device,
random_state
):
# Set random states
np.random.seed(random_state)
torch.manual_seed(random_state)
torch.cuda.manual_seed_all(random_state)
# load data
df = pd.read_csv(dataset)
with open(train_idx_file, 'r') as f:
train_idx = np.array(list(map(int, f.readlines())))
with open(val_idx_file, 'r') as f:
val_idx = np.array(list(map(int, f.readlines())))
all_idx = np.append(train_idx, val_idx)
df_train = df.iloc[all_idx]
train_texts = df_train['texts'].to_numpy()
train_labels = df_train['labels'].to_numpy()
train_data = TextDataset(train_texts, train_labels)
train_data_loader = DataLoader(dataset=train_data, batch_size=batch_size, shuffle=False)
df_val = df.iloc[val_idx]
val_texts = df_val['texts'].to_numpy()
val_labels = df_val['labels'].to_numpy()
val_data = TextDataset(val_texts, val_labels)
val_data_loader = DataLoader(dataset=val_data, batch_size=val_batch_size, shuffle=False)
# insert code here!
results = []
param_grid = ParameterGrid(hyperparam_grid)
for run_idx, params in enumerate(param_grid):
print(f'Run: {run_idx+1}/{len(list(param_grid))}')
print("Running with params:")
pprint(params)
# init lm model & tokenizer
lm_model = AutoModelForMaskedLM.from_pretrained(base_model, return_dict=True, output_hidden_states=True)
tokenizer = AutoTokenizer.from_pretrained(base_model, return_dict=True, output_hidden_states=True)
lm_model.to(device)
# init clustering model
model, initial_centroids, initial_embeddings = init_model(
lm_model=lm_model,
tokenizer=tokenizer,
data_loader=train_data_loader,
embedding_extractor=params['embedding_extractor'],
n_clusters=np.unique(train_labels).shape[0],
device=device
)
# init optimizer & scheduler
opt = torch.optim.RMSprop(
params=model.parameters(),
lr=params['lr'], # hier weitermachen
eps=1e-8
)
total_steps = len(train_data_loader) * n_epochs
scheduler = get_linear_schedule_with_warmup(
optimizer=opt,
num_warmup_steps=int(len(train_data_loader) * 0.5),
num_training_steps=total_steps
)
# train the model
hist = train(
n_epochs=n_epochs,
model=model,
optimizer=opt,
scheduler=scheduler,
annealing_alphas=params['annealing_alphas'],
train_data_loader=train_data_loader,
eval_data_loader=val_data_loader,
clustering_loss_weight=params['clustering_loss_weight'],
early_stopping=early_stopping,
early_stopping_tol=early_stopping_tol,
verbose=True
)
# do eval
run_results = {**{f'param_{key}': value for key, value in params.items()}}
predicted_labels, true_labels = evaluate(
model=model,
eval_data_loader=val_data_loader,
verbose=True
)
best_matching, accuracy = cluster_accuracy(true_labels, predicted_labels)
ari = adjusted_rand_score(true_labels, predicted_labels)
nmi = normalized_mutual_info_score(true_labels, predicted_labels)
purity = purity_score(y_true=true_labels, y_pred=predicted_labels)
run_results['best_matching'] = best_matching
run_results['accuracy'] = accuracy
run_results['ari'] = ari
run_results['nmi'] = nmi
# save train hist
os.makedirs(result_dir, exist_ok=True)
results.append(run_results)
result_df = pd.DataFrame.from_records(results)
result_df.to_csv(os.path.join(result_dir, 'opt_results_ag_news_subset10.csv'), index=False)
with open(os.path.join(result_dir, f'train_hist_run{run_idx}.h'), 'wb') as f:
pickle.dump(hist, file=f)