def cluster_embeddings(Convo1_Embeddings, Convo2_Embeddings, df_labels1,
df_labels2, common_speaker):
Agg_EmbeddingList = torch.cat([Convo1_Embeddings, Convo2_Embeddings],
dim=0).cpu().numpy()
kmeans = KMeans(n_clusters=2, n_init=20, algorithm="elkan")
ac = AgglomerativeClustering(n_clusters=2)
pca = PCA(n_components=2)
Convo2_Embeddings = StandardScaler().fit_transform(
Convo2_Embeddings.cpu().numpy())
principalComponents = pca.fit_transform(Convo2_Embeddings)
#kmeans.fit(principalComponents)_pre
#ac.fit(principalComponents)
ac = ac.fit_predict(principalComponents)
#y_kmeans = kmeans.predict(principalComponents)
#cluster_labels_1 = y_kmeans#[0:df_labels1.shape[1]]ta
cluster_labels_2 = ac #y_kmeans#[df_labels1.shape[1]+1:]
print(cluster_labels_2)
#common_speaker_1 = df_labels1.loc[common_speaker,:].values.tolist()
common_speaker_2 = df_labels2.loc[common_speaker, :].values.tolist()
speaker_labels = list(df_labels1.index.values) + list(
df_labels2.index.values)
#second_speaker_1_label = [x for x in list(df_labels1.index.values) if common_speaker not in x][0]
second_speaker_2_label = [
x for x in list(df_labels2.index.values) if common_speaker not in x
][0]
#second_speaker_1 = df_labels1.loc[second_speaker_1_label,:].values.tolist()
second_speaker_2 = df_labels2.loc[
second_speaker_2_label, :].values.tolist()
#df_conv_1 = pd.DataFrame(list(zip(cluster_labels_1, common_speaker_1, second_speaker_1)), columns=['cluster', common_speaker, second_speaker_1_label])
df_conv_1 = pd.DataFrame()
df_conv_2 = pd.DataFrame(
list(zip(cluster_labels_2, common_speaker_2, second_speaker_2)),
columns=['cluster', common_speaker, second_speaker_2_label])
return df_conv_1, df_conv_2
def run_dimensionality_reduction(
feature_extractor,
dataset_types,
reduction_model,
already_fit=False,
):
print("Getting features")
feature_datasets = FeatureDatasets(feature_extractor)
xs = []
ys = []
if type(dataset_types) == DatasetType:
dataset_types = [dataset_types]
for dataset_type in dataset_types:
if dataset_type == DatasetType.Competition:
new_xs, new_ys = feature_datasets.get_features(dataset_type), None
else:
new_xs, new_ys = feature_datasets.get_features_and_labels(
dataset_type)
xs.append(new_xs)
if new_ys is not None:
ys.append(new_ys)
xs = torch.cat(xs)
if ys:
ys = torch.cat(ys)
print("Scaling data")
xs = StandardScaler().fit_transform(xs.cpu().detach())
if already_fit:
print("Using pre-fit model")
reduced_features = reduction_model.transform(xs)
else:
print("Running " + str(reduction_model))
reduced_features = reduction_model.fit_transform(xs, ys)
print(reduction_model.explained_variance_ratio_)
print("Aggregating principal components")
principal_df = pd.DataFrame(data=reduced_features, columns=["pc1", "pc2"])
target_df = pd.DataFrame(data=ys, columns=["target"])
final_df = pd.concat([principal_df, target_df], axis=1)
return final_df
