File size: 16,331 Bytes
2443328 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 |
## LIBRARIES ###
## Data
import numpy as np
from numpy.core.numeric import outer
import pandas as pd
import torch
import pickle
from tqdm import tqdm
from math import floor
from collections import defaultdict
from transformers import AutoTokenizer
#pd.set_option('precision', 2)
#pd.options.display.float_format = '${:,.2f}'.format
# Analysis
# from gensim.models.doc2vec import Doc2Vec
# from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score
import nltk
from nltk.cluster import KMeansClusterer
import scipy.spatial.distance as sdist
from scipy.spatial import distance_matrix
# nltk.download('punkt') #make sure that punkt is downloaded
# App & Visualization
import streamlit as st
import altair as alt
import plotly.graph_objects as go
from streamlit_vega_lite import altair_component
# utils
from random import sample
from seal import utils as ut
def down_samp(embedding):
"""Down sample a data frame for altiar visualization """
# total number of positive and negative sentiments in the class
#embedding = embedding.groupby('slice').apply(lambda x: x.sample(frac=0.3))
total_size = embedding.groupby(['slice', 'label'], as_index=False).count()
user_data = 0
# if 'Your Sentences' in str(total_size['slice']):
# tmp = embedding.groupby(['slice'], as_index=False).count()
# val = int(tmp[tmp['slice'] == "Your Sentences"]['source'])
# user_data = val
max_sample = total_size.groupby('slice').max()['content']
# # down sample to meeting altair's max values
# # but keep the proportional representation of groups
down_samp = 1/(sum(max_sample.astype(float))/(1000-user_data))
max_samp = max_sample.apply(lambda x: floor(
x*down_samp)).astype(int).to_dict()
max_samp['Your Sentences'] = user_data
# # sample down for each group in the data frame
embedding = embedding.groupby('slice').apply(
lambda x: x.sample(n=max_samp.get(x.name))).reset_index(drop=True)
# # order the embedding
return(embedding)
#down sample low loss points only so misclassified examples are not down sampled in viz
def down_samp_ll(embedding):
df_ll = embedding[embedding['slice'] == 'low-loss']
#if(len(df_ll)<5000):
# return embedding
#else:
df_hl = embedding[embedding['slice'] == 'high-loss']
down_samp = len(df_ll) - (1000-len(df_hl))
df_ll.sample(n=down_samp)
embedding.drop(df_ll.index)
return embedding
def data_comparison(df):
selection = alt.selection_multi(fields=['cluster', 'label'])
color = alt.condition(alt.datum.slice == 'high-loss', alt.Color('cluster:N', scale=alt.Scale(
domain=df.cluster.unique().tolist()), legend=None), alt.value("lightgray"))
opacity = alt.condition(selection, alt.value(0.7), alt.value(0.25))
# basic chart
scatter = alt.Chart(df).mark_point(size=100, filled=True).encode(
x=alt.X('x:Q', axis=None),
y=alt.Y('y:Q', axis=None),
color=color,
shape=alt.Shape('label:N', scale=alt.Scale(
range=['circle', 'diamond'])),
tooltip=['cluster:N', 'slice:N', 'content:N', 'label:N', 'pred:N'],
opacity=opacity
).properties(
width=1000,
height=800
).interactive()
legend = alt.Chart(df).mark_point(size=100, filled=True).encode(
x=alt.X("label:N"),
y=alt.Y('cluster:N', axis=alt.Axis(
orient='right'), sort='ascending', title=''),
shape=alt.Shape('label:N', scale=alt.Scale(
range=['circle', 'diamond']), legend=None),
color=color,
).add_selection(
selection
)
layered = scatter | legend
layered = layered.configure_axis(
grid=False
).configure_view(
strokeOpacity=0
)
content = legend.encode(text='content:N')
return layered
def viz_panel(embedding_df):
""" Visualization Panel Layout"""
all_metrics = {}
st.warning("**Error group visualization**")
with st.expander("How to read this chart:"):
st.markdown("* Each **point** is an input example.")
st.markdown("* Gray points have low-loss and the colored have high-loss. High-loss instances are clustered using **kmeans** and each color represents a cluster.")
st.markdown(
"* The **shape** of each point reflects the label category -- positive (diamond) or negative sentiment (circle).")
#st.altair_chart(data_comparison(down_samp(embedding_df)), use_container_width=True)
viz = data_comparison(embedding_df)
st.altair_chart(viz, use_container_width=True)
@st.cache()
def frequent_tokens(data, tokenizer, loss_quantile=0.95, top_k=200, smoothing=0.005):
unique_tokens = []
tokens = []
for row in tqdm(data['content']):
tokenized = tokenizer(row, padding=True, truncation=True, return_tensors='pt')
tokens.append(tokenized['input_ids'].flatten())
unique_tokens.append(torch.unique(tokenized['input_ids']))
losses = data['loss'].astype(float)
high_loss = losses.quantile(loss_quantile)
loss_weights = np.where(losses > high_loss,losses,0.0)
loss_weights = loss_weights / loss_weights.sum()
token_frequencies = defaultdict(float)
token_frequencies_error = defaultdict(float)
weights_uniform = np.full_like(loss_weights, 1 / len(loss_weights))
for i in tqdm(range(len(data))):
for token in unique_tokens[i]:
token_frequencies[token.item()] += weights_uniform[i]
token_frequencies_error[token.item()] += loss_weights[i]
token_lrs = {k: (smoothing+token_frequencies_error[k]) / (
smoothing+token_frequencies[k]) for k in token_frequencies}
tokens_sorted = list(map(lambda x: x[0], sorted(
token_lrs.items(), key=lambda x: x[1])[::-1]))
top_tokens = []
for i, (token) in enumerate(tokens_sorted[:top_k]):
top_tokens.append(['%10s' % (tokenizer.decode(token)), '%.4f' % (token_frequencies[token]), '%.4f' % (
token_frequencies_error[token]), '%4.2f' % (token_lrs[token])])
return pd.DataFrame(top_tokens, columns=['token', 'freq', 'error-freq', 'ratio'])
def load_precached_groups(data_ll, df_list, num_clusters, group_dict_path, group_idx_path, num_points=1000):
merged = dynamic_groups(df_list, num_clusters)
down_samp = len(data_ll) - (num_points-len(merged))
sample_idx = data_ll.sample(n=down_samp)
data_ll = data_ll.drop(sample_idx.index)
# put all the low loss data in one bigger cluster
data_ll['cluster'] = merged.loc[merged['cluster'].idxmax()].cluster + 1
merged = pd.concat([merged, data_ll])
# merged['cluster'] = merged['cluster'].astype('str')
# with open(group_dict_path, 'rb') as f:
# group_dict = pickle.load(f)
# with open(group_idx_path, 'rb') as f:
# group_idx_dict = pickle.load(f)
# for k,v in group_idx_dict.items():
# label = group_dict.get(k)
# merged.loc[merged.index.isin(v), ['cluster']] = label
return merged
def dynamic_groups(df_list, num_clusters):
merged = pd.DataFrame()
ind = 0
for df in df_list:
kmeans_df, assigned_clusters = kmeans(df, num_clusters=num_clusters)
kmeans_df['cluster'] = kmeans_df['cluster'] + ind*num_clusters
ind = ind+1
merged = pd.concat([merged, kmeans_df])
return merged
@st.cache(ttl=600)
def get_data(inference, emb):
preds = inference.outputs.numpy()
losses = inference.losses.numpy()
embeddings = pd.DataFrame(emb, columns=['x', 'y'])
num_examples = len(losses)
# dataset_labels = [dataset[i]['label'] for i in range(num_examples)]
return pd.concat([pd.DataFrame(np.transpose(np.vstack([dataset[:num_examples]['content'],
dataset[:num_examples]['label'], preds, losses])), columns=['content', 'label', 'pred', 'loss']), embeddings], axis=1)
def kmeans(data, num_clusters=3):
X = np.array(data['embedding'].to_list())
kclusterer = KMeansClusterer(
num_clusters, distance=nltk.cluster.util.cosine_distance,
repeats=25, avoid_empty_clusters=True)
assigned_clusters = kclusterer.cluster(X, assign_clusters=True)
data['cluster'] = pd.Series(
assigned_clusters, index=data.index).astype('int')
data['centroid'] = data['cluster'].apply(lambda x: kclusterer.means()[x])
return data, assigned_clusters
def distance_from_centroid(row):
return sdist.norm(row['embedding'] - row['centroid'].tolist())
@st.cache(ttl=600)
def craft_prompt(cluster_df):
instruction = "In this task, we'll assign a short and precise label to a cluster of documents based on the topics or concepts most relevant to these documents. The documents are all subsets of a sentiment classification dataset.\n"
if len(cluster_df) > 10:
content = cluster_df['content'].str[:600].tolist()
else:
content = cluster_df['content'].str[:1000].tolist()
examples = '\n - '.join(content)
text = instruction + '- ' + examples + '\n Cluster label:'
return text.strip()
@st.cache(ttl=600)
def topic_distribution(weights, smoothing=0.01):
topic_frequencies = defaultdict(float)
topic_frequencies_error = defaultdict(float)
weights_uniform = np.full_like(weights, 1 / len(weights))
num_examples = len(weights)
for i in range(num_examples):
example = dataset[i]
category = example['title']
topic_frequencies[category] += weights_uniform[i]
topic_frequencies_error[category] += weights[i]
topic_ratios = {c: (smoothing + topic_frequencies_error[c]) / (
smoothing + topic_frequencies[c]) for c in topic_frequencies}
categories_sorted = map(lambda x: x[0], sorted(
topic_ratios.items(), key=lambda x: x[1], reverse=True))
topic_distr = []
for category in categories_sorted:
topic_distr.append(['%.3f' % topic_frequencies[category], '%.3f' %
topic_frequencies_error[category], '%.2f' % topic_ratios[category], '%s' % category])
return pd.DataFrame(topic_distr, columns=['Overall frequency', 'Error frequency', 'Ratio', 'Category'])
def populate_session(dataset, model):
data_df = read_file_to_df(
'./assets/data/'+dataset + '_' + model+'.parquet')
if model == 'albert-base-v2-yelp-polarity':
tokenizer = AutoTokenizer.from_pretrained('textattack/'+model)
else:
tokenizer = AutoTokenizer.from_pretrained(model)
# if "user_data" not in st.session_state:
# st.session_state["user_data"] = data_df
# if "selected_slice" not in st.session_state:
# st.session_state["selected_slice"] = None
return tokenizer
@st.cache(allow_output_mutation=True)
def read_file_to_df(file):
return pd.read_parquet(file)
if __name__ == "__main__":
### STREAMLIT APP CONGFIG ###
st.set_page_config(layout="wide", page_title="Interactive Error Analysis")
ut.init_style()
lcol, rcol = st.columns([5, 2])
# ******* loading the mode and the data
#st.sidebar.mardown("<h4>Interactive Error Analysis</h4>", unsafe_allow_html=True)
dataset = st.sidebar.selectbox(
"Dataset",
["amazon_polarity", "yelp_polarity", "imdb"],
index=1
)
model = st.sidebar.selectbox(
"Model",
["distilbert-base-uncased-finetuned-sst-2-english",
"albert-base-v2-yelp-polarity", "distilbert-imdb"],
)
### LOAD DATA AND TOKENIZER VARIABLES ###
##uncomment the next next line to run dynamically and not from file
#tokenizer = populate_session(dataset, model)
if dataset == 'imdb':
data_df = read_file_to_df('./assets/data/imdb_distilbert.parquet')
else:
data_df = read_file_to_df(
'./assets/data/'+dataset + '_' + model+'.parquet')
data_df = data_df[:20000]
loss_quantile = st.sidebar.slider(
"Loss Quantile", min_value=0.9, max_value=1.0, step=0.01, value=0.98
)
data_df['loss'] = data_df['loss'].astype(float)
data_df['pred'] = data_df['pred'].astype(int)
losses = data_df['loss']
high_loss = losses.quantile(loss_quantile)
data_df['slice'] = np.where(data_df['loss'] >= high_loss, 'high-loss', 'low-loss')
# drop rows that are not hl
data_hl = pd.DataFrame(data_df[data_df['slice'] == 'high-loss'])
#data_hl = data_hl.drop(data_hl[data_hl.pred==data_hl.label].index)
data_ll = pd.DataFrame(data_df[data_df['slice'] == 'low-loss'])
# this is to allow clustering over each error type. fp, fn for binary classification
df_list = [d for _, d in data_hl.groupby(['label'])]
run_kmeans = st.sidebar.radio(
"Cluster error group?", ('True', 'False'), index=0)
num_clusters = st.sidebar.slider(
"# clusters", min_value=1, max_value=60, step=1, value=3)
num_points = st.sidebar.slider(
"# data points to visualize", min_value=1000, max_value=5000, step=100, value=1000)
selected_cluster = st.sidebar.number_input(
label='Cluster #:', max_value=num_clusters-1, min_value=0)
if run_kmeans == 'True':
with st.spinner(text='running kmeans...'):
group_dict_path = './assets/data/cluster-labels/'+dataset+'.pkl'
group_idx_path = './assets/data/cluster-labels/'+dataset+'_idx.pkl'
#data_hl_path = './assets/data/high-loss/'+dataset+'.parquet'
merged = load_precached_groups(data_ll, df_list, int(
(num_clusters/2)), group_dict_path, group_idx_path, num_points=num_points)
#dynamic_groups(df_list,)
#tmp = pd.concat([data_ll, merged], axis =0, ignore_index=True)
cluster_content = craft_prompt(
merged.loc[merged['cluster'] == selected_cluster])
with lcol:
st.markdown('<h5>Error Groups</h5>', unsafe_allow_html=True)
with st.expander("How to read this table:"):
st.markdown(
"* *Error groups* refers to the subset of evaluation dataset the model performs poorly on.")
st.markdown(
"* The table displays model error groups on the evaluation dataset, sorted by loss.")
st.markdown(
"* Each row is an input example that includes the label, model pred, loss, and error group.")
with st.spinner(text='loading error groups...'):
#dataframe=read_file_to_df('./assets/data/'+dataset+ '_'+ model+'_error-slices.parquet')
#uncomment the next next line to run dynamically and not from file
dataframe = merged[['content', 'label', 'pred', 'loss', 'cluster']].sort_values(
by=['loss'], ascending=False)
#table_html = dataframe.to_html(columns=['content', 'label', 'pred', 'loss', 'cluster'], max_rows=50)
#table_html = table_html.replace("<th>", '<th align="left">') # left-align the headers
st.write(dataframe.style.format(
{'loss': '{:.2f}'}), width=1000, height=300)
with rcol:
with st.spinner(text='loading...'):
st.markdown('<h5>Word Distribution in Error Groups</h5>',
unsafe_allow_html=True)
#uncomment the next two lines to run dynamically and not from file
# if model == 'albert-base-v2-yelp-polarity':
# tokenizer = AutoTokenizer.from_pretrained('textattack/'+model)
# else:
# tokenizer = AutoTokenizer.from_pretrained(model)
# commontokens = frequent_tokens(data_df, tokenizer, loss_quantile=loss_quantile)
if dataset == 'imdb':
commontokens = read_file_to_df('./assets/data/imdb_distilbert_commontokens.parquet')
else:
commontokens = read_file_to_df(
'./assets/data/'+dataset + '_' + model+'_commontokens.parquet')
with st.expander("How to read this table:"):
st.markdown(
"* The table displays the most frequent tokens in error groups, relative to their frequencies in the val set.")
st.write(commontokens)
with st.spinner(text='loading visualization...'):
viz_panel(merged)
st.sidebar.download_button(
data=cluster_content,
label="Build prompt from data",
file_name='prompt'
)
|