dataset-explore / app.py
gabrielaltay's picture
Create app.py
c94b761
"""
BigBIO Dataset Explorer Demo
"""
from collections import Counter
from collections import defaultdict
import string
from datasets import load_dataset
from loguru import logger
import numpy as np
import pandas as pd
import plotly.express as px
import spacy
from spacy import displacy
import streamlit as st
from bigbio.dataloader import BigBioConfigHelpers
from bigbio.hf_maps import BATCH_MAPPERS_TEXT_FROM_SCHEMA
from sklearn.feature_extraction.text import CountVectorizer
st.set_page_config(layout="wide")
IBM_COLORS = [
"#648fff",
"#dc267f",
"#ffb000",
"#fe6100",
"#785ef0",
"#000000",
"#ffffff",
]
def get_html(html: str):
"""Convert HTML so it can be rendered."""
WRAPPER = """<div style="overflow-x: auto; border: 1px solid #e6e9ef; border-radius: 0.25rem; padding: 1rem;\
margin-bottom: 2.5rem">{}</div>"""
# Newlines seem to mess with the rendering
html = html.replace("\n", " ")
return WRAPPER.format(html)
@st.cache()
def load_conhelps():
conhelps = BigBioConfigHelpers()
logger.info(conhelps)
conhelps = conhelps.filtered(lambda x: not x.is_large)
conhelps = conhelps.filtered(lambda x: x.is_bigbio_schema)
conhelps = conhelps.filtered(lambda x: not x.is_local)
return conhelps
def update_axis_font(fig):
fig.update_layout(
xaxis = dict(title_font = dict(size=20)),
yaxis = dict(title_font = dict(size=20)),
)
return fig
def draw_histogram(hist_data, col_name, histnorm=None, nbins=25, xmax=None, loc=st):
fig = px.histogram(
hist_data,
x=col_name,
color="split",
color_discrete_sequence=IBM_COLORS,
marginal="box", # or violin, rug
barmode="group",
hover_data=hist_data.columns,
histnorm=histnorm,
nbins=nbins,
range_x=(0, xmax) if xmax else None,
)
fig = update_axis_font(fig)
loc.plotly_chart(fig, use_container_width=True)
def draw_bar(bar_data, x, y, loc=st):
fig = px.bar(
bar_data,
x=x,
y=y,
color="split",
color_discrete_sequence=IBM_COLORS,
barmode="group",
hover_data=bar_data.columns,
)
fig = update_axis_font(fig)
loc.plotly_chart(fig, use_container_width=True)
def parse_metrics(metadata, loc):
for split, meta in metadata.items():
for key, val in meta.__dict__.items():
if isinstance(val, int):
loc.metric(label=f"{split}-{key}", value=val)
def parse_counters(metadata):
meta = metadata["train"] # using the training counter to fetch the names
counters = []
for k, v in meta.__dict__.items():
if "counter" in k and len(v) > 0:
counters.append(k)
return counters
# generate the df for histogram
def parse_label_counter(metadata, counter_type):
hist_data = []
for split, m in metadata.items():
metadata_counter = getattr(m, counter_type)
for k, v in metadata_counter.items():
row = {}
row["labels"] = k
row[counter_type] = v
row["split"] = split
hist_data.append(row)
return pd.DataFrame(hist_data)
# load BigBioConfigHelpers
#==================================
logger.info("about to call load_conhelps")
conhelps = load_conhelps()
logger.info("exiting call load_conhelps")
config_name_to_conhelp = {ch.config.name: ch for ch in conhelps}
ds_display_names = sorted(list(set([ch.display_name for ch in conhelps])))
ds_display_name_to_config_names = defaultdict(list)
for ch in conhelps:
ds_display_name_to_config_names[ch.display_name].append(ch.config.name)
# dataset selection
#==================================
st.sidebar.title("Dataset Selection")
ds_display_name = st.sidebar.selectbox("dataset name", ds_display_names, index=0)
config_names = ds_display_name_to_config_names[ds_display_name]
config_name = st.sidebar.selectbox("config name", config_names)
conhelp = config_name_to_conhelp[config_name]
st.header(f"Dataset stats for {ds_display_name}")
@st.cache()
def load_data(conhelp):
metadata = conhelp.get_metadata()
dsd = conhelp.load_dataset()
dsd = dsd.map(
BATCH_MAPPERS_TEXT_FROM_SCHEMA[conhelp.bigbio_schema_caps.lower()],
batched=True)
return dsd, metadata
@st.cache()
def count_vectorize(dsd):
cv = CountVectorizer()
xcvs = {}
dfs_tok_per_samp = []
for split, ds in dsd.items():
xcv = cv.fit_transform(ds['text'])
token_counts = np.asarray(xcv.sum(axis=1)).flatten()
df = pd.DataFrame(token_counts, columns=["tokens per sample"])
df["split"] = split
dfs_tok_per_samp.append(df)
xcvs[split] = xcv
df_tok_per_samp = pd.concat(dfs_tok_per_samp)
return xcvs, df_tok_per_samp
dsd_load_state = st.info(f"Loading {ds_display_name} - {config_name} ...")
dsd, metadata = load_data(conhelp)
dsd_load_state.empty()
cv_load_state = st.info(f"Count Vectorizing {ds_display_name} - {config_name} ...")
xcvs, df_tok_per_samp = count_vectorize(dsd)
cv_load_state.empty()
st.sidebar.subheader(f"BigBIO Schema = {conhelp.bigbio_schema_caps}")
st.sidebar.subheader("Tasks Supported by Dataset")
tasks = conhelp.tasks
tasks = [string.capwords(task.replace("_", " ")) for task in tasks]
st.sidebar.markdown(
"""
{}
""".format(
"\n".join([
f"- {task}" for task in tasks
]))
)
st.sidebar.subheader("Languages")
langs = conhelp.languages
st.sidebar.markdown(
"""
{}
""".format("\n".join([f"- {lang}" for lang in langs]))
)
st.sidebar.subheader("Home Page")
st.sidebar.write(conhelp.homepage)
st.sidebar.subheader("Description")
st.sidebar.write(conhelp.description)
st.sidebar.subheader("Citation")
st.sidebar.markdown(f"""\
```
{conhelp.citation}
````
"""
)
st.sidebar.subheader("Counts")
parse_metrics(metadata, st.sidebar)
# dataframe display
#if "train" in dsd.keys():
# st.subheader("Sample Preview")
# df = pd.DataFrame.from_dict(dsd["train"])
# st.write(df.head(10))
# draw token distribution
st.subheader("Sample Length Distribution")
max_xmax = int(df_tok_per_samp["tokens per sample"].max())
xmax = st.slider("xmax", min_value=0, max_value=max_xmax, value=max_xmax)
histnorms = ['percent', 'probability', 'density', 'probability density', None]
histnorm = st.selectbox("histnorm", histnorms)
draw_histogram(df_tok_per_samp, "tokens per sample", histnorm=histnorm, xmax=xmax, loc=st)
st.subheader("Counter Distributions")
counters = parse_counters(metadata)
counter_type = st.selectbox("counter_type", counters)
label_df = parse_label_counter(metadata, counter_type)
label_max = int(label_df[counter_type].max() - 1)
label_min = int(label_df[counter_type].min())
filter_value = st.slider("minimum cutoff", label_min, label_max)
label_df = label_df[label_df[counter_type] >= filter_value]
# draw bar chart for counter
draw_bar(label_df, "labels", counter_type, st)
st.subheader("Sample Explorer")
split = st.selectbox("split", list(dsd.keys()))
sample_index = st.number_input(
"sample index",
min_value=0,
max_value=len(dsd[split])-1,
value=0,
)
sample = dsd[split][sample_index]
if conhelp.bigbio_schema_caps == "KB":
nlp = spacy.blank("en")
text = sample["text"]
doc = nlp(text)
spans = []
for bb_ent in sample["entities"]:
span = doc.char_span(
bb_ent["offsets"][0][0],
bb_ent["offsets"][0][1],
label=bb_ent["type"],
)
spans.append(span)
doc.spans["sc"] = spans
html = displacy.render(
doc,
style="span",
options={
"colors": {
et: clr for et,clr in zip(
metadata[split].entities_type_counter.keys(),
IBM_COLORS*10
)
}
},
)
style = "<style>mark.entity { display: inline-block }</style>"
st.write(f"{style}{get_html(html)}", unsafe_allow_html=True)
st.write(sample)