Benjamin Bossan
Add possibility to load existing model card
433130f
# HF space creator starting from an sklearn model
from __future__ import annotations
import base64
import glob
import io
import os
import pickle
import re
import shutil
from pathlib import Path
from tempfile import mkdtemp
import pandas as pd
import sklearn
import streamlit as st
from huggingface_hub import hf_hub_download
from sklearn.base import BaseEstimator
import skops.io as sio
from skops import card, hub_utils
st.set_page_config(layout="wide")
st.title("Skops space creator for sklearn")
PLACEHOLDER = "[More Information Needed]"
PLOT_PREFIX = "__plot__:"
# store session state
if "custom_sections" not in st.session_state:
st.session_state.custom_sections = {}
# the tmp_path is used to upload the sklearn model to
tmp_path = Path(mkdtemp(prefix="skops-"))
# the hf_path is the actual repo used for init()
hf_path = Path(mkdtemp(prefix="skops-"))
# a hacky way to "persist" custom sections
CUSTOM_SECTIONS_CACHE_FILE = ".custom-sections.json"
def _clear_custom_section_cache():
st.session_state.custom_sections.clear()
def _remove_custom_section(key):
section_names = list(st.session_state.custom_sections.keys())
for section_name in section_names:
if (
(section_name == key)
or section_name.startswith(key + "/")
or section_name.startswith(key + " /")
):
del st.session_state.custom_sections[section_name]
def _clear_repo(path):
for file_path in glob.glob(str(Path(path) / "*")):
if os.path.isfile(file_path) or os.path.islink(file_path):
os.unlink(file_path)
elif os.path.isdir(file_path):
shutil.rmtree(file_path)
def _write_plot(plot_name, plot_file):
with open(plot_name, "wb") as f:
f.write(plot_file)
def init_repo():
_clear_repo(hf_path)
try:
file_name = tmp_path / "model.skops"
sio.dump(model, file_name)
reqs = [r.strip().rstrip(",") for r in requirements.splitlines()]
hub_utils.init(
model=file_name,
dst=hf_path,
task=task,
data=data,
requirements=reqs,
)
except Exception as exc:
print("Uh oh, something went wrong when initializing the repo:", exc)
def load_model():
if model_file is None:
return
bytes_data = model_file.getvalue()
model = pickle.loads(bytes_data)
assert isinstance(model, BaseEstimator), "model must be an sklearn model"
return model
def load_data():
if data_file is None:
return
bytes_data = io.BytesIO(data_file.getvalue())
df = pd.read_csv(bytes_data)
return df
def _parse_metrics(metrics):
metrics_table = {}
for line in metrics.splitlines():
line = line.strip()
name, _, val = line.partition("=")
try:
# try to coerce to float but don't error if it fails
val = float(val.strip())
except ValueError:
pass
metrics_table[name.strip()] = val
return metrics_table
def _load_model_card_from_repo(repo_id: str) -> Card:
path = hf_hub_download(repo_id, "README.md")
return card.parse_modelcard(path)
def _create_model_card():
init_repo()
if model_card_repo: # load existing model card
model_card = _load_model_card_from_repo(model_card_repo)
else: # create new model card
metadata = card.metadata_from_config(hf_path)
model_card = card.Card(model=model, metadata=metadata)
if model_description:
model_card.add(**{"Model description": model_description})
if intended_uses:
model_card.add(
**{"Model description/Intended uses & limitations": intended_uses}
)
if metrics:
metrics_table = _parse_metrics(metrics)
model_card.add_metrics(**metrics_table)
if authors:
model_card.add(**{"Model Card Authors": authors})
if contact:
model_card.add(**{"Model Card Contact": contact})
if citation:
model_card.add(**{"Citation": citation})
if st.session_state.custom_sections:
for key, val in st.session_state.custom_sections.items():
if not key:
continue
if key.startswith(PLOT_PREFIX):
key = key[len(PLOT_PREFIX) :] # noqa
model_card.add_plot(**{key: val})
else:
model_card.add(**{key: val})
return model_card
def _process_card_for_rendering(rendered: str) -> tuple[str, str]:
idx = rendered[1:].index("\n---") + 1
metadata = rendered[3:idx]
rendered = rendered[idx + 4 :] # noqa: E203
# below is a hack to display the images in streamlit
# https://discuss.streamlit.io/t/image-in-markdown/13274/10 The problem is
# that streamlit does not display images in markdown, so we need to replace
# them with html. However, we only want that in the rendered markdown, not
# in the card that is produced for the hub
def markdown_images(markdown):
# example image markdown:
# ![Test image](images/test.png "Alternate text")
images = re.findall(
r'(!\[(?P<image_title>[^\]]+)\]\((?P<image_path>[^\)"\s]+)\s*([^\)]*)\))',
markdown,
)
return images
def img_to_bytes(img_path):
img_bytes = Path(img_path).read_bytes()
encoded = base64.b64encode(img_bytes).decode()
return encoded
def img_to_html(img_path, img_alt):
img_format = img_path.split(".")[-1]
img_html = (
f'<img src="data:image/{img_format.lower()};'
f'base64,{img_to_bytes(img_path)}" '
f'alt="{img_alt}" '
'style="max-width: 100%;">'
)
return img_html
def markdown_insert_images(markdown):
images = markdown_images(markdown)
for image in images:
image_markdown = image[0]
image_alt = image[1]
image_path = image[2]
markdown = markdown.replace(
image_markdown, img_to_html(image_path, image_alt)
)
return markdown
rendered_with_img = markdown_insert_images(rendered)
return metadata, rendered_with_img
def display_model_card(model_card):
if not model_card:
return
rendered = model_card.render()
metadata, rendered = _process_card_for_rendering(rendered)
# idx = rendered[1:].index("\n---") + 1
# metadata = rendered[3:idx]
# rendered = rendered[idx + 4 :] # noqa: E203
# strip metadata
with st.expander("show metadata"):
st.text(metadata)
st.markdown(rendered, unsafe_allow_html=True)
def download_model_card(model_card):
if model_card is not None:
return model_card.render()
return ""
def add_custom_section():
# this is required to "refresh" these variables...
global section_name, section_content
section_name = st.session_state.key_section_name
section_content = st.session_state.key_section_content
if not section_name or not section_content:
return
st.session_state.custom_sections[section_name] = section_content
def add_custom_plot():
# this is required to "refresh" these variables...
global section_name, section_content
plot_name = st.session_state.key_plot_name
plot_file = st.session_state.key_plot_file
if not plot_name or not plot_file:
return
# store plot in temp repo
file_name = plot_file.name.replace(" ", "_")
file_path = str(tmp_path / file_name)
with open(file_path, "wb") as f:
f.write(plot_file.getvalue())
st.session_state.custom_sections[str(PLOT_PREFIX + plot_name)] = file_path
with st.sidebar:
# This contains every element required to edit the model card
model = None
data = None
section_name = None
section_content = None
st.title("Model Card Editor")
model_file = st.file_uploader("Upload a model*", on_change=load_model)
data_file = st.file_uploader(
"Upload X data (csv)*", type=["csv"], on_change=load_data
)
task = st.selectbox(
label="Choose the task type*",
options=[
"tabular-classification",
"tabular-regression",
"text-classification",
"text-regression",
],
on_change=init_repo,
)
requirements = st.text_area(
label="Requirements*",
value=f"scikit-learn=={sklearn.__version__}\n",
on_change=init_repo,
)
if model_file is not None:
model = load_model()
if data_file is not None:
data = load_data()
if model is not None and data is not None:
init_repo()
model_card_repo = st.text_input(
"Optional: HF repo to load model card from (e.g. 'gpt2'), "
"leave empty to use default skops template",
value="",
)
# DEFAULT SKOPS SECTIONS
if not model_card_repo:
model_description = st.text_input("Model description", value=PLACEHOLDER)
intended_uses = st.text_area(
"Intended uses & limitations", height=2, value=PLACEHOLDER
)
metrics = st.text_area("Metrics (e.g. 'accuracy = 0.95'), one metric per line")
authors = st.text_area(
"Authors",
value="This model card is written by following authors:\n\n" + PLACEHOLDER,
)
contact = st.text_area(
"Contact",
value="You can contact the model card authors through following channels:\n\n"
+ PLACEHOLDER,
)
citation = st.text_area(
"Citation",
value="Below you can find information related to citation.\n\nBibTex:\n\n```\n"
+ PLACEHOLDER
+ "\n```",
height=5,
)
else:
model_description = None
intended_uses = None
metrics = None
authors = None
contact = None
citation = None
# ADD A CUSTOM SECTIONS
with st.form("custom-section", clear_on_submit=True):
section_name = st.text_input(
"Section name (use '/' for subsections, e.g. 'Model description/My new"
" section')",
key="key_section_name",
)
section_content = st.text_area(
"Content of the new section", key="key_section_content"
)
submit_new_section = st.form_submit_button(
"Create new section", on_click=add_custom_section
)
# ADD A PLOT
with st.form("custom-plots", clear_on_submit=True):
plot_name = st.text_input(
"Section name (use '/' for subsections, e.g. 'Model description/My new"
" plot')",
key="key_plot_name",
)
plot_file = st.file_uploader("Upload a figure*", key="key_plot_file")
submit_new_plot = st.form_submit_button("Add plot", on_click=add_custom_plot)
for key in st.session_state.custom_sections:
if not key:
continue
if key.startswith(PLOT_PREFIX):
st.button(
f"Remove plot '{key[len(PLOT_PREFIX):]}'",
on_click=_remove_custom_section,
args=(key,),
)
else:
st.button(
f"Remove section '{key}'", on_click=_remove_custom_section, args=(key,)
)
if st.session_state.custom_sections:
st.button(
f"Remove all ({len(st.session_state.custom_sections)}) custom elements",
on_click=_clear_custom_section_cache,
)
model_card = None
if model is None:
st.text("*add a model to render the model card*")
if data is None:
st.text("*add data to render the model card")
if (model is not None) and (data is not None):
model_card = _create_model_card()
# this contains the rendered model card
rendered = download_model_card(model_card)
if rendered:
st.download_button(label="Download model card (markdown format)", data=rendered)
display_model_card(model_card)