Spaces:
Running
Running
Tristan Thrush
fixed cuttoff issue for wide leaderboards, made leaderboard data updating asynchronous, made streamlit set the url to match the selected dataset
80f2297
| import pandas as pd | |
| from tqdm.auto import tqdm | |
| import streamlit as st | |
| from huggingface_hub import HfApi, hf_hub_download | |
| from huggingface_hub.repocard import metadata_load | |
| from ascending_metrics import ascending_metrics | |
| import numpy as np | |
| from st_aggrid import AgGrid, GridOptionsBuilder, JsCode | |
| from os.path import exists | |
| import threading | |
| def get_model_ids(): | |
| api = HfApi() | |
| models = api.list_models(filter="model-index") | |
| model_ids = [x.modelId for x in models] | |
| return model_ids | |
| def get_metadata(model_id): | |
| try: | |
| readme_path = hf_hub_download(model_id, filename="README.md") | |
| return metadata_load(readme_path) | |
| except Exception: | |
| # 404 README.md not found or problem loading it | |
| return None | |
| def parse_metric_value(value): | |
| if isinstance(value, str): | |
| "".join(value.split("%")) | |
| try: | |
| value = float(value) | |
| except: # noqa: E722 | |
| value = None | |
| elif isinstance(value, list): | |
| if len(value) > 0: | |
| value = value[0] | |
| else: | |
| value = None | |
| value = round(value, 2) if isinstance(value, float) else None | |
| return value | |
| def parse_metrics_rows(meta): | |
| if not isinstance(meta["model-index"], list) or len(meta["model-index"]) == 0 or "results" not in meta["model-index"][0]: | |
| return None | |
| for result in meta["model-index"][0]["results"]: | |
| if "dataset" not in result or "metrics" not in result or "type" not in result["dataset"]: | |
| continue | |
| dataset = result["dataset"]["type"] | |
| if "args" not in result["dataset"]: | |
| continue | |
| row = {"dataset": dataset} | |
| for metric in result["metrics"]: | |
| type = metric["type"].lower().strip() | |
| value = parse_metric_value(metric.get("value", None)) | |
| if value is None: | |
| continue | |
| if type not in row or value < row[type]: | |
| # overwrite the metric if the new value is lower (e.g. with LM) | |
| row[type] = value | |
| yield row | |
| def get_data_wrapper(): | |
| def get_data(): | |
| data = [] | |
| model_ids = get_model_ids() | |
| for model_id in tqdm(model_ids): | |
| meta = get_metadata(model_id) | |
| if meta is None: | |
| continue | |
| for row in parse_metrics_rows(meta): | |
| if row is None: | |
| continue | |
| row["model_id"] = model_id | |
| data.append(row) | |
| dataframe = pd.DataFrame.from_records(data) | |
| dataframe.to_pickle("cache.pkl") | |
| if exists("cache.pkl"): | |
| # If we have saved the results previously, call an asynchronous process | |
| # to fetch the results and update the saved file. Don't make users wait | |
| # while we fetch the new results. Instead, display the old results for | |
| # now. The new results should be loaded when this method | |
| # is called again. | |
| dataframe = pd.read_pickle("cache.pkl") | |
| t = threading.Thread(name='get_data procs', target=get_data) | |
| t.start() | |
| else: | |
| # We have to make the users wait during the first startup of this app. | |
| get_data() | |
| dataframe = pd.read_pickle("cache.pkl") | |
| return dataframe | |
| dataframe = get_data_wrapper() | |
| selectable_datasets = list(set(dataframe.dataset.tolist())) | |
| st.markdown("# 🤗 Leaderboards") | |
| query_params = st.experimental_get_query_params() | |
| default_dataset = "common_voice" | |
| if "dataset" in query_params: | |
| if len(query_params["dataset"]) > 0 and query_params["dataset"][0] in selectable_datasets: | |
| default_dataset = query_params["dataset"][0] | |
| dataset = st.sidebar.selectbox( | |
| "Dataset", | |
| selectable_datasets, | |
| index=selectable_datasets.index(default_dataset), | |
| ) | |
| st.experimental_set_query_params(**{"dataset": [dataset]}) | |
| dataset_df = dataframe[dataframe.dataset == dataset] | |
| dataset_df = dataset_df.dropna(axis="columns", how="all") | |
| selectable_metrics = list(filter(lambda column: column not in ("model_id", "dataset"), dataset_df.columns)) | |
| default_metric = st.sidebar.radio( | |
| "Default Metric", | |
| selectable_metrics, | |
| ) | |
| dataset_df = dataset_df.filter(["model_id"] + selectable_metrics) | |
| dataset_df = dataset_df.dropna(thresh=2) # Want at least two non-na values (one for model_id and one for a metric). | |
| dataset_df = dataset_df.sort_values(by=default_metric, ascending=default_metric in ascending_metrics) | |
| dataset_df = dataset_df.replace(np.nan, '-') | |
| st.markdown( | |
| "Please click on the model's name to be redirected to its model card." | |
| ) | |
| st.markdown( | |
| "Want to beat the leaderboard? Don't see your model here? Simply request an automatic evaluation [here](https://huggingface.co/spaces/autoevaluate/autoevaluate)." | |
| ) | |
| # Make the default metric appear right after model names | |
| cols = dataset_df.columns.tolist() | |
| cols.remove(default_metric) | |
| cols = cols[:1] + [default_metric] + cols[1:] | |
| dataset_df = dataset_df[cols] | |
| # Make the leaderboard | |
| gb = GridOptionsBuilder.from_dataframe(dataset_df) | |
| gb.configure_column( | |
| "model_id", | |
| cellRenderer=JsCode('''function(params) {return '<a target="_blank" href="https://huggingface.co/'+params.value+'">'+params.value+'</a>'}'''), | |
| ) | |
| for name in selectable_metrics: | |
| gb.configure_column(name, type=["numericColumn","numberColumnFilter","customNumericFormat"], precision=2, aggFunc='sum') | |
| gb.configure_column( | |
| default_metric, | |
| cellStyle=JsCode('''function(params) { return {'backgroundColor': '#FFD21E'}}''') | |
| ) | |
| go = gb.build() | |
| AgGrid(dataset_df, gridOptions=go, allow_unsafe_jscode=True, fit_columns_on_grid_load=True) | |