import pandas as pd import streamlit as st from huggingface_hub import HfApi from utils import ascending_metrics, metric_ranges, LANGUAGES import numpy as np from st_aggrid import AgGrid, GridOptionsBuilder, JsCode, ColumnsAutoSizeMode from os.path import exists import threading st.set_page_config(layout="wide") def get_model_infos(): api = HfApi() model_infos = api.list_models(filter="model-index", cardData=True) return model_infos 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, 4) if isinstance(value, float) else None return value def parse_metrics_rows(meta, only_verified=False): 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 not isinstance(result, dict) or "dataset" not in result or "metrics" not in result or "type" not in result["dataset"]: continue dataset = result["dataset"]["type"] if dataset == "": continue row = {"dataset": dataset, "split": "-unspecified-", "config": "-unspecified-"} if "split" in result["dataset"]: row["split"] = result["dataset"]["split"] if "config" in result["dataset"]: row["config"] = result["dataset"]["config"] no_results = True incorrect_results = False for metric in result["metrics"]: name = metric["type"].lower().strip() if name in ("model_id", "dataset", "split", "config", "pipeline_tag", "only_verified"): # Metrics are not allowed to be named "dataset", "split", "config", "pipeline_tag" continue value = parse_metric_value(metric.get("value", None)) if value is None: continue if name in row: new_metric_better = value < row[name] if name in ascending_metrics else value > row[name] if name not in row or new_metric_better: # overwrite the metric if the new value is better. if only_verified: if "verified" in metric and metric["verified"]: no_results = False row[name] = value if name in metric_ranges: if value < metric_ranges[name][0] or value > metric_ranges[name][1]: incorrect_results = True else: no_results = False row[name] = value if name in metric_ranges: if value < metric_ranges[name][0] or value > metric_ranges[name][1]: incorrect_results = True if no_results or incorrect_results: continue yield row @st.cache(ttl=0) def get_data_wrapper(): def get_data(dataframe=None, verified_dataframe=None): data = [] verified_data = [] print("getting model infos") model_infos = get_model_infos() print("got model infos") for model_info in model_infos: meta = model_info.cardData if meta is None: continue for row in parse_metrics_rows(meta): if row is None: continue row["model_id"] = model_info.id row["pipeline_tag"] = model_info.pipeline_tag row["only_verified"] = False data.append(row) for row in parse_metrics_rows(meta, only_verified=True): if row is None: continue row["model_id"] = model_info.id row["pipeline_tag"] = model_info.pipeline_tag row["only_verified"] = True 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() st.markdown("# 🤗 Whisper Event: Final Leaderboard") query_params = st.experimental_get_query_params() if "first_query_params" not in st.session_state: st.session_state.first_query_params = query_params first_query_params = st.session_state.first_query_params default_config = first_query_params.get("config", [None])[0] default_metric = first_query_params.get("metric", [None])[0] only_verified_results = False task = "automatic-speech-recognition" dataset = "mozilla-foundation/common_voice_11_0" split = "test" dataframe = dataframe[dataframe.only_verified == only_verified_results] current_query_params = {"dataset": [dataset], "only_verified": [int(only_verified_results)], "task": [task], "split": [split]} st.experimental_set_query_params(**current_query_params) dataset_df = dataframe[dataframe.dataset == dataset] dataset_df = dataset_df[dataset_df.split == split] dataset_df = dataset_df.dropna(axis="columns", how="all") selectable_metrics = ["wer", "cer"] sorting_metric = "wer" selectable_configs = list(set(dataset_df["config"])) selectable_configs.sort(key=lambda name: name.lower()) selectable_configs.remove("-unspecified-") selectable_configs = [config for config in selectable_configs if config in LANGUAGES] visual_configs = [f"{config}: {LANGUAGES[config]}" for config in selectable_configs] dataset_df = dataset_df[["config", "model_id"] + selectable_metrics] all_ds = [] for config in selectable_configs: dataset_df_ds = dataset_df[dataset_df.config == config] dataset_df_ds = dataset_df_ds.dropna(thresh=2) # Want at least two non-na values (one for model_id and one for a metric) # Make the default metric appear right after model names and dataset names cols = dataset_df_ds.columns.tolist() cols.remove(sorting_metric) sorting_metric_index = 1 if dataset != "-any-" else 2 cols = cols[:sorting_metric_index] + [sorting_metric] + cols[sorting_metric_index:] dataset_df_ds = dataset_df_ds[cols] # Sort the leaderboard, giving the sorting metric highest priority and then ordering by other metrics in the case of equal values. dataset_df_ds = dataset_df_ds.sort_values(by=cols[sorting_metric_index:], ascending=[metric in ascending_metrics for metric in cols[sorting_metric_index:]]) dataset_df_ds = dataset_df_ds.replace(np.nan, '-') all_ds.append(dataset_df_ds.iloc[0]) all_ds = pd.DataFrame(all_ds, columns=["config", "model_id", "wer", "cer"]) language_names = [LANGUAGES[config] for config in selectable_configs] all_ds.insert(1, "language", language_names) # Make the leaderboard gb = GridOptionsBuilder.from_dataframe(all_ds) gb.configure_default_column(sortable=False) gb.configure_column( "model_id", cellRenderer=JsCode('''function(params) {return ''+params.value+''}'''), ) for name in selectable_metrics: gb.configure_column(name, type=["numericColumn","numberColumnFilter","customNumericFormat"], precision=2, aggFunc='sum') gb.configure_column( sorting_metric, sortable=True, #cellStyle=JsCode('''function(params) { return {'backgroundColor': '#FFD21E'}}''') ) go = gb.build() fit_columns = len(all_ds.columns) < 10 AgGrid(all_ds, gridOptions=go, height=28*len(all_ds) + (35 if fit_columns else 41), allow_unsafe_jscode=True, enable_enterprise_modules=False, columns_auto_size_mode=ColumnsAutoSizeMode.FIT_CONTENTS)