app.py

from huggingface_hub import HfApi
import pandas as pd
import os
import streamlit as st
import altair as alt
import numpy as np
import datetime
from huggingface_hub import Repository

from transformers.models.auto.configuration_auto import CONFIG_MAPPING_NAMES
from transformers.models.auto.modeling_auto import (
    MODEL_FOR_CTC_MAPPING_NAMES,
    MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES,
    MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES,
    MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING_NAMES,
    MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES,
    MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES,
    MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES,
    MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES,
)

audio_models = MODEL_FOR_CTC_MAPPING_NAMES.keys() + MODEL_FOR_AUDIO_CLASSIFICATION_MAPPING_NAMES.keys() + MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES.keys()

vision_models = MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES.keys() + MODEL_FOR_INSTANCE_SEGMENTATION_MAPPING_NAMES.keys() + MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING_NAMES.keys() + MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING_NAMES.keys() + MODEL_FOR_CAUSAL_IMAGE_MODELING_MAPPING_NAMES.keys() + MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING_NAMES.keys()

today = datetime.date.today()
year, week, _ = today.isocalendar()

DATASET_REPO_URL = (
    "https://huggingface.co/datasets/patrickvonplaten/model-archs-downloads-space-data"
)
DATA_FILENAME = f"data_{week}_{year}.csv"
DATA_FILE = os.path.join("data", DATA_FILENAME)


def retrieve_model_stats():
    hf_api = HfApi()
    all_stats = {}
    total_downloads = 0

    for model_name in list(CONFIG_MAPPING_NAMES.keys()):
        if model_name in audio_models:
            modality = "audio"
        elif model_name in vision_models:
            modality = "vision"
        else:
            modality = "text"

        model_stats = {
            "num_downloads": 0,
            "%_of_all_downloads": 0,
            "num_models": 0,
            "download_per_model": 0,
            "modality": modality,
        }
        models = hf_api.list_models(filter=model_name)

        model_stats["num_models"] = len(models)
        model_stats["num_downloads"] = sum(
            [m.downloads for m in models if hasattr(m, "downloads")]
        )
        if len(models) > 0:
            model_stats["download_per_model"] = round(
                model_stats["num_downloads"] / len(models), 2
            )
        total_downloads += model_stats["num_downloads"]

        # save in overall dict
        all_stats[model_name] = model_stats

    for model_name in list(CONFIG_MAPPING_NAMES.keys()):
        all_stats[model_name]["%_of_all_downloads"] = (
            round(all_stats[model_name]["num_downloads"] / total_downloads, 5) * 100
        )  # noqa: E501
        downloads = all_stats[model_name]["num_downloads"]
        all_stats[model_name]["num_downloads"] = f"{downloads:,}"

    sorted_results = dict(
        reversed(sorted(all_stats.items(), key=lambda d: d[1]["%_of_all_downloads"]))
    )
    dataframe = pd.DataFrame.from_dict(sorted_results, orient="index")

    # give header to model names
    result = "model_names" + dataframe.to_csv()
    return result


repo = Repository(local_dir="data", clone_from=DATASET_REPO_URL)

if not os.path.isfile(DATA_FILE):
    print("Create datafile...")
    result = retrieve_model_stats()

    if not os.path.isfile(DATA_FILE):
        with open(DATA_FILE, "w") as f:
            f.write(result)

        commit_url = repo.push_to_hub()
        print(commit_url)

with open(DATA_FILE, "r") as f:
    dataframe = pd.read_csv(DATA_FILE)

int_downloads = np.array(
    [int(x.replace(",", "")) for x in dataframe["num_downloads"].values]
)

st.title(f"Transformers stats for year {year} and week {week}")
# print top 20 downloads
source = pd.DataFrame(
    {
        "Number of total downloads": int_downloads[:20],
        "Model architecture name": dataframe["model_names"].values[:20],
    }
)
bar_chart = (
    alt.Chart(source)
    .mark_bar()
    .encode(
        y="Number of total downloads",
        x=alt.X("Model architecture name", sort=None),
    )
)
st.title("Top 20 downloads last 30 days")
st.altair_chart(bar_chart, use_container_width=True)

# print bottom 20 downloads
source = pd.DataFrame(
    {
        "Number of total downloads": int_downloads[-20:],
        "Model architecture name": dataframe["model_names"].values[-20:],
    }
)
bar_chart = (
    alt.Chart(source)
    .mark_bar()
    .encode(
        y="Number of total downloads",
        x=alt.X("Model architecture name", sort=None),
    )
)
st.title("Bottom 20 downloads last 30 days")
st.altair_chart(bar_chart, use_container_width=True)

# print all stats
st.title("All stats last 30 days")
st.table(dataframe)