src/display/utils.py

from dataclasses import dataclass, make_dataclass
from enum import Enum

from src.tasks import Categories, Tasks


def fields(raw_class):
    return [v for k, v in raw_class.__dict__.items() if k[:2] != "__" and k[-2:] != "__"]


# These classes are for user facing column names,
# to avoid having to change them all around the code
# when a modif is needed
@dataclass
class ColumnContent:
    name: str
    type: str
    displayed_by_default: bool
    hidden: bool = False
    never_hidden: bool = False
    task_language: str = ""
    task_domain: str = ""
    dummy: bool = False


## Leaderboard columns
auto_eval_column_dict = []
defaults_dict = {}
# Init
auto_eval_column_dict.append(["model_type_symbol", ColumnContent])
defaults_dict["model_type_symbol"] = ColumnContent("T", "str", True, never_hidden=True)
auto_eval_column_dict.append(["model", ColumnContent])
defaults_dict["model"] = ColumnContent("Model", "markdown", True, never_hidden=True)
# Scores: by default only the total average and average by category are displayed
auto_eval_column_dict.append(["average", ColumnContent])
defaults_dict["average"] = ColumnContent(
    name="Avg Performance", type="number", displayed_by_default=True, task_domain="average"
)
# Eval time sum columns
eval_time_sum_configs = [
    ("sum_t", "Sum t", ""),
    ("sum_t_es", "Sum t ES", "ES"),
    ("sum_t_ca", "Sum t CA", "CA"),
    ("sum_t_eu", "Sum t EU", "EU"),
    ("sum_t_gl", "Sum t GL", "GL"),
    ("sum_t_va", "Sum t VA", "VA"),
    ("sum_t_pt", "Sum t PT", "PT"),
]
for field_name, col_name, lang in eval_time_sum_configs:
    auto_eval_column_dict.append([field_name, ColumnContent])
    defaults_dict[field_name] = ColumnContent(
        name=col_name, type="number", displayed_by_default=True, task_language=lang, task_domain="average"
    )
# CO2 sum columns
co2_sum_configs = [
    ("sum_kg_co2", "Sum kg CO2", ""),
    ("sum_kg_co2_es", "Sum kg CO2 ES", "ES"),
    ("sum_kg_co2_ca", "Sum kg CO2 CA", "CA"),
    ("sum_kg_co2_eu", "Sum kg CO2 EU", "EU"),
    ("sum_kg_co2_gl", "Sum kg CO2 GL", "GL"),
    ("sum_kg_co2_va", "Sum kg CO2 VA", "VA"),
    ("sum_kg_co2_pt", "Sum kg CO2 PT", "PT"),
]
for field_name, col_name, lang in co2_sum_configs:
    auto_eval_column_dict.append([field_name, ColumnContent])
    defaults_dict[field_name] = ColumnContent(
        name=col_name, type="number", displayed_by_default=True, task_language=lang, task_domain="average"
    )
for category in Categories:
    auto_eval_column_dict.append([category.name, ColumnContent])
    defaults_dict[category.name] = ColumnContent(
        name=category.value.col_name,
        type="number",
        displayed_by_default=True,
        task_language=category.value.language.name,
        task_domain="average",
    )
for task in Tasks:
    auto_eval_column_dict.append([task.name, ColumnContent])
    defaults_dict[task.name] = ColumnContent(
        name=task.value.col_name,
        type="number",
        displayed_by_default=True,
        task_language=task.value.language.name,
        task_domain=task.value.domain.name,
    )
# Model information
auto_eval_column_dict.append(["model_type", ColumnContent])
defaults_dict["model_type"] = ColumnContent("Type", "str", False)
auto_eval_column_dict.append(["architecture", ColumnContent])
defaults_dict["architecture"] = ColumnContent("Architecture", "str", False)
auto_eval_column_dict.append(["weight_type", ColumnContent])
defaults_dict["weight_type"] = ColumnContent("Weight type", "str", False, hidden=True)
auto_eval_column_dict.append(["precision", ColumnContent])
defaults_dict["precision"] = ColumnContent("Precision", "str", False)
auto_eval_column_dict.append(["license", ColumnContent])
defaults_dict["license"] = ColumnContent("Hub License", "str", False)
auto_eval_column_dict.append(["params", ColumnContent])
defaults_dict["params"] = ColumnContent("#Params (B)", "number", False)
auto_eval_column_dict.append(["likes", ColumnContent])
defaults_dict["likes"] = ColumnContent("Hub ❤️", "number", False)
auto_eval_column_dict.append(["still_on_hub", ColumnContent])
defaults_dict["still_on_hub"] = ColumnContent("Available on the hub", "bool", False)
auto_eval_column_dict.append(["revision", ColumnContent])
defaults_dict["revision"] = ColumnContent("Model sha", "str", False, hidden=False)
# Dummy column for the search bar (hidden by the custom CSS)
auto_eval_column_dict.append(["dummy", ColumnContent])
defaults_dict["dummy"] = ColumnContent("model_name_for_query", "str", False, dummy=True)

# We use make dataclass to dynamically fill the scores from Tasks
# Create without defaults to avoid mutable default issue
AutoEvalColumn = make_dataclass("AutoEvalColumn", auto_eval_column_dict, frozen=True)

# Set class attributes (these are accessed as AutoEvalColumn.field_name, not as instance attributes)
for field_name, default_value in defaults_dict.items():
    setattr(AutoEvalColumn, field_name, default_value)


@dataclass
class TaskDetails:
    name: str
    display_name: str = ""
    symbol: str = ""  # emoji


## For the queue columns in the submission tab
@dataclass(frozen=True)
class EvalQueueColumn:  # Queue column
    model = ColumnContent("model", "markdown", True)
    revision = ColumnContent("revision", "str", True)
    private = ColumnContent("private", "bool", True)
    precision = ColumnContent("precision", "str", True)
    weight_type = ColumnContent("weight_type", "str", "Original")
    status = ColumnContent("status", "str", True)


## All the model information that we might need
@dataclass
class ModelDetails:
    name: str
    display_name: str = ""
    symbol: str = ""  # emoji


class ModelType(Enum):
    PT = ModelDetails(name="pretrained", symbol="🟢")
    FT = ModelDetails(name="fine-tuned", symbol="🔶")
    IFT = ModelDetails(name="instruction-tuned", symbol="⭕")
    RL = ModelDetails(name="RL-tuned", symbol="🟦")
    Unknown = ModelDetails(name="", symbol="?")

    def to_str(self, separator=" "):
        return f"{self.value.symbol}{separator}{self.value.name}"

    @staticmethod
    def from_str(type):
        if "fine-tuned" in type or "🔶" in type:
            return ModelType.FT
        if "pretrained" in type or "🟢" in type:
            return ModelType.PT
        if "RL-tuned" in type or "🟦" in type:
            return ModelType.RL
        if "instruction-tuned" in type or "⭕" in type:
            return ModelType.IFT
        return ModelType.Unknown


class WeightType(Enum):
    Adapter = ModelDetails("Adapter")
    Original = ModelDetails("Original")
    Delta = ModelDetails("Delta")


class Precision(Enum):
    float16 = ModelDetails("float16")
    bfloat16 = ModelDetails("bfloat16")
    float32 = ModelDetails("float32")
    qt_8bit = ModelDetails("8bit")
    qt_4bit = ModelDetails("4bit")
    qt_GPTQ = ModelDetails("GPTQ")
    Unknown = ModelDetails("?")

    def from_str(precision):
        if precision in ["torch.float16", "float16"]:
            return Precision.float16
        if precision in ["torch.bfloat16", "bfloat16"]:
            return Precision.bfloat16
        if precision in ["float32"]:
            return Precision.float32
        if precision in ["8bit"]:
            return Precision.qt_8bit
        if precision in ["4bit"]:
            return Precision.qt_4bit
        if precision in ["GPTQ", "None"]:
            return Precision.qt_GPTQ
        return Precision.Unknown


# Column selection
COLS = [c.name for c in fields(AutoEvalColumn) if not c.hidden]

EVAL_COLS = [c.name for c in fields(EvalQueueColumn)]
EVAL_TYPES = [c.type for c in fields(EvalQueueColumn)]

BENCHMARK_COLS = [t.value.col_name for t in Tasks]