TheRock/rocm-libraries/dnn-providers/integration-tests/reference_data_scripts/utilities/graph.py

## Copyright © Advanced Micro Devices, Inc., or its affiliates.
## SPDX-License-Identifier:  MIT

import torch
from tensor import TensorAttributes, dump_data_as_binary, load_data_from_binary
from common import DTypeConverter, NODE_REGISTRY
from batchnorm_inference import BatchnormInference
from itertools import chain
import json
import os


def node_from_dict(node_dict: dict, tensors: dict[int, TensorAttributes]):
    if node_dict["type"] not in NODE_REGISTRY.keys():
        raise RuntimeError("Unsupported node type: " + node_dict["type"])

    return NODE_REGISTRY[node_dict["type"]].from_dict(node_dict, tensors)


class Graph:
    def __init__(
        self,
        nodes,
        dtype: torch.dtype = None,
        io_data_type: torch.dtype = None,
        compute_data_type: torch.dtype = None,
        intermediate_data_type: torch.dtype = None,
        name: str = "",
    ):
        if dtype != None and (
            io_data_type == None
            and compute_data_type == None
            and intermediate_data_type == None
        ):
            io_data_type = compute_data_type = intermediate_data_type = dtype
        elif dtype != None:
            raise ValueError(
                "type must be set, or all of io_data_type, compute_data_type and intermediate_data_type must be"
            )

        self.nodes = nodes
        self.tensors = dict()
        for node in nodes:
            inputs = node.inputs.__dict__.values()
            outputs = node.outputs.__dict__.values()
            tensors = filter(
                lambda x: type(x) == TensorAttributes, chain(inputs, outputs)
            )
            self.tensors.update({tensor.uid: tensor for tensor in tensors})

        self.io_data_type = io_data_type
        self.compute_data_type = compute_data_type
        self.intermediate_data_type = intermediate_data_type
        self.name = name

    # TODO Implement execute for more complicated graphs
    # def execute():

    def as_dict(self):
        tensors = [tensor.as_dict() for tensor in self.tensors.values()]
        return {
            "nodes": [node.as_dict() for node in self.nodes],
            "tensors": tensors,
            "io_data_type": DTypeConverter.to_string(self.io_data_type),
            "compute_data_type": DTypeConverter.to_string(self.compute_data_type),
            "intermediate_data_type": DTypeConverter.to_string(
                self.intermediate_data_type
            ),
            "name": self.name,
        }

    def save(self, base_filename: str):
        with open(base_filename + ".json", "w") as file:
            json.dump(self.as_dict(), file)
        for tensor in self.tensors.values():
            dump_data_as_binary(
                "{}.tensor{}.bin".format(base_filename, tensor.uid), tensor
            )

    @staticmethod
    def from_file(base_path: str):
        with open(base_path + ".json") as file:
            d = json.load(file)

        tensor_list = [
            TensorAttributes.from_dict(tensorDict) for tensorDict in d["tensors"]
        ]
        tensor_map = {tensor.uid: tensor for tensor in tensor_list}

        nodes = [node_from_dict(node_dict, tensor_map) for node_dict in d["nodes"]]

        for uid, tensor in tensor_map.items():
            filepath = "{}.tensor{}".format(base_path, uid)
            if os.path.exists(filepath):
                load_data_from_binary(filepath, tensor)

        io_data_type = DTypeConverter.from_string(d["io_data_type"])
        compute_data_type = DTypeConverter.from_string(d["compute_data_type"])
        intermediate_data_type = DTypeConverter.from_string(d["intermediate_data_type"])

        return Graph(
            nodes,
            io_data_type=io_data_type,
            compute_data_type=compute_data_type,
            intermediate_data_type=intermediate_data_type,
            name=d["name"],
        )