Function dump_internal_results

ppq/api/fsys.py:197–235

(
    graph: BaseGraph, dataloader: Iterable,
    dump_dir: str, executing_device: str, sample: bool = True)

def dump_internal_results(
    graph: BaseGraph, dataloader: Iterable,
    dump_dir: str, executing_device: str, sample: bool = True):

    i_dir  = os.path.join(dump_dir, 'inputs')
    o_dir  = os.path.join(dump_dir, 'outputs')

    create_dir(i_dir)
    create_dir(o_dir)

    # 找出所有量化点，抽出所有中间结果.
    for var in graph.variables.values():
        if isinstance(var, QuantableVariable):
            if (var.source_op_config is not None and
                var.source_op_config.state == QuantizationStates.ACTIVATED):
                graph.outputs[var.name] = var # 直接标记为网络输出

    executor = TorchExecutor(graph, device=executing_device)
    for batch_idx, batch in tqdm(enumerate(dataloader),
                                 total=len(dataloader), desc='Dumping Results ...'):
        batch = batch.to(executing_device)
        outputs = executor.forward(batch)

        create_dir(os.path.join(o_dir, str(batch_idx)))
        for name, output in zip(graph.outputs, outputs):

            # 如果数字太多就抽样
            if output.numel() > 10000 and sample:
                output = tensor_random_fetch(
                    tensor=output, seed=10086, # 保证随机种子一致才能比对结果
                    num_of_fetches=4096)

            dump_to_file(
                filename=os.path.join(o_dir, str(batch_idx), name + '.dat'),
                data=output, format='.dat')

        dump_to_file(
            filename=os.path.join(i_dir, str(batch_idx) + '.npy'),
            data=batch, format='.npy')


def split_result_to_directory(raw_dir: str, to_dir: str):