Error Training COCO: RuntimeError: result type Float can't be cast to the desired output type long long

[kells1986]

I'm just getting to grips with the library and the codebase, all looks great so far.

I'm trying to train on COCO just to verify my understanding. However I'm running into the error: RuntimeError: result type Float can't be cast to the desired output type long long

The code itself is fairly minimal:

import yolort
import torch
import torchvision
import pytorch_lightning as pl
import yolort.trainer
import os
from PIL import Image
import yolort.data.transforms as T
import yolort.data.data_module

lightning_task = yolort.trainer.lightning_task.DefaultTask(num_classes=80)

def get_transform(train):
    transforms = []
    transforms.append(T.PILToTensor())
    transforms.append(T.ConvertImageDtype(torch.float))
    if train:
        transforms.append(T.RandomHorizontalFlip(0.5))
    return T.Compose(transforms)

dataset = yolort.data.coco.COCODetection("<path to>/Downloads/val2017/",
                                         "<path to>/Downloads/annotations/instances_val2017.json",
                                         get_transform(train=False))

train_dataset, val_dataset = torch.utils.data.random_split(dataset, [4500, 500])

detection_data_module = yolort.data.data_module.DetectionDataModule(train_dataset=train_dataset,
                                                                    val_dataset=val_dataset)

trainer = pl.Trainer(max_epochs=2,
                    log_every_n_steps=10
                    )

trainer.fit(lightning_task,
            train_dataloaders=detection_data_module.train_dataloader(),
            val_dataloaders=detection_data_module.val_dataloader())

The trainer.fit call gets past the sanity checking stage, but fails before the first step of the first epoch.

Am I missing something?

Is there a minimal training example available somewhere?

[kells1986]

This appears to have been fixed by installing yolort from GitHub rather than running pip install -U yolort.

pip install -U yolort installed yolort == 0.6.3 on my machine.

Cloning the repo and then running pip install -e . appears to have resolved the error I was seeing