def _accumulate_predictions_from_multiple_gpus(predictions_per_gpu):
all_predictions = all_gather(predictions_per_gpu)
if not is_main_process():
return
predictions = {}
for p in all_predictions:
predictions.update(p)
return predictions
def after_train_epoch(self, trainer):
if not self.every_n_epochs(trainer, self.interval):
return
trainer.model.eval()
results = [None for _ in range(len(self.dataset))]
detections = {}
if trainer.rank == 0:
prog_bar = torchie.ProgressBar(len(self.dataset))
for idx in range(trainer.rank, len(self.dataset), trainer.world_size):
data = self.dataset[idx]
data_gpu = scatter(collate_kitti([data], samples_per_gpu=1),
[torch.cuda.current_device()])[0]
# compute output
with torch.no_grad():
output = trainer.model(data_gpu, return_loss=False)
token = output["metadata"]["token"]
for k, v in output.items():
if k not in [
"metadata",
]:
output[k] = v.to(cpu_device)
detections.update({
token: output,
})
detections[idx] = result
batch_size = trainer.world_size
if trainer.rank == 0:
for _ in range(batch_size):
prog_bar.update()
all_predictions = all_gather(detections)
if trainer.rank != 0:
return
predictions = {}
for p in all_predictions:
predictions.update(p)
result_dict, _ = self.dataset.evaluation(predictions, None)
for k, v in result_dict["results"].items():
print(f"Evaluation {k}: {v}")
