it would seem they are related but i cant see the further details of this bug. Either doing a manual artefact upload with task or turning tensor board tracking off in the hugging face trainer both seemed to enable json tracking within the checkpoints. But I would have thought the tensorboard behavior wasnt desired.

called with:

task = Task.init(
    project_name=project_name, task_name=task_name, output_uri="fileserver_address"
)


task.connect(config)


checkpoint = config.get("model_path")

image_processor = AutoImageProcessor.from_pretrained(
        checkpoint,
        num_labels=config.get("class_number"),
    )

best_model = training(checkpoint, image_processor)

you can see here its missing some files

need to work out if i need to reconfigure something and re-train or are my files (actual model tensor) recoverable

default_outputdir in the conf is set to the filerserver address same as above pointing

training function:

def training(checkpoint, image_processor):

    data_test_train, labels, label_to_id, id_to_label = pre_process()

    model = AutoModelForImageClassification.from_pretrained(
        checkpoint,
        num_labels=len(labels),
        id2label=id_to_label,
        label2id=label_to_id,
        ),
    )
    def metrics(eval_pred):
        metric_val = config.get("eval_metric")
        metric = evaluate.load(metric_val)
        predictions, labels = eval_pred
        predictions = np.argmax(predictions, axis=1)
        if metric_val == "accuracy":
            return metric.compute(predictions=predictions, references=labels)
        else:
            return metric.compute(
                predictions=predictions, references=labels, average="weighted"
            )

    data_collator = DefaultDataCollator()

    training_args = TrainingArguments(
        output_dir="somefolder",
        remove_unused_columns=False,
        eval_strategy="epoch",
        save_strategy="epoch",
        learning_rate=config.get("learning_rate"),
        per_device_train_batch_size=config.get("train_batch_size"),
        gradient_accumulation_steps=config.get("gradient_steps"),
        per_device_eval_batch_size=config.get("eval_batch_size"),
        num_train_epochs=config.get("epochs"),
        warmup_ratio=config.get("warmup_ratio"),
        logging_steps=config.get("logging_steps"),
        save_total_limit=config.get("save_total"),
        load_best_model_at_end=True,
        report_to="tensorboard",
        metric_for_best_model=config.get("eval_metric"),
        push_to_hub=False,
    )
    trainer = Trainer(
        model=model.to(device),
        args=training_args,
        data_collator=data_collator,
        train_dataset=data_test_train["train"],
        eval_dataset=data_test_train["test"],
        tokenizer=image_processor,
        compute_metrics=metrics,
    )
    print("training")
    trainer.train()
    print("save_model")
    trainer.save_model()
    best_model = trainer.state.best_model_checkpoint
    print(best_model)
    classifification_report(data_test_train["test"], model, best_model)
    return best_model

Try save_safetensors=False in TrainingArguments . Not sure if clearML supports safetensors

further context. it saves the optimizer.pt , scheduler.pt , rng_state_path, and training_args.bin. but i cant locate the model.safetensors or meta jsons

Hi @<1730033904972206080:profile|FantasticSeaurchin8>
Is this only relates to this
https://github.com/coqui-ai/Trainer/issues/7
Or is it a clearml sdk issue?

Hi @<1730033904972206080:profile|FantasticSeaurchin8> , can you provide a sample script that reproduces this behaviour?

So we have managed to get whole checkpoint files to save by removing the save_total_limit from training, this seems to save checkpoint folders with all files in it. however now we have a ballooning server.

did discover this None
and wondering if there's some nuance in autotracking that needs to be circumvented

But that doesn't explain why the model JSON files are missing.
@<1523701070390366208:profile|CostlyOstrich36> do you have any ideas? Thank you

None ?

console output:

clearml.Task - INFO - Completed model upload to file_server/training.e5f99149b9b/models/optimizer.pt
clearml.Task - INFO - Completed model upload to file_server/training.e5f99149b9b/models/scheduler.pt
clearml.Task - INFO - Completed model upload to file_server/training.e5f99149b9b/models/rng_state.pth
save_model
somemodel/checkpoint-198
clearml.Task - INFO - Completed model upload to file_server/training.e5f99149b9b/models/training_args.bin

hmm i can probably provide snippets

queued with:

task = Task.create(
    project_name="name",
    task_name="training",
    repo="repo",
    branch="branch",
    script="training_script",
    packages=package_list,
    docker="docker_gpu_image",
    docker_args=["--network=host"],
)
task.output_uri = "filer_server"
task.enqueue(task, "training")

so turning report_to="tensorboard", off seemed to solve the issue...as in the training run saves checkpoints as you would expect. That doesnt seem like desired behavior..