Callbacks are a way of adding additional methods to the finetuning process. The methods are executed when certain events occur and there are several callback classes, each serving a different function by providing different methods for different events.
A run can be assigned multiple callbacks using the optional callbacks parameter when it is created.
The EvaluationCallback is used to calculate performance metrics for the model being tuned at the end of each epoch.
In order to evaluate the model, two additional data sets - a query dataset and an index dataset - need to be provided as arguments.
If no index set is provided, the query dataset is reused instead. Below is an example of the metrics as they are printed at the end of finetuning:
The evaluation callback is triggered at the end of each epoch, in which the model is evaluated using the query_data and index_data datasets that were provided when the callback was created.
These datasets can be provided in the same way the train_data and eval_data parameters of the fit() method; either as a path to a CSV file, a DocumentArray or the name of a DocumentArray that has been pushed on the Jina AI Cloud. See Prepare Training Data for more information about how to prepare your data.
It is worth noting that the evaluation callback and the eval_data parameter of the fit method do not do the same thing.
The eval_data parameter is used to evaluate the loss of the model.
On the other hand, the evaluation callback is used to evaluate the quality of the searches using metrics such as average precision and recall.
These search metrics can be used by other callbacks if the evaluation callback is first in the list of callbacks when creating a run.
Evaluation callback with two models
Usually, you don’t need to provide the name of a model to the evalution callback.
The callback just takes the model which is fine-tuned.
However, if multiple models are involved in the fine-tuning process, like this is the case for CLIP models, it needs to be clear which model is used to encode the documents in query_data and index_data.
This can be specified by the model attribute of the callback.
If a different model should be used for the index_data, you can set this via the index_model attribute.
For an example, see Text-to-Image Search via CLIP.
During the fine-tuning process, the evaluation metrics are displayed in the logs, which you can retrieve via the logs() function.
After the fine-tuning job has finished, the evaluation metrics can be retrieved from the cloud by calling the metrics() function.
This function returns a JSON object with the metrics before and after fine-tuning.
Alternatively, you can also retrieve the metrics via the display_metrics() function, which prints the evaluation results in the form of a table to the console.
If you want to compare the K most similar retrieval results (Top-K) before and after fine-tuning, you can set the gather_examples parameter of the evaluation callback to True.
In this case, the evaluation callback will store the top-k results for each query document before and after fine-tuning.
You can retrieve them with the example_results() function.
Alternatively, you can use the display_examples() function to display a table of the Top-K results before and after fine-tuning to the console.
This callback evaluates the performance of the model at the end of each epoch, and keeps a record of the best perfoming model across all epochs. Once fitting is finished the best performing model is saved instead of the most recent model. The definition of best is based on two parameters:
monitor: The metric that is used to compare models to each other. By default this value is val_loss, the loss function calculated using the evaluation data, however the loss calculated on the training data can be used instead with train_loss; any metric that is recorded by the evaluation callback can also be used.
mode: Whether the monitored metric should be maximised (max) or minimised (min). By default the mode is set to auto, meaning that it will automatically choose the correct mode depending on the chosen metric: ‘min’ if the metric is loss and ‘max’ if the metric is one recorded by the evaluation callback.
The console output below shows how the evaluation loss of the model is monitored between each epoch, and how the best performing model is tracked. Since the final model has a higher loss than the previously recorded best model, the best model will be saved instead of the latest one.
Similarly to the best model checkpoint callback, the early stopping callback measures a given metric at the end of every epoch. Unlike the best model checkpoint callback, the early stopping callback does not save the best model; only the monitored metric is recorded between runs in order to assess the rate of improvement.
Below is some example output for a run with the early stopping callback followed by the output for the same run without the early stopping callback, and then the python code used to create the run. The output for the run with early stopping finished after just ten epochs whereas the other run finished all twenty epochs, resulting in nearly twice the runtime. That said, the resulting loss value of the early stopping run is only 0.284, compared to the full run’s 0.272, less than five percent higher. The early stopping callback can be used in this way to reduce the amount of training time while still showing improvement.
The early stopping callback triggers at the end of training and evaluation batches to record the loss, and at the end of each epoch to evaluate the model and compare it to the best so far. Whether it stops training at the end of an epoch depends on several parameters:
minimum_delta: The minimum amount of improvement that a model can have over the previous best model to be considered worthwhile, zero by default, meaning that the training will not stop early unless the performance starts to decrease
patience: The number of consecutive rounds without improvement before the training is stopped, two by default.
baseline: an optional parameter that is used to compare the model’s score against instead of the best previous model when checking for improvement. This baseline does not get changed over the course of a run.
Finetuner allows you to utilize Weights & Biases for experiment tracking and visualization.
The WandBLogger uses Weights & Biases Anonymous Mode
to track a Finetuner Run. The benefits of anonymous mode are: you do not need to share your
Weights & Biases api_key with us since no login is required.
Use WandBLogger together with EvaluationCallback
The WandBLogger will track the training loss, plus the evaluation loss if eval_data is not None.
If you use EvaluationCallback together with WandBLogger, search metrics will be tracked as well.
Such as mrr, precision, recall, etc.
fromfinetuner.callbackimportWandBLogger,EvaluationCallbackrun=finetuner.fit(model='resnet50',run_name='resnet-tll-early-6',train_data='finetuner/tll-train-da',epochs=5,learning_rate=1e-6,callbacks=[EvaluationCallback(query_data='finetuner/tll-test-query-da',index_data='finetuner/tll-test-index-da'),WandBLogger(),])# stream the logs, or use run.logs() to get logsforentryinrun.stream_logs():print(entry)
You can find the Weights & Biases entry in the logs, copy the link of View run at: