In this tutorial, we will learn about the different details around finetuning from pretrained models like loading from checkpoints, loading a model from the model zoo and doing validation/inference using finetuned models. We will walk-through the tutorial by training a Visual BERT model and train/validate on Hateful Memes dataset.
Follow the prerequisites for installation and dataset here.
VisualBERT model is pretrained on V+L multimodal data. We will use a pretrained model on COCO Captions. To begin finetuning our VisualBERT model we will load a model pretrained on COCO Captions and finetune that on Hateful Memes.
The config file contains two important changes :
checkpoint.resume_pretrained specifies if we want to resume from a pretrained model using the pretrained state dict mappings defined in
checkpoint.resume_zoo specifies which pretrained model from our model zoo we want to use for this. In this case, we will use
checkpoint.pretrained_state_mapping specifies how a pretrained model will be loaded and mapped to which keys of the target model. We use it since we only want to load specific layers from the pretrained model. In the case of VisualBERT model, we want to load the pretrained
bert layers. This is specified in our defaults.yaml:
This will ensure only the
model.bert layers of the COCO pretrained model gets loaded.
We can also use the default config for VisualBERT on hateful memes directly and override the pretrained options through command line args:
After running the training our model will be saved in
env.save_dir if overriden. This will be the directory structure:
Instead of loading from the model zoo we can also load from a file:
checkpoint.resume_file can also be used when loading a model file for evaluation or generating predictions. We will see more example usage of this later.
To resume the training in case it gets intterupted, run:
checkpoint.resume=True, MMF will try to load automatically the last saved checkpoint in the
env.save_dir experiment folder
Instead of the last saved checkpoint, we can also resume from the "best" checkpoint based on
training.early_stop.criteria if enabled in the config. This can be achieved using
In the config early stopping parameters are as follows:
Optionally, you can limit the maximum number of checkpoint files that are saved.
In the config, this is managed with the parameter
When the parameter is set to -1, every eligible checkpoint is saved; otherwise, only the last
max_to_keep checkpoints are kept at any given time.
After we finish the training we will load the trained model for validation:
Note that here we specify
run_type=val so that we are running only validation. We also use
checkpoint.resume_file to load the trained model.
We will next load the trained model to generate prediction results:
This will generate a submission file in csv format that can be used for submission to the Hateful Memes challenge.