gobbli.model.transformer.model module¶
-
class
gobbli.model.transformer.model.
Transformer
(data_dir=None, load_existing=False, use_gpu=False, nvidia_visible_devices='all', logger=None, **kwargs)[source]¶ Bases:
gobbli.model.base.BaseModel
,gobbli.model.mixin.TrainMixin
,gobbli.model.mixin.PredictMixin
,gobbli.model.mixin.EmbedMixin
Classifier/embedding wrapper for any of the Transformers from transformers.
Create a model.
- Parameters
data_dir¶ (
Optional
[Path
]) – Optional path to a directory used to store model data. If not given, a unique directory under GOBBLI_DIR will be created and used.load_existing¶ (
bool
) – If True,data_dir
should be a directory that was previously used to create a model. Parameters will be loaded to match the original model, and user-specified model parameters will be ignored. If False, the data_dir must be empty if it already exists.use_gpu¶ (
bool
) – If True, use the nvidia-docker runtime (https://github.com/NVIDIA/nvidia-docker) to expose NVIDIA GPU(s) to the container. Will cause an error if the computer you’re running on doesn’t have an NVIDIA GPU and/or doesn’t have the nvidia-docker runtime installed.nvidia_visible_devices¶ (
str
) – Which GPUs to make available to the container; ignored ifuse_gpu
is False. If not ‘all’, should be a comma-separated string: ex.1,2
.logger¶ (
Optional
[Logger
]) – If passed, use this logger for logging instead of the default module-level logger.**kwargs¶ – Additional model-specific parameters to be passed to the model’s
init()
method.
-
build
()¶ Perform any pre-setup that needs to be done before running the model (building Docker images, etc).
-
property
class_weights_dir
¶ The root directory used to store initial model weights (before fine-tuning). These should generally be some pretrained weights made available by model developers. This directory will NOT be created by default; models should download their weights and remove the weights directory if the download doesn’t finish properly.
Most models making use of this directory will have multiple sets of weights and will need to store those in subdirectories under this directory.
- Return type
Path
- Returns
The path to the class-wide weights directory.
-
data_dir
()¶ - Return type
Path
- Returns
The main data directory unique to this instance of the model.
-
embed
(embed_input, embed_dir_name=None)¶ Generates embeddings using a model and the params in the given
gobbli.io.EmbedInput
.- Parameters
embed_input¶ (
EmbedInput
) – Contains various parameters needed to determine how to generate embeddings and what data to generate embeddings for.embed_dir_name¶ (
Optional
[str
]) – Optional name to store embedding input and output under. The directory is always created under the model’sdata_dir
. If a name is not given, a unique name is generated via a UUID. If a name is given, that directory must not already exist.
- Return type
- Returns
Output of training.
-
embed_dir
()¶ The directory to be used for data related to embedding (weights, embeddings, etc)
- Return type
Path
- Returns
Path to the embedding data directory.
-
property
host_cache_dir
¶ Directory to be used for downloaded transformers files. Should be the same across all instances of the class, since these are generally static model weights/config files that can be reused.
-
property
image_tag
¶ - Return type
str
- Returns
The Docker image tag to be used for the transformer container.
-
property
info_path
¶ - Return type
Path
- Returns
The path to the model’s info file, containing information about the model including the type of model, gobbli version it was trained using, etc.
-
init
(params)[source]¶ See
gobbli.model.base.BaseModel.init()
.Transformer parameters:
transformer_model
(str
): Name of a transformer model architecture to use. For training/prediction, the value should be one such thatfrom transformers import <value>ForSequenceClassification
is a valid import. ex value = “Bert” ->from transformers import BertForSequenceClassification
. Note this means only a subset of the transformers models are supported for these tasks – search the docs to see which ones you can use. For embedding generation, the import is<value>Model
, so any transformer model is supported.transformer_weights
(str
): Name of the pretrained weights to use. See the transformers docs for supported values. These depend on thetransformer_model
chosen.config_overrides
(dict
): Dictionary of keys and values that will override config for the model.max_seq_length
: Truncate all sequences to this length after tokenization. Used to save memory.lr
: Learning rate for the AdamW optimizer.adam_eps
: Epsilon value for the AdamW optimizer.gradient_accumulation_steps
: Number of iterations to accumulate gradients before updating the model. Used to allow larger effective batch sizes for models too big to fit a large batch on the GPU. The “effective batch size” isgradient_accumulation_steps
*TrainInput.params.train_batch_size
. If you encounter memory errors while training, try decreasing the batch size and increasinggradient_accumulation_steps
. For example, if a training batch size of 32 causes memory errors, try decreasing batch size to 16 and increasinggradient_accumulation_steps
to 2. If you still have problems with memory, you can drop batch size to 8 andgradient_accumulation_steps
to 4, and so on.
Note that gobbli relies on transformers to perform validation on these parameters, so initialization errors may not be caught until model runtime.
-
property
logger
¶ - Return type
Logger
- Returns
A logger for derived models to use.
-
property
metadata_path
¶ - Return type
Path
- Returns
The path to the model’s metadata file containing model-specific parameters.
-
classmethod
model_class_dir
()¶ - Return type
Path
- Returns
A directory shared among all classes of the model.
-
predict
(predict_input, predict_dir_name=None)¶ Runs prediction on new data using params containing in the given
gobbli.io.PredictInput
.- Parameters
predict_input¶ (
PredictInput
) – Contains various parameters needed to determine how to run prediction and what data to predict for.predict_dir_name¶ (
Optional
[str
]) – Optional name to store prediction input and output under. The directory is always created under the model’sdata_dir
. If a name is not given, a unique name is generated via a UUID. If a name is given, that directory must not already exist.
- Return type
-
predict_dir
()¶ The directory to be used for data related to prediction (weights, predictions, etc)
- Return type
Path
- Returns
Path to the prediction data directory.
-
train
(train_input, train_dir_name=None)¶ Trains a model using params in the given
gobbli.io.TrainInput
. The training process varies depending on the model, but in general, it includes the following steps:Update weights using the training dataset
Evaluate performance on the validation dataset.
Repeat for a number of epochs.
When finished, report loss/accuracy and return the trained weights.
- Parameters
train_input¶ (
TrainInput
) – Contains various parameters needed to determine how to train and what data to train on.train_dir_name¶ (
Optional
[str
]) – Optional name to store training input and output under. The directory is always created under the model’sdata_dir
. If a name is not given, a unique name is generated via a UUID. If a name is given, that directory must not already exist.
- Return type
- Returns
Output of training.
-
train_dir
()¶ The directory to be used for data related to training (data files, etc).
- Return type
Path
- Returns
Path to the training data directory.
-
property
weights_dir
¶ The directory containing weights for a specific instance of the model. This is the class weights directory by default, but subclasses might define this property to return a subdirectory based on a set of pretrained model weights.
- Return type
Path
- Returns
The instance-specific weights directory.