mlflow.pytorch

The mlflow.pytorch module provides an API for logging and loading PyTorch models. This module exports PyTorch models with the following flavors:

PyTorch (native) format

This is the main flavor that can be loaded back into PyTorch.

mlflow.pyfunc

Produced for use by generic pyfunc-based deployment tools and batch inference.

mlflow.pytorch.autolog(log_every_n_epoch=1, log_every_n_step=None, log_models=True, disable=False, exclusive=False, disable_for_unsupported_versions=False, silent=False, registered_model_name=None)[source]

Note

Autologging is known to be compatible with the following package versions: 1.0.5 <= pytorch-lightning <= 1.7.7. Autologging may not succeed when used with package versions outside of this range.

Enables (or disables) and configures autologging from PyTorch Lightning to MLflow.

Autologging is performed when you call the fit method of pytorch_lightning.Trainer().

Explore the complete PyTorch MNIST for an expansive example with implementation of additional lightening steps.

Note: Autologging is only supported for PyTorch Lightning models, i.e., models that subclass pytorch_lightning.LightningModule. In particular, autologging support for vanilla PyTorch models that only subclass torch.nn.Module is not yet available.

Parameters

  • log_every_n_epoch – If specified, logs metrics once every n epochs. By default, metrics are logged after every epoch.

  • log_every_n_step – If specified, logs batch metrics once every n global step. By default, metrics are not logged for steps. Note that setting this to 1 can cause performance issues and is not recommended.

  • log_models – If True, trained models are logged as MLflow model artifacts. If False, trained models are not logged.

  • disable – If True, disables the PyTorch Lightning autologging integration. If False, enables the PyTorch Lightning autologging integration.

  • exclusive – If True, autologged content is not logged to user-created fluent runs. If False, autologged content is logged to the active fluent run, which may be user-created.

  • disable_for_unsupported_versions – If True, disable autologging for versions of pytorch and pytorch-lightning that have not been tested against this version of the MLflow client or are incompatible.

  • silent – If True, suppress all event logs and warnings from MLflow during PyTorch Lightning autologging. If False, show all events and warnings during PyTorch Lightning autologging.

  • registered_model_name – If given, each time a model is trained, it is registered as a new model version of the registered model with this name. The registered model is created if it does not already exist.

Example 
import os

import pytorch_lightning as pl
import torch
from torch.nn import functional as F
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision.datasets import MNIST

try:
    from torchmetrics.functional import accuracy
except ImportError:
    from pytorch_lightning.metrics.functional import accuracy

import mlflow.pytorch
from mlflow import MlflowClient

# For brevity, here is the simplest most minimal example with just a training
# loop step, (no validation, no testing). It illustrates how you can use MLflow
# to auto log parameters, metrics, and models.

class MNISTModel(pl.LightningModule):
    def __init__(self):
        super().__init__()
        self.l1 = torch.nn.Linear(28 * 28, 10)

    def forward(self, x):
        return torch.relu(self.l1(x.view(x.size(0), -1)))

    def training_step(self, batch, batch_nb):
        x, y = batch
        logits = self(x)
        loss = F.cross_entropy(logits, y)
        pred = logits.argmax(dim=1)
        acc = accuracy(pred, y)

        # Use the current of PyTorch logger
        self.log("train_loss", loss, on_epoch=True)
        self.log("acc", acc, on_epoch=True)
        return loss

    def configure_optimizers(self):
        return torch.optim.Adam(self.parameters(), lr=0.02)

def print_auto_logged_info(r):

    tags = {k: v for k, v in r.data.tags.items() if not k.startswith("mlflow.")}
    artifacts = [f.path for f in MlflowClient().list_artifacts(r.info.run_id, "model")]
    print("run_id: {}".format(r.info.run_id))
    print("artifacts: {}".format(artifacts))
    print("params: {}".format(r.data.params))
    print("metrics: {}".format(r.data.metrics))
    print("tags: {}".format(tags))

# Initialize our model
mnist_model = MNISTModel()

# Initialize DataLoader from MNIST Dataset
train_ds = MNIST(os.getcwd(), train=True,
    download=True, transform=transforms.ToTensor())
train_loader = DataLoader(train_ds, batch_size=32)

# Initialize a trainer
trainer = pl.Trainer(max_epochs=20, progress_bar_refresh_rate=20)

# Auto log all MLflow entities
mlflow.pytorch.autolog()

# Train the model
with mlflow.start_run() as run:
    trainer.fit(mnist_model, train_loader)

# fetch the auto logged parameters and metrics
print_auto_logged_info(mlflow.get_run(run_id=run.info.run_id))
Output 
run_id: 42caa17b60cb489c8083900fb52506a7
artifacts: ['model/MLmodel', 'model/conda.yaml', 'model/data']
params: {'betas': '(0.9, 0.999)',
         'weight_decay': '0',
         'epochs': '20',
         'eps': '1e-08',
         'lr': '0.02',
         'optimizer_name': 'Adam', '
         amsgrad': 'False'}
metrics: {'acc_step': 0.0,
          'train_loss_epoch': 1.0917967557907104,
          'train_loss_step': 1.0794280767440796,
          'train_loss': 1.0794280767440796,
          'acc_epoch': 0.0033333334140479565,
          'acc': 0.0}
tags: {'Mode': 'training'}
../_images/pytorch_lightening_autolog.png

PyTorch autologged MLflow entities

mlflow.pytorch.get_default_conda_env()[source]
Returns

The default Conda environment as a dictionary for MLflow Models produced by calls to save_model() and log_model().

Example 
import mlflow.pytorch

# Log PyTorch model
with mlflow.start_run() as run:
    mlflow.pytorch.log_model(model, "model")

# Fetch the associated conda environment
env = mlflow.pytorch.get_default_conda_env()
print("conda env: {}".format(env))
Output 
conda env {'name': 'mlflow-env',
           'channels': ['conda-forge'],
           'dependencies': ['python=3.7.5',
                            {'pip': ['torch==1.5.1',
                                     'mlflow',
                                     'cloudpickle==1.6.0']}]}
mlflow.pytorch.get_default_pip_requirements()[source]
Returns

A list of default pip requirements for MLflow Models produced by this flavor. Calls to save_model() and log_model() produce a pip environment that, at minimum, contains these requirements.

mlflow.pytorch.load_model(model_uri, dst_path=None, **kwargs)[source]

Load a PyTorch model from a local file or a run.

Parameters

  • model_uri –

    The location, in URI format, of the MLflow model, for example:

    • /Users/me/path/to/local/model

    • relative/path/to/local/model

    • s3://my_bucket/path/to/model

    • runs:/<mlflow_run_id>/run-relative/path/to/model

    • models:/<model_name>/<model_version>

    • models:/<model_name>/<stage>

    For more information about supported URI schemes, see Referencing Artifacts.

  • dst_path – The local filesystem path to which to download the model artifact. This directory must already exist. If unspecified, a local output path will be created.

  • kwargs – kwargs to pass to torch.load method.

Returns

A PyTorch model.

Example 
import torch
import mlflow.pytorch

# Class defined here
class LinearNNModel(torch.nn.Module):
    ...

# Initialize our model, criterion and optimizer
...

# Training loop
...

# Log the model
with mlflow.start_run() as run:
    mlflow.pytorch.log_model(model, "model")

# Inference after loading the logged model
model_uri = "runs:/{}/model".format(run.info.run_id)
loaded_model = mlflow.pytorch.load_model(model_uri)
for x in [4.0, 6.0, 30.0]:
    X = torch.Tensor([[x]])
    y_pred = loaded_model(X)
    print("predict X: {}, y_pred: {:.2f}".format(x, y_pred.data.item()))
Output 
predict X: 4.0, y_pred: 7.57
predict X: 6.0, y_pred: 11.64
predict X: 30.0, y_pred: 60.48
mlflow.pytorch.load_state_dict(state_dict_uri, **kwargs)[source]

Note

Experimental: This method may change or be removed in a future release without warning.

Load a state_dict from a local file or a run.

Parameters

  • state_dict_uri –

    The location, in URI format, of the state_dict, for example:

    • /Users/me/path/to/local/state_dict

    • relative/path/to/local/state_dict

    • s3://my_bucket/path/to/state_dict

    • runs:/<mlflow_run_id>/run-relative/path/to/state_dict

    For more information about supported URI schemes, see Referencing Artifacts.

  • kwargs – kwargs to pass to torch.load.

Returns

A state_dict

Example 
with mlflow.start_run():
    artifact_path = "model"
    mlflow.pytorch.log_state_dict(model.state_dict(), artifact_path)
    state_dict_uri = mlflow.get_artifact_uri(artifact_path)

state_dict = mlflow.pytorch.load_state_dict(state_dict_uri)
mlflow.pytorch.log_model(pytorch_model, artifact_path, conda_env=None, code_paths=None, pickle_module=None, registered_model_name=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list, csr_matrix, csc_matrix] = None, await_registration_for=300, requirements_file=None, extra_files=None, pip_requirements=None, extra_pip_requirements=None, **kwargs)[source]

Log a PyTorch model as an MLflow artifact for the current run.

Warning

Log the model with a signature to avoid inference errors. If the model is logged without a signature, the MLflow Model Server relies on the default inferred data type from NumPy. However, PyTorch often expects different defaults, particularly when parsing floats. You must include the signature to ensure that the model is logged with the correct data type so that the MLflow model server can correctly provide valid input.

Parameters

  • pytorch_model –

    PyTorch model to be saved. Can be either an eager model (subclass of torch.nn.Module) or scripted model prepared via torch.jit.script or torch.jit.trace.

    The model accept a single torch.FloatTensor as input and produce a single output tensor.

    If saving an eager model, any code dependencies of the model’s class, including the class definition itself, should be included in one of the following locations:

    • The package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_paths parameter.

  • artifact_path – Run-relative artifact path.

  • conda_env –

    Either a dictionary representation of a Conda environment or the path to a conda environment yaml file. If provided, this describes the environment this model should be run in. At minimum, it should specify the dependencies contained in get_default_conda_env(). If None, a conda environment with pip requirements inferred by mlflow.models.infer_pip_requirements() is added to the model. If the requirement inference fails, it falls back to using get_default_pip_requirements(). pip requirements from conda_env are written to a pip requirements.txt file and the full conda environment is written to conda.yaml. The following is an example dictionary representation of a conda environment:

    {
    "name": "mlflow-env",
    "channels": ["conda-forge"],
    "dependencies": [
        "python=3.7.0",
        {
            "pip": [
                "torch==x.y.z"
            ],
        },
    ],
}

  • code_paths – A list of local filesystem paths to Python file dependencies (or directories containing file dependencies). These files are prepended to the system path when the model is loaded.

  • pickle_module – The module that PyTorch should use to serialize (“pickle”) the specified pytorch_model. This is passed as the pickle_module parameter to torch.save(). By default, this module is also used to deserialize (“unpickle”) the PyTorch model at load time.

  • registered_model_name – If given, create a model version under registered_model_name, also creating a registered model if one with the given name does not exist.

  • signature –

    ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models.signature import infer_signature
train = df.drop_column("target_label")
predictions = ... # compute model predictions
signature = infer_signature(train, predictions)

  • input_example – Input example provides one or several instances of valid model input. The example can be used as a hint of what data to feed the model. The given example can be a Pandas DataFrame where the given example will be serialized to json using the Pandas split-oriented format, or a numpy array where the example will be serialized to json by converting it to a list. Bytes are base64-encoded.

  • await_registration_for – Number of seconds to wait for the model version to finish being created and is in READY status. By default, the function waits for five minutes. Specify 0 or None to skip waiting.

  • requirements_file –

    Warning

    requirements_file has been deprecated. Please use pip_requirements instead.

    A string containing the path to requirements file. Remote URIs are resolved to absolute filesystem paths. For example, consider the following requirements_file string:

    requirements_file = "s3://my-bucket/path/to/my_file"

    In this case, the "my_file" requirements file is downloaded from S3. If None, no requirements file is added to the model.

  • extra_files –

    A list containing the paths to corresponding extra files. Remote URIs are resolved to absolute filesystem paths. For example, consider the following extra_files list -

    extra_files = [“s3://my-bucket/path/to/my_file1”,

    ”s3://my-bucket/path/to/my_file2”]

    In this case, the "my_file1 & my_file2" extra file is downloaded from S3.

    If None, no extra files are added to the model.

  • pip_requirements – Either an iterable of pip requirement strings (e.g. ["torch", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes the environment this model should be run in. If None, a default list of requirements is inferred by mlflow.models.infer_pip_requirements() from the current software environment. If the requirement inference fails, it falls back to using get_default_pip_requirements(). Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

  • extra_pip_requirements –

    Either an iterable of pip requirement strings (e.g. ["pandas", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes additional pip requirements that are appended to a default set of pip requirements generated automatically based on the user’s current software environment. Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

    Warning

    The following arguments can’t be specified at the same time:

    • conda_env

    • pip_requirements

    • extra_pip_requirements

    This example demonstrates how to specify pip requirements using pip_requirements and extra_pip_requirements.

  • kwargs – kwargs to pass to torch.save method.

Returns

A ModelInfo instance that contains the metadata of the logged model.

Example 
import numpy as np
import torch
import mlflow.pytorch

class LinearNNModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = torch.nn.Linear(1, 1)  # One in and one out

    def forward(self, x):
        y_pred = self.linear(x)
        return y_pred

def gen_data():
    # Example linear model modified to use y = 2x
    # from https://github.com/hunkim/PyTorchZeroToAll
    # X training data, y labels
    X = torch.arange(1.0, 25.0).view(-1, 1)
    y = torch.from_numpy(np.array([x * 2 for x in X])).view(-1, 1)
    return X, y

# Define model, loss, and optimizer
model = LinearNNModel()
criterion = torch.nn.MSELoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.001)

# Training loop
epochs = 250
X, y = gen_data()
for epoch in range(epochs):
    # Forward pass: Compute predicted y by passing X to the model
    y_pred = model(X)

    # Compute the loss
    loss = criterion(y_pred, y)

    # Zero gradients, perform a backward pass, and update the weights.
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

# Log the model
with mlflow.start_run() as run:
    mlflow.pytorch.log_model(model, "model")

    # convert to scripted model and log the model
    scripted_pytorch_model = torch.jit.script(model)
    mlflow.pytorch.log_model(scripted_pytorch_model, "scripted_model")

# Fetch the logged model artifacts
print("run_id: {}".format(run.info.run_id))
for artifact_path in ["model/data", "scripted_model/data"]:
    artifacts = [f.path for f in MlflowClient().list_artifacts(run.info.run_id,
                artifact_path)]
    print("artifacts: {}".format(artifacts))
Output 
run_id: 1a1ec9e413ce48e9abf9aec20efd6f71
artifacts: ['model/data/model.pth',
            'model/data/pickle_module_info.txt']
artifacts: ['scripted_model/data/model.pth',
            'scripted_model/data/pickle_module_info.txt']
../_images/pytorch_logged_models.png

PyTorch logged models

mlflow.pytorch.log_state_dict(state_dict, artifact_path, **kwargs)[source]

Note

Experimental: This method may change or be removed in a future release without warning.

Log a state_dict as an MLflow artifact for the current run.

Warning

This function just logs a state_dict as an artifact and doesn’t generate an MLflow Model.

Parameters

  • state_dict – state_dict to be saved.

  • artifact_path – Run-relative artifact path.

  • kwargs – kwargs to pass to torch.save.

Example 
# Log a model as a state_dict
with mlflow.start_run():
    state_dict = model.state_dict()
    mlflow.pytorch.log_state_dict(state_dict, artifact_path="model")

# Log a checkpoint as a state_dict
with mlflow.start_run():
    state_dict = {
        "model": model.state_dict(),
        "optimizer": optimizer.state_dict(),
        "epoch": epoch,
        "loss": loss,
    }
    mlflow.pytorch.log_state_dict(state_dict, artifact_path="checkpoint")
mlflow.pytorch.save_model(pytorch_model, path, conda_env=None, mlflow_model=None, code_paths=None, pickle_module=None, signature: mlflow.models.signature.ModelSignature = None, input_example: Union[pandas.core.frame.DataFrame, numpy.ndarray, dict, list, csr_matrix, csc_matrix] = None, requirements_file=None, extra_files=None, pip_requirements=None, extra_pip_requirements=None, **kwargs)[source]

Save a PyTorch model to a path on the local file system.

Parameters

  • pytorch_model –

    PyTorch model to be saved. Can be either an eager model (subclass of torch.nn.Module) or scripted model prepared via torch.jit.script or torch.jit.trace.

    The model accept a single torch.FloatTensor as input and produce a single output tensor.

    If saving an eager model, any code dependencies of the model’s class, including the class definition itself, should be included in one of the following locations:

    • The package(s) listed in the model’s Conda environment, specified by the conda_env parameter.

    • One or more of the files specified by the code_paths parameter.

  • path – Local path where the model is to be saved.

  • conda_env –

    Either a dictionary representation of a Conda environment or the path to a conda environment yaml file. If provided, this describes the environment this model should be run in. At minimum, it should specify the dependencies contained in get_default_conda_env(). If None, a conda environment with pip requirements inferred by mlflow.models.infer_pip_requirements() is added to the model. If the requirement inference fails, it falls back to using get_default_pip_requirements(). pip requirements from conda_env are written to a pip requirements.txt file and the full conda environment is written to conda.yaml. The following is an example dictionary representation of a conda environment:

    {
    "name": "mlflow-env",
    "channels": ["conda-forge"],
    "dependencies": [
        "python=3.7.0",
        {
            "pip": [
                "torch==x.y.z"
            ],
        },
    ],
}

  • mlflow_model – mlflow.models.Model this flavor is being added to.

  • code_paths – A list of local filesystem paths to Python file dependencies (or directories containing file dependencies). These files are prepended to the system path when the model is loaded.

  • pickle_module – The module that PyTorch should use to serialize (“pickle”) the specified pytorch_model. This is passed as the pickle_module parameter to torch.save(). By default, this module is also used to deserialize (“unpickle”) the PyTorch model at load time.

  • signature –

    ModelSignature describes model input and output Schema. The model signature can be inferred from datasets with valid model input (e.g. the training dataset with target column omitted) and valid model output (e.g. model predictions generated on the training dataset), for example:

    from mlflow.models.signature import infer_signature
train = df.drop_column("target_label")
predictions = ... # compute model predictions
signature = infer_signature(train, predictions)

  • input_example – Input example provides one or several instances of valid model input. The example can be used as a hint of what data to feed the model. The given example can be a Pandas DataFrame where the given example will be serialized to json using the Pandas split-oriented format, or a numpy array where the example will be serialized to json by converting it to a list. Bytes are base64-encoded.

  • requirements_file –

    Warning

    requirements_file has been deprecated. Please use pip_requirements instead.

    A string containing the path to requirements file. Remote URIs are resolved to absolute filesystem paths. For example, consider the following requirements_file string:

    requirements_file = "s3://my-bucket/path/to/my_file"

    In this case, the "my_file" requirements file is downloaded from S3. If None, no requirements file is added to the model.

  • extra_files –

    A list containing the paths to corresponding extra files. Remote URIs are resolved to absolute filesystem paths. For example, consider the following extra_files list -

    extra_files = [“s3://my-bucket/path/to/my_file1”,

    ”s3://my-bucket/path/to/my_file2”]

    In this case, the "my_file1 & my_file2" extra file is downloaded from S3.

    If None, no extra files are added to the model.

  • pip_requirements – Either an iterable of pip requirement strings (e.g. ["torch", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes the environment this model should be run in. If None, a default list of requirements is inferred by mlflow.models.infer_pip_requirements() from the current software environment. If the requirement inference fails, it falls back to using get_default_pip_requirements(). Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

  • extra_pip_requirements –

    Either an iterable of pip requirement strings (e.g. ["pandas", "-r requirements.txt", "-c constraints.txt"]) or the string path to a pip requirements file on the local filesystem (e.g. "requirements.txt"). If provided, this describes additional pip requirements that are appended to a default set of pip requirements generated automatically based on the user’s current software environment. Both requirements and constraints are automatically parsed and written to requirements.txt and constraints.txt files, respectively, and stored as part of the model. Requirements are also written to the pip section of the model’s conda environment (conda.yaml) file.

    Warning

    The following arguments can’t be specified at the same time:

    • conda_env

    • pip_requirements

    • extra_pip_requirements

    This example demonstrates how to specify pip requirements using pip_requirements and extra_pip_requirements.

  • kwargs – kwargs to pass to torch.save method.

Example 
import os

import torch
import mlflow.pytorch

# Class defined here
class LinearNNModel(torch.nn.Module):
    ...

# Initialize our model, criterion and optimizer
...

# Training loop
...

# Save PyTorch models to current working directory
with mlflow.start_run() as run:
    mlflow.pytorch.save_model(model, "model")

    # Convert to a scripted model and save it
    scripted_pytorch_model = torch.jit.script(model)
    mlflow.pytorch.save_model(scripted_pytorch_model, "scripted_model")

# Load each saved model for inference
for model_path in ["model", "scripted_model"]:
    model_uri = "{}/{}".format(os.getcwd(), model_path)
    loaded_model = mlflow.pytorch.load_model(model_uri)
    print("Loaded {}:".format(model_path))
    for x in [6.0, 8.0, 12.0, 30.0]:
        X = torch.Tensor([[x]])
        y_pred = loaded_model(X)
        print("predict X: {}, y_pred: {:.2f}".format(x, y_pred.data.item()))
    print("--")
Output 
Loaded model:
predict X: 6.0, y_pred: 11.90
predict X: 8.0, y_pred: 15.92
predict X: 12.0, y_pred: 23.96
predict X: 30.0, y_pred: 60.13
--
Loaded scripted_model:
predict X: 6.0, y_pred: 11.90
predict X: 8.0, y_pred: 15.92
predict X: 12.0, y_pred: 23.96
predict X: 30.0, y_pred: 60.13
mlflow.pytorch.save_state_dict(state_dict, path, **kwargs)[source]

Note

Experimental: This method may change or be removed in a future release without warning.

Save a state_dict to a path on the local file system

Parameters

  • state_dict – state_dict to be saved.

  • path – Local path where the state_dict is to be saved.

  • kwargs – kwargs to pass to torch.save.