Skip to content

_load_model

load_model

Load saved weights for a given model.

This method can be used with TensorFlow models: 

m = fe.build(fe.architecture.tensorflow.LeNet, optimizer_fn="adam")
fe.backend.save_model(m, save_dir="tmp", model_name="test")
fe.backend.load_model(m, weights_path="tmp/test.h5")

This method can be used with PyTorch models: 

m = fe.build(fe.architecture.pytorch.LeNet, optimizer_fn="adam")
fe.backend.save_model(m, save_dir="tmp", model_name="test")
fe.backend.load_model(m, weights_path="tmp/test.pt")

Parameters:

Name Type Description Default
model Union[Model, Module]

A neural network instance to load.

 required
weights_path str

Path to the model weights.

 required
load_optimizer bool

Whether to load optimizer. If True, then it will load file in the path.

 False

Raises:

Type Description
ValueError

If model is an unacceptable data type.
Source code in fastestimator/fastestimator/backend/_load_model.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
def load_model(model: Union[tf.keras.Model, torch.nn.Module], weights_path: str, load_optimizer: bool = False):
    """Load saved weights for a given model.

    This method can be used with TensorFlow models:
    ```python
    m = fe.build(fe.architecture.tensorflow.LeNet, optimizer_fn="adam")
    fe.backend.save_model(m, save_dir="tmp", model_name="test")
    fe.backend.load_model(m, weights_path="tmp/test.h5")
    ```

    This method can be used with PyTorch models:
    ```python
    m = fe.build(fe.architecture.pytorch.LeNet, optimizer_fn="adam")
    fe.backend.save_model(m, save_dir="tmp", model_name="test")
    fe.backend.load_model(m, weights_path="tmp/test.pt")
    ```

    Args:
        model: A neural network instance to load.
        weights_path: Path to the `model` weights.
        load_optimizer: Whether to load optimizer. If True, then it will load <weights_opt> file in the path.

    Raises:
        ValueError: If `model` is an unacceptable data type.
    """
    if load_optimizer:
        assert hasattr(model, "fe_compiled") and model.fe_compiled, "model must be built by fe.build if loading optimizers"

    if not os.path.exists(weights_path):
        raise ValueError("Weights path doesn't exist: ", weights_path)

    if isinstance(model, tf.keras.Model):
        model.load_weights(weights_path)
        if load_optimizer:
            assert model.current_optimizer, "optimizer does not exist"
            optimizer_path = "{}_opt.pkl".format(os.path.splitext(weights_path)[0])
            assert os.path.exists(optimizer_path), "cannot find optimizer path: {}".format(optimizer_path)
            with open(optimizer_path, 'rb') as f:
                state_dict = pickle.load(f)
            model.current_optimizer.set_weights(state_dict['weights'])
            weight_decay = None
            if hasattr(model.current_optimizer, "weight_decay") and tf.keras.backend.get_value(
                    model.current_optimizer.weight_decay) is not None:
                weight_decay = state_dict['weight_decay']
            set_lr(model, state_dict['lr'], weight_decay=weight_decay)
    elif isinstance(model, torch.nn.Module):
        if isinstance(model, torch.nn.DataParallel):
            model.module.load_state_dict(preprocess_torch_weights(weights_path))
        else:
            model.load_state_dict(preprocess_torch_weights(weights_path))
        if load_optimizer:
            assert model.current_optimizer, "optimizer does not exist"
            optimizer_path = "{}_opt.pt".format(os.path.splitext(weights_path)[0])
            assert os.path.exists(optimizer_path), "cannot find optimizer path: {}".format(optimizer_path)
            model.current_optimizer.load_state_dict(torch.load(optimizer_path))
    else:
        raise ValueError("Unrecognized model instance {}".format(type(model)))

preprocess_torch_weights

Preprocess the torch weights dictionary.

This method is used to remove the any DataParallel artifacts in torch weigths.

Parameters:

Name Type Description Default
weights_path str

Path to the model weights.

 required
Source code in fastestimator/fastestimator/backend/_load_model.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
def preprocess_torch_weights(weights_path: str) -> OrderedDict:
    """Preprocess the torch weights dictionary.

    This method is used to remove the any DataParallel artifacts in torch weigths.

    Args:
        weights_path: Path to the model weights.
    """
    new_state_dict = OrderedDict()
    for key, value in torch.load(weights_path, map_location='cpu' if torch.cuda.device_count() == 0 else None).items():
        # remove `module.`
        new_key = key
        if key.startswith('module.'):
            new_key = key[7:]
        new_state_dict[new_key] = value

    return new_state_dict