Rename the parameters of a PyTorch module's saved state dict. Last tested with PyTorch 1.0.1.

rename_state_dict_keys.py

import torch
from collections import OrderedDict


def rename_state_dict_keys(source, key_transformation, target=None):
    """
    source             -> Path to the saved state dict.
    key_transformation -> Function that accepts the old key names of the state
                          dict as the only argument and returns the new key name.
    target (optional)  -> Path at which the new state dict should be saved
                          (defaults to `source`)

    Example:
    Rename the key `layer.0.weight` `layer.1.weight` and keep the names of all
    other keys.

    ```py
    def key_transformation(old_key):
        if old_key == "layer.0.weight":
            return "layer.1.weight"

        return old_key

    rename_state_dict_keys(state_dict_path, key_transformation)
    ```
    """
    if target is None:
        target = source

    state_dict = torch.load(source)
    new_state_dict = OrderedDict()

    for key, value in state_dict.items():
        new_key = key_transformation(key)
        new_state_dict[new_key] = value

    torch.save(new_state_dict, target)

test_rename_state_dict_keys.py

import unittest
import torch
import os
from rename_state_dict_keys import rename_state_dict_keys


class TestRenameStateDictKeys(unittest.TestCase):

    def test_renaming(self):
        """
        Define some modules for the test.
        """
        import torch.nn as nn
        from torch_state_control.nn import StatefulModule


        class SimpleModule(nn.Module):

            def __init__(self):
                super().__init__()
                self.layer = nn.Sequential(
                    nn.Linear(in_features=2, out_features=1, bias=False)
                )

            def forward(self, x):
                return self.layer(x)

        class SimpleModuleWithDropout(nn.Module):

            def __init__(self):
                super().__init__()
                self.layer = nn.Sequential(
                    nn.Dropout(p=0),
                    nn.Linear(in_features=2, out_features=1, bias=False)
                )

            def forward(self, x):
                return self.layer(x)


        """
        Try to load the state dict of SimpleModule into SimpleModuleWithDropout
        by renaming the parameters for the linear layer.
        """
        state_dict_path = './state_dict.torch'
        x = torch.tensor([1.0, 10])
        simple_module = SimpleModule()
        simple_module_with_dropout = SimpleModuleWithDropout()
        torch.save(simple_module.state_dict(), state_dict_path)

        # The test only works if at this point the results are different.
        self.assertNotEqual(simple_module(x), simple_module_with_dropout(x))

        # Before renaming, loading the state dict is expected to fail.
        with self.assertRaisesRegex(RuntimeError, 'Missing key'):
            loaded_state_dict = torch.load(state_dict_path)
            simple_module_with_dropout.load_state_dict(loaded_state_dict)

        # Rename the parameters.
        def key_transformation(old_key):
            if old_key == "layer.0.weight":
                return "layer.1.weight"

            return old_key
        rename_state_dict_keys(state_dict_path, key_transformation)

        # Loading the state dict should succeed now due to the renaming.
        loaded_state_dict = torch.load(state_dict_path)
        simple_module_with_dropout.load_state_dict(loaded_state_dict)

        # Since both modules should have the same parameter values now, the
        # results should be equal.
        self.assertEqual(simple_module(x), simple_module_with_dropout(x))

        # Clean up.
        os.remove(state_dict_path)

if __name__ == '__main__':
    unittest.main()

Can we rename the labels also after training the model?

I have a trained model. During training there was some problem with string characters. So i converted my labels into numbers like:

red : 0

blue: 1

green: 2

Now is it possible to rename my label back to actual label names. Hours of training. Would be helpful if anyone has an idea.