init commit

Author: Soohwan Kim

.gitignore

@@ -6,6 +6,16 @@ __pycache__/
 # C extensions
 *.so
 
+.DS_Store
+*.bin
+*.zip
+*.idea
+venv/*
+*.pyc
+.idea
+.ipynb_checkpoints
+.DS_Store/*
+
 # Distribution / packaging
 .Python
 build/

README.md

@@ -1,2 +1,189 @@
 # pytorch-lr-scheduler
 PyTorch implementation of some learning rate schedulers for deep learning researcher.
+  
+## Usage
+  
+### `ReduceLROnPlateauScheduler`  
+  
+- Example code
+  
+```python
+import torch
+
+from lr_scheduler.reduce_lr_on_plateau_lr_scheduler import ReduceLROnPlateauScheduler
+
+if __name__ == '__main__':
+    max_epochs, steps_in_epoch = 10, 10000
+
+    model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
+    optimizer = torch.optim.Adam(model, 1e-4)
+
+    scheduler = ReduceLROnPlateauScheduler(optimizer, patience=1, factor=0.3)
+
+    for epoch in range(max_epochs):
+        for timestep in range(steps_in_epoch):
+            ...
+            ...
+        
+        val_loss = validate()
+        scheduler.step(val_loss)
+```
+  
+- Visualize
+![](images/ReduceLROnPlateauScheduler.png)
+
+### `TransformerLRScheduler`
+  
+- Example code
+  
+```python
+import torch
+
+from lr_scheduler.transformer_lr_scheduler import TransformerLRScheduler
+
+if __name__ == '__main__':
+    max_epochs, steps_in_epoch = 10, 10000
+
+    model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
+    optimizer = torch.optim.Adam(model, 1e-10)
+
+    scheduler = TransformerLRScheduler(
+        optimizer=optimizer, 
+        init_lr=1e-10, 
+        peak_lr=0.1,
+        final_lr=1e-4, 
+        final_lr_scale=0.05,
+        warmup_steps=3000, 
+        decay_steps=17000,
+    )
+
+    for epoch in range(max_epochs):
+        for timestep in range(steps_in_epoch):
+            ...
+            ...
+            scheduler.step()
+```
+
+- Visualize
+![](images/TransformerLRScheduler.png)
+
+### `TriStageLRScheduler`
+  
+- Example code
+  
+```python
+import torch
+
+from lr_scheduler.tri_stage_lr_scheduler import TriStageLRScheduler
+
+if __name__ == '__main__':
+    max_epochs, steps_in_epoch = 10, 10000
+
+    model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
+    optimizer = torch.optim.Adam(model, 1e-10)
+
+    scheduler = TriStageLRScheduler(
+        optimizer, 
+        init_lr=1e-10, 
+        peak_lr=1e-4, 
+        final_lr=1e-7, 
+        init_lr_scale=0.01, 
+        final_lr_scale=0.05,
+        warmup_steps=30000, 
+        hold_steps=70000, 
+        decay_steps=100000,
+        total_steps=200000,
+    )
+
+    for epoch in range(max_epochs):
+        for timestep in range(steps_in_epoch):
+            ...
+            ...
+            scheduler.step()
+```
+  
+- Visualize
+  
+![](images/TriStageLRScheduler.png)
+  
+### `WarmupReduceLROnPlateauScheduler`
+  
+- Example code
+```python
+import torch
+
+from lr_scheduler.warmup_reduce_lr_on_plateau_scheduler import WarmupReduceLROnPlateauScheduler
+
+if __name__ == '__main__':
+    max_epochs, steps_in_epoch = 10, 10000
+
+    model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
+    optimizer = torch.optim.Adam(model, 1e-10)
+
+    scheduler = WarmupReduceLROnPlateauScheduler(
+        optimizer, 
+        init_lr=1e-10, 
+        peak_lr=1e-4, 
+        warmup_steps=30000, 
+        patience=1,
+        factor=0.3,
+    )
+
+    for epoch in range(max_epochs):
+        for timestep in range(steps_in_epoch):
+            ...
+            ...
+            if timestep < warmup_steps:
+                scheduler.step()
+                
+        val_loss = validate()
+        scheduler.step(val_loss)
+```
+  
+- Visualize
+  
+![](images/WarmupReduceLROnPlateauScheduler.png)
+  
+### `WarmupLRScheduler`
+  
+- Example code
+  
+```python
+import torch
+
+from lr_scheduler.warmup_lr_scheduler import WarmupLRScheduler
+
+if __name__ == '__main__':
+    max_epochs, steps_in_epoch = 10, 10000
+
+    model = [torch.nn.Parameter(torch.randn(2, 2, requires_grad=True))]
+    optimizer = torch.optim.Adam(model, 1e-10)
+
+    scheduler = WarmupLRScheduler(
+        optimizer, 
+        init_lr=1e-10, 
+        peak_lr=1e-4, 
+        warmup_steps=4000,
+    )
+
+    for epoch in range(max_epochs):
+        for timestep in range(steps_in_epoch):
+            ...
+            ...
+            scheduler.step()
+```
+  
+- Visualize
+  
+![](images/WarmupLRScheduler.png)
+  
+## Troubleshoots and Contributing
+If you have any questions, bug reports, and feature requests, please [open an issue](https://github.com/sooftware/openspeech/issues) on Github.   
+  
+I appreciate any kind of feedback or contribution.  Feel free to proceed with small issues like bug fixes, documentation improvement.  For major contributions and new features, please discuss with the collaborators in corresponding issues.
+  
+### Code Style
+I follow [PEP-8](https://www.python.org/dev/peps/pep-0008/) for code style. Especially the style of docstrings is important to generate documentation. 
+  
+### License
+This project is licensed under the MIT LICENSE - see the [LICENSE.md](https://github.com/sooftware/OpenSpeech/blob/master/LICENSE) file for details

images/ReduceLROnPlateauScheduler.png

@@ -0,0 +1,45 @@
+# MIT License
+#
+# Copyright (c) 2021 Soohwan Kim
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+
+class LearningRateScheduler(object):
+    r"""
+    Provides inteface of learning rate scheduler.
+
+    Note:
+        Do not use this class directly, use one of the sub classes.
+    """
+    def __init__(self, optimizer, lr):
+        self.optimizer = optimizer
+        self.lr = lr
+
+    def step(self, *args, **kwargs):
+        raise NotImplementedError
+
+    @staticmethod
+    def set_lr(optimizer, lr):
+        for g in optimizer.param_groups:
+            g['lr'] = lr
+
+    def get_lr(self):
+        for g in self.optimizer.param_groups:
+            return g['lr']

images/TransformerLRScheduler.png

@@ -0,0 +1,69 @@
+# MIT License
+#
+# Copyright (c) 2021 Soohwan Kim
+#
+# Permission is hereby granted, free of charge, to any person obtaining a copy
+# of this software and associated documentation files (the "Software"), to deal
+# in the Software without restriction, including without limitation the rights
+# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+# copies of the Software, and to permit persons to whom the Software is
+# furnished to do so, subject to the following conditions:
+#
+# The above copyright notice and this permission notice shall be included in all
+# copies or substantial portions of the Software.
+#
+# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+# SOFTWARE.
+
+from omegaconf import DictConfig
+from torch.optim import Optimizer
+
+from lr_scheduler.lr_scheduler import LearningRateScheduler
+
+
+class ReduceLROnPlateauScheduler(LearningRateScheduler):
+    r"""
+    Reduce learning rate when a metric has stopped improving. Models often benefit from reducing the learning rate by
+    a factor of 2-10 once learning stagnates. This scheduler reads a metrics quantity and if no improvement is seen
+    for a ‘patience’ number of epochs, the learning rate is reduced.
+
+    Args:
+        optimizer (Optimizer): Optimizer.
+        lr (float): Initial learning rate.
+        patience (int): Number of epochs with no improvement after which learning rate will be reduced.
+        factor (float): Factor by which the learning rate will be reduced. new_lr = lr * factor.
+    """
+    def __init__(
+            self,
+            optimizer: Optimizer,
+            lr: float,
+            patience: int = 1,
+            factor: float = 0.3,
+    ) -> None:
+        super(ReduceLROnPlateauScheduler, self).__init__(optimizer, lr)
+        self.lr = lr
+        self.patience = patience
+        self.factor = factor
+        self.val_loss = 100.0
+        self.count = 0
+
+    def step(self, val_loss: float):
+        if val_loss is not None:
+            if self.val_loss < val_loss:
+                self.count += 1
+                self.val_loss = val_loss
+            else:
+                self.count = 0
+                self.val_loss = val_loss
+
+            if self.patience == self.count:
+                self.count = 0
+                self.lr *= self.factor
+                self.set_lr(self.optimizer, self.lr)
+
+        return self.lr