adam comments

labml_nn/optimizers/__init__.py

@@ -67,14 +67,13 @@ class WeightDecay:
     def defaults(self):
         return dict(weight_decay=self.weight_decay)
 
-    def __call__(self, param: torch.nn.Parameter, group: Dict[str, any]):
-        grad = param.grad.data
-
+    def __call__(self, param: torch.nn.Parameter, grad: torch.Tensor, group: Dict[str, any]):
         if self.weight_decouple:
             if not self.absolute:
                 param.data.mul_(1.0 - group['lr'] * group['weight_decay'])
             else:
                 param.data.mul_(1.0 - group['weight_decay'])
+            return grad
         else:
             if group['weight_decay'] != 0:
-                grad.add_(param.data, alpha=group['weight_decay'])
+                return grad.add(param.data, alpha=group['weight_decay'])

labml_nn/optimizers/ada_belief.py

@@ -81,7 +81,7 @@ class AdaBelief(RAdam):
             return m, v
 
     def step_param(self, state: Dict[str, any], group: Dict[str, any], grad: torch.Tensor, param: torch.nn.Parameter):
-        self.weight_decay(param, group)
+        grad = self.weight_decay(param, grad, group)
         m, v = self.get_mv(state, group, grad)
         state['step'] += 1
 

labml_nn/optimizers/adam.py

@@ -6,6 +6,33 @@ This is an implementation of popular optimizer *Adam* from paper
 
 We extend the class `GenericAdaptiveOptimizer` defined in [__init__.py](index.html)
 to implement the Adam optimizer.
+
+*Adam* update is,
+
+\begin{align}
+m_t &\leftarrow \beta_1 m_{t-1} + (1 - \beta_1) \cdot g_t \\
+v_t &\leftarrow \beta_2 v_{t-1} + (1 - \beta_2) \cdot g_t^2 \\
+\hat{m}_t &\leftarrow \frac{m_t}/{1-\beta_1^t} \\
+\hat{v}_t &\leftarrow \frac{v_t}/{1-\beta_2^t} \\
+\theta_t &\leftarrow \theta_{t-1} - \alpha \cdot \frac{\hat{m}_t}{\sqrt{\hat{v}_t} + \epsilon}
+\end{align}
+
+where $\alpha$, $\beta_1$, $\beta_2$ and $\epsilon$ are scalar hyper parameters.
+$m_t$ and $v_t$ are first and second order moments.
+$\hat{m}_t$  and $\hat{v}_t$ are biased corrected moments.
+$\epsilon$ is used as a fix for division by zero error, but also acts as a form of a hyper-parameter
+that acts against variance in gradients.
+
+Effective step taken assuming $\epsilon = 0$ is,
+$$\Delta t = \alpha \cdot \frac{\hat{m}_t}{\hat{v}_t}$$
+This is bounded by,
+$$\vert \Delta t \vert \le \alpha \cdot \frac{1 - \beta_1}{\sqrt{1-\beta_2}}$$
+when $1-\beta_1 \gt \sqrt{1-\beta_2}$
+and
+$$\vert \Delta t\vert  \le \alpha$$
+otherwise.
+And in most common scenarios,
+$$\vert \Delta t \vert \approx \alpha$$
 """
 
 import math
@@ -28,6 +55,7 @@ class Adam(GenericAdaptiveOptimizer):
         * `params` is the list of parameters
         * 'lr' is the learning rate $\alpha$
         * `betas` is a tuple of ($\beta_1$, $\beta_2$)
+        * `eps` is $\hat{\epsilon}$
         * `weight_decay` is an instance of class `WeightDecay` defined in [__init__.py](index.html)
         * `defaults` is a dictionary of default for group values.
          This is useful when you want to extend the class `Adam`.
@@ -101,8 +129,6 @@ class Adam(GenericAdaptiveOptimizer):
         This computes the following
 
         \begin{align}
-        \hat{m}_t &\leftarrow \frac{m_t}/{1-\beta_1^t} \\
-        \hat{v}_t &\leftarrow \frac{v_t}/{1-\beta_2^t} \\
         \theta_t &\leftarrow \theta_{t-1} - \alpha \cdot \frac{\hat{m}_t}{\sqrt{\hat{v}_t} + \epsilon}
         \end{align}
 
@@ -114,12 +140,12 @@ class Adam(GenericAdaptiveOptimizer):
         \theta_t &\leftarrow \theta_{t-1} - \alpha \cdot
                 \frac{m_t / (1-\beta_1^t)}{\sqrt{v_t/(1-\beta_2^t)} + \epsilon} \\
         \theta_t &\leftarrow \theta_{t-1} - \alpha \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t} \cdot
-                \frac{m_t}{\sqrt{v_t} + \epsilon'} \\
+                \frac{m_t}{\sqrt{v_t} + \hat{\epsilon}} \\
         \end{align}
 
         where
-        $$\epsilon` = (1-\beta_2^t) \epsilon \approx \epsilon$$
-        since $\beta_2 \approx 1$
+        $$\hat{\epsilon} = (1-\beta_2^t) \epsilon$$
+        is what we should specify as the hyper-parameter.
         """
 
         # Get $\beta_1$ and $\beta_2$
@@ -134,7 +160,7 @@ class Adam(GenericAdaptiveOptimizer):
         # $\alpha \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t}$
         step_size = self.get_lr(state, group) * math.sqrt(bias_correction2) / bias_correction1
         # $\theta_t \leftarrow \theta_{t-1} - \alpha \frac{\sqrt{1-\beta_2^t}}{1-\beta_1^t} \cdot
-        #  \frac{m_t}{\sqrt{v_t} + \epsilon}$
+        #  \frac{m_t}{\sqrt{v_t} + \hat{\epsilon}}$
         param.data.addcdiv_(m, denominator, value=-step_size)
 
     def step_param(self, state: Dict[str, any], group: Dict[str, any], grad: torch.Tensor, param: torch.nn.Parameter):
@@ -148,7 +174,7 @@ class Adam(GenericAdaptiveOptimizer):
         """
 
         # Calculate weight decay
-        self.weight_decay(param, group)
+        grad = self.weight_decay(param, grad, group)
 
         # Get $m_t$ and $v_t$
         m, v = self.get_mv(state, group, grad)

labml_nn/optimizers/radam.py

@@ -19,7 +19,7 @@ class RAdam(AMSGrad):
         super().__init__(params, lr, betas, eps, weight_decay, amsgrad, defaults)
 
     def step_param(self, state: Dict[str, any], group: Dict[str, any], grad: torch.Tensor, param: torch.nn.Parameter):
-        self.weight_decay(param, group)
+        grad = self.weight_decay(param, grad, group)
 
         m, v = self.get_mv(state, group, grad)
         state['step'] += 1