remove RMSprop, nobody uses it anymore

README.md

@@ -81,7 +81,7 @@ Change to `DEBUG=4` to see the generated code.
 
 ## Neural networks?
 
-It turns out, a decent autograd tensor library is 90% of what you need for neural networks. Add an optimizer (SGD, RMSprop, and Adam implemented) from tinygrad.nn.optim, write some boilerplate minibatching code, and you have all you need.
+It turns out, a decent autograd tensor library is 90% of what you need for neural networks. Add an optimizer (SGD, Adam, AdamW implemented) from tinygrad.nn.optim, write some boilerplate minibatching code, and you have all you need.
 
 ### Neural network example (from test/models/test_mnist.py)
 

test/models/test_mnist.py

@@ -108,12 +108,5 @@ class TestMNIST(unittest.TestCase):
     train(model, X_train, Y_train, optimizer, steps=600)
     assert evaluate(model, X_test, Y_test) > 0.94   # CPU gets 0.9494 sometimes
 
-  def test_rmsprop(self):
-    np.random.seed(1337)
-    model = TinyBobNet()
-    optimizer = optim.RMSprop(model.parameters(), lr=0.0002, alpha=0.9)
-    train(model,  X_train, Y_train, optimizer, steps=400)
-    assert evaluate(model, X_test, Y_test) > 0.95
-
 if __name__ == '__main__':
   unittest.main()

test/test_optim.py

@@ -2,7 +2,7 @@ import numpy as np
 import torch
 import unittest
 from tinygrad.tensor import Tensor
-from tinygrad.nn.optim import Adam, SGD, RMSprop, AdamW
+from tinygrad.nn.optim import Adam, SGD, AdamW
 
 np.random.seed(1337)
 x_init = np.random.randn(1,4).astype(np.float32)
@@ -39,7 +39,6 @@ class TestOptim(unittest.TestCase):
       np.testing.assert_allclose(x, y, atol=atol, rtol=rtol)
 
   def _test_sgd(self, steps, opts, atol, rtol): self._test_optim(SGD, torch.optim.SGD, steps, opts, atol, rtol)
-  def _test_rmsprop(self, steps, opts, atol, rtol): self._test_optim(RMSprop, torch.optim.RMSprop, steps, opts, atol, rtol)
   def _test_adam(self, steps, opts, atol, rtol): self._test_optim(Adam, torch.optim.Adam, steps, opts, atol, rtol)
   def _test_adamw(self, steps, opts, atol, rtol): self._test_optim(AdamW, torch.optim.AdamW, steps, opts, atol, rtol)
 
@@ -55,19 +54,14 @@ class TestOptim(unittest.TestCase):
   def test_multistep_sgd_nesterov_momentum(self): self._test_sgd(10, {'lr': 0.001, 'momentum': 0.9, 'nesterov': True}, 1e-5, 0)
   def test_multistep_sgd_high_lr_nesterov_momentum(self): self._test_sgd(10, {'lr': 10, 'momentum': 0.9, 'nesterov': True}, 1e-5, 3e-4)
 
-  def test_rmsprop(self): self._test_rmsprop(1, {'lr': 0.001, 'alpha': 0.99}, 1e-5, 0)
-  def test_rmsprop_high_lr(self): self._test_rmsprop(1, {'lr': 10, 'alpha': 0.99}, 1e-5, 1e-5)
   def test_adam(self): self._test_adam(1, {'lr': 0.001}, 1e-5, 0)
   def test_adam_high_lr(self): self._test_adam(1, {'lr': 10}, 1e-5, 1e-5)
   def test_adamw(self): self._test_adamw(1, {'lr': 0.001}, 1e-5, 0)
   def test_adamw_high_lr(self): self._test_adamw(1, {'lr': 10}, 1e-5, 1e-5)
 
-  def test_multistep_rmsprop(self): self._test_rmsprop(10, {'lr': 0.001}, 1e-5, 0)
-  def test_multistep_rmsprop_high_lr(self): self._test_rmsprop(10, {'lr': 10}, 1e-5, 3e-4)
-
   def test_multistep_adam(self): self._test_adam(10, {'lr': 0.001}, 1e-5, 0)
   def test_multistep_adam_high_lr(self): self._test_adam(10, {'lr': 10}, 1e-5, 3e-4)
-  
+
   def test_multistep_adamw(self): self._test_adamw(10, {'lr': 0.001}, 1e-5, 0)
   def test_multistep_adamw_high_lr(self): self._test_adamw(10, {'lr': 10}, 1e-5, 3e-4)
 

tinygrad/nn/optim.py

@@ -11,13 +11,6 @@ class Optimizer:
     self.params: List[Tensor] = [x for x in params if x.requires_grad]
     self.buffers: List[Tensor] = [x for x in params if not x.requires_grad]   # buffers are still realized
 
-  # TODO: this probably shouldn't change the gradients, just the ones used by the optimizer
-  def clipnorm(self, amount=1):
-    for param in self.params:
-      assert param.grad is not None
-      # clipnorm is the L2 norm, not value: is this right?
-      param.grad.assign(param.grad.clip(-(amount**2), (amount**2)))
-
   def zero_grad(self):
     for param in self.params: param.grad = None
 
@@ -44,21 +37,6 @@ class SGD(Optimizer):
       t.assign(t.detach() - g * self.lr)
     self.realize(self.b)
 
-class RMSprop(Optimizer):
-  def __init__(self, params: List[Tensor], lr=0.001, alpha=0.99, eps=1e-8):
-    super().__init__(params)
-    self.lr, self.alpha, self.eps = lr, alpha, eps
-
-    self.v = [Tensor.zeros(*t.shape, device=t.device, requires_grad=False) for t in self.params]
-
-  def step(self) -> None:
-    for i, t in enumerate(self.params):
-      assert t.grad is not None
-      g = t.grad.realize()
-      self.v[i].assign(self.alpha * self.v[i] + (1.0 - self.alpha) * (g * g)).realize()
-      t.assign(t.detach() - (g * self.lr).div(self.v[i].sqrt() + self.eps))
-    self.realize(self.v)
-
 class AdamW(Optimizer):
   def __init__(self, params: List[Tensor], lr=0.001, b1=0.9, b2=0.999, eps=1e-8, wd=0.01):
     super().__init__(params)