File size: 10,036 Bytes
ad63418
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
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
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
import torch
from torch.optim import Optimizer
import math

class CustomAdamW(Optimizer):
    def __init__(self, params, lr = 1e-3, betas = (0.9, 0.95), eps = 1e-8, weight_decay = 0.1):
        defaults = dict(lr = lr, betas = betas, eps = eps, weight_decay = weight_decay)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure = None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            beta1, beta2 = group["betas"]
            eps = group["eps"]
            weight_decay = group["weight_decay"]

            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError("CustomAdamW does not support sparse gradients")

                state = self.state[p]

                if len(state) == 0:
                    state["step"] = 0
                    if "m" not in state:
                        state["m"] = torch.zeros_like(p.data)
                    if "v" not in state:
                        state["v"] = torch.zeros_like(p.data)

                state["step"] += 1

                step = state["step"]

                ## Moment Updation
                state["m"] = beta1 * state["m"] + (1 - beta1) * grad
                state["v"] = beta2 * state["v"] + (1 - beta2) * (grad ** 2)

                ## Bias Correction
                bias_correction1 = 1.0 - beta1 ** step
                m_hat = state["m"] / bias_correction1

                bias_correction2 = 1.0 - beta2 ** step
                v_hat = state["v"] / bias_correction2

                ## Weight Update
                p.data = p.data - lr * m_hat / (torch.sqrt(v_hat) + eps) - lr * p.data * weight_decay

        return loss

class CustomCautious(Optimizer):
    def __init__(self, params, lr = 1e-3, betas = (0.9, 0.95), eps = 1e-8, weight_decay = 0.1):
        defaults = dict(lr = lr, betas = betas, eps = eps, weight_decay = weight_decay)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure = None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            beta1, beta2 = group["betas"]
            eps = group["eps"]
            weight_decay = group["weight_decay"]

            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError("CustomAdamW does not support sparse gradients")

                state = self.state[p]

                if len(state) == 0:
                    state["step"] = 0
                    if "m" not in state:
                        state["m"] = torch.zeros_like(p.data)
                    if "v" not in state:
                        state["v"] = torch.zeros_like(p.data)

                state["step"] += 1

                step = state["step"]

                ## Moment Updation
                state["m"] = beta1 * state["m"] + (1 - beta1) * grad
                state["v"] = beta2 * state["v"] + (1 - beta2) * (grad ** 2)

                ## Bias Correction
                bias_correction1 = 1.0 - beta1 ** step
                m_hat = state["m"] / bias_correction1

                bias_correction2 = 1.0 - beta2 ** step
                v_hat = state["v"] / bias_correction2

                ## Calculate u
                u = m_hat / (torch.sqrt(v_hat) + eps)

                ## Calculate s
                s = (u * grad > 0).float()

                ## Calculate u_hat
                u_hat = (u * s) / torch.clamp(torch.mean(s), min = 1e-8)

                ## Weight Update
                p.data = p.data - lr * u_hat - lr * weight_decay * p.data

        return loss

class CustomLion(Optimizer):
    def __init__(self, params, lr = 2e-4, betas = (0.9, 0.99), weight_decay = 0.5):
        defaults = dict(lr = lr, betas = betas, weight_decay = weight_decay)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure = None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            beta1, beta2 = group["betas"]
            weight_decay = group["weight_decay"]

            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError("CustomAdamW does not support sparse gradients")

                state = self.state[p]

                if len(state) == 0:
                    state["step"] = 0
                    if "m" not in state:
                        state["m"] = torch.zeros_like(p.data)

                state["step"] += 1

                ## Calculate m_hat
                m_hat = beta1 * state["m"] + (1 - beta1) * grad

                ## Moment Updation
                state["m"] = beta2 * state["m"] + (1 - beta2) * grad

                ## Weight update
                p.data = p.data - lr * torch.sign(m_hat) - lr * weight_decay * p.data

        return loss

def newtonschulz_orthogonalization(X, steps = 5, eps = 1e-7):
    if X.ndim != 2:
        raise ValueError(f"expected 2D matrix, but got a tensor with {X.ndim} dimensions")

    rows, cols = X.shape
    is_transposed = False
    if rows > cols:
        X = X.T
        is_transposed = True

    norm = torch.linalg.matrix_norm(X)

    X = X / (norm + eps)

    a, b, c = (3.4445, -4.7750, 2.0315)

    for _ in range(1, steps + 1):
        first_term = a * X
        A = X @ X.T
        second_term = ((b * A) + c * (A @ A)) @ X
        X = first_term + second_term

    if is_transposed:
        X = X.T

    return X
    
    
class CustomScion(Optimizer):
    def __init__(self, params, lr = 1e-3, beta = 0.9, weight_decay = 0.01, steps = 5, eps = 1e-7):
        defaults = dict(lr = lr, beta = beta, weight_decay = weight_decay, steps = steps, eps = eps)
        super().__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure = None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            beta = group["beta"]
            weight_decay = group["weight_decay"]
            steps = group["steps"]
            eps = group["eps"]


            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError("CustomScion does not support sparse gradients")

                state = self.state[p]

                if len(state) == 0:
                    state["step"] = 0
                    if "m" not in state:
                        state["m"] = torch.zeros_like(p.data)

                state["step"] += 1

                ## Momentum update
                state["m"] = beta * state["m"] + grad

                ## Calculate u
                u = beta * state["m"] + grad

                ## Apply NewtonSchulz Orthogonalization
                if p.ndim == 2:
                    u = newtonschulz_orthogonalization(u, steps = steps, eps = eps)
                    rows, cols = p.data.shape
                    s = math.sqrt(max(1, rows / cols))
                    u = u * s
                else:
                    u = torch.sign(u)

                ## Weight update
                p.data = p.data - lr * u - lr * weight_decay * p.data

        return loss


class CustomSophia(Optimizer):
    def __init__(self, params, lr = 1e-4, betas = (0.965, 0.99), rho = 0.04, weight_decay = 0.1, eps = 1e-12):
        defaults = dict(lr = lr, betas = betas, rho = rho, weight_decay = weight_decay, eps = eps)
        super().__init__(params, defaults)

    @torch.no_grad()
    def update_hessian(self):
        for group in self.param_groups:
            beta1, beta2 = group["betas"]
            for p in group["params"]:
                if p.grad is None:
                    continue

                state = self.state[p]
                grad = p.grad
                if "h" not in state:
                    state["h"] = torch.zeros_like(p.data)
                state["h"] = beta2 * state["h"] + (1 - beta2) * (grad ** 2)

    @torch.no_grad()
    def step(self, closure = None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group["lr"]
            beta1, beta2 = group["betas"]
            rho = group["rho"]
            weight_decay = group["weight_decay"]
            eps = group["eps"]


            for p in group["params"]:
                if p.grad is None:
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError("CustomSophia does not support sparse gradients")

                state = self.state[p]

                if "step" not in state:
                    state["step"] = 0
                if "m" not in state:
                    state["m"] = torch.zeros_like(p.data)
                if "h" not in state:
                    state["h"] = torch.zeros_like(p.data)

                state["step"] += 1

                state["m"] = beta1 * state["m"] + (1 - beta1) * grad

                denom = torch.clamp(rho * state["h"], min = eps)
                ratio = torch.clamp(state["m"] / denom, min = -1.0, max = 1.0)

                p.data = p.data - lr * ratio - lr * weight_decay * p.data

        return loss