File size: 5,617 Bytes
faa3682
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import math

import jax
import jax.numpy as jnp
import ninjax as nj
import optax

from . import internal
from . import nets

f32 = jnp.float32
i32 = jnp.int32
sg = jax.lax.stop_gradient


class Optimizer(nj.Module):

  summary_depth: int = 2

  def __init__(self, modules, opt):
    modules = modules if isinstance(modules, (list, tuple)) else (modules,)
    self.modules = modules
    self.opt = opt
    self.step = nj.Variable(jnp.array, 0, i32, name='step')
    self.scaling = (nets.COMPUTE_DTYPE == jnp.float16)
    if self.scaling:
      self.opt = optax.apply_if_finite(self.opt, max_consecutive_errors=1000)
      self.grad_scale = nj.Variable(jnp.array, 1e4, f32, name='grad_scale')
      self.good_steps = nj.Variable(jnp.array, 0, i32, name='good_steps')

  def __call__(self, lossfn, *args, has_aux=False, **kwargs):
    metrics = {}

    def lossfn2(*args, **kwargs):
      outs = lossfn(*args, **kwargs)
      loss, aux = outs if has_aux else (outs, None)
      assert loss.dtype == f32, (self.name, loss.dtype)
      assert loss.shape == (), (self.name, loss.shape)
      if self.scaling:
        loss *= sg(self.grad_scale.read())
      return loss, aux

    loss, params, grads, aux = nj.grad(
        lossfn2, self.modules, has_aux=True)(*args, **kwargs)
    if self.scaling:
      loss *= 1 / self.grad_scale.read()

    counts = {k: math.prod(v.shape) for k, v in params.items()}
    if nj.creating():
      print(self._summarize_params(counts, self.summary_depth))

    axes = internal.get_data_axes()
    if axes:
      grads = jax.tree.map(lambda x: jax.lax.pmean(x, axes), grads)

    if self.scaling:
      invscale = 1 / self.grad_scale.read()
      grads = jax.tree.map(lambda x: x * invscale, grads)

    state = self.sub('state', nj.Tree, self.opt.init, params)
    updates, new_state = self.opt.update(grads, state.read(), params)
    nj.context().update(optax.apply_updates(params, updates))
    state.write(new_state)
    grad_norm = optax.global_norm(grads)
    if self.scaling:
      self._update_scale(grads, jnp.isfinite(grad_norm))
      grad_norm = jnp.where(jnp.isfinite(grad_norm), grad_norm, jnp.nan)
      self.step.write(self.step.read() + i32(jnp.isfinite(grad_norm)))
      metrics['grad_scale'] = self.grad_scale.read()
      metrics['grad_overflow'] = f32(~jnp.isfinite(grad_norm))
    else:
      self.step.write(self.step.read() + 1)
    metrics['loss'] = loss.mean()
    metrics['updates'] = self.step.read()
    metrics['grad_norm'] = grad_norm
    metrics['grad_rms'] = nets.rms(grads)
    metrics['update_rms'] = nets.rms(updates)
    metrics['param_rms'] = nets.rms([x.values for x in self.modules])
    metrics['param_count'] = jnp.array(list(counts.values()), f32).sum()
    metrics = {f'{self.name}/{k}': v for k, v in metrics.items()}
    return (metrics, aux) if has_aux else metrics

  def _update_scale(self, grads, finite):
    keep = (finite & (self.good_steps.read() < 1000))
    incr = (finite & (self.good_steps.read() >= 1000))
    decr = ~finite
    self.good_steps.write(i32(keep) * (self.good_steps.read() + 1))
    self.grad_scale.write(jnp.clip(
        f32(keep) * self.grad_scale.read() +
        f32(incr) * self.grad_scale.read() * 2 +
        f32(decr) * self.grad_scale.read() / 2, 1e-4, 1e5))
    return finite

  def _summarize_params(self, counts, depth):
    lines = []
    pfxs = []
    for key in counts:
      parts = key.split('/')
      pfxs += ['/'.join(parts[: i + 1]) for i in range(min(len(parts), depth))]
    subcounts = {
        prefix: sum(v for k, v in counts.items() if k.startswith(prefix))
        for prefix in set(pfxs)}
    lines = [f'Optimizer {self.name} has {sum(counts.values()):,} params:']
    for prefix, count in sorted(subcounts.items(), key=lambda x: -x[1]):
      lines.append(f'{count:>14,} {prefix}')
    return '\n'.join(lines)


def clip_by_agc(clip=0.3, pmin=1e-3):

  def init_fn(params):
    return ()

  def update_fn(updates, state, params=None):
    def fn(param, update):
      unorm = jnp.linalg.norm(update.flatten(), 2)
      pnorm = jnp.linalg.norm(param.flatten(), 2)
      upper = clip * jnp.maximum(pmin, pnorm)
      return update * (1 / jnp.maximum(1.0, unorm / upper))
    updates = jax.tree.map(fn, params, updates) if clip else updates
    return updates, ()

  return optax.GradientTransformation(init_fn, update_fn)


def scale_by_rms(beta=0.999, eps=1e-8):

  def init_fn(params):
    nu = jax.tree.map(lambda t: jnp.zeros_like(t, f32), params)
    step = jnp.zeros((), i32)
    return (step, nu)

  def update_fn(updates, state, params=None):
    step, nu = state
    step = optax.safe_int32_increment(step)
    nu = jax.tree.map(
        lambda v, u: beta * v + (1 - beta) * (u * u), nu, updates)
    nu_hat = optax.bias_correction(nu, beta, step)
    updates = jax.tree.map(
        lambda u, v: u / (jnp.sqrt(v) + eps), updates, nu_hat)
    return updates, (step, nu)

  return optax.GradientTransformation(init_fn, update_fn)


def scale_by_momentum(beta=0.9, nesterov=False):

  def init_fn(params):
    mu = jax.tree.map(lambda t: jnp.zeros_like(t, f32), params)
    step = jnp.zeros((), i32)
    return (step, mu)

  def update_fn(updates, state, params=None):
    step, mu = state
    step = optax.safe_int32_increment(step)
    mu = optax.update_moment(updates, mu, beta, 1)
    if nesterov:
      mu_nesterov = optax.update_moment(updates, mu, beta, 1)
      mu_hat = optax.bias_correction(mu_nesterov, beta, step)
    else:
      mu_hat = optax.bias_correction(mu, beta, step)
    return mu_hat, (step, mu)

  return optax.GradientTransformation(init_fn, update_fn)