File size: 7,962 Bytes
9965bf6
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
Helpers to train with 16-bit precision.
"""

import numpy as np
import torch as th
import torch.nn as nn
from torch._utils import _flatten_dense_tensors, _unflatten_dense_tensors

from . import logger

INITIAL_LOG_LOSS_SCALE = 20.0


def convert_module_to_f16(l):
    """
    Convert primitive modules to float16.
    """
    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
        l.weight.data = l.weight.data.half()
        if l.bias is not None:
            l.bias.data = l.bias.data.half()


def convert_module_to_f32(l):
    """
    Convert primitive modules to float32, undoing convert_module_to_f16().
    """
    if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Conv3d)):
        l.weight.data = l.weight.data.float()
        if l.bias is not None:
            l.bias.data = l.bias.data.float()


def make_master_params(param_groups_and_shapes):
    """
    Copy model parameters into a (differently-shaped) list of full-precision
    parameters.
    """
    master_params = []
    for param_group, shape in param_groups_and_shapes:
        master_param = nn.Parameter(
            _flatten_dense_tensors(
                [param.detach().float() for (_, param) in param_group]
            ).view(shape)
        )
        master_param.requires_grad = True
        master_params.append(master_param)
    return master_params


def model_grads_to_master_grads(param_groups_and_shapes, master_params):
    """
    Copy the gradients from the model parameters into the master parameters
    from make_master_params().
    """
    for master_param, (param_group, shape) in zip(
        master_params, param_groups_and_shapes
    ):
        master_param.grad = _flatten_dense_tensors(
            [param_grad_or_zeros(param) for (_, param) in param_group]
        ).view(shape)


def master_params_to_model_params(param_groups_and_shapes, master_params):
    """
    Copy the master parameter data back into the model parameters.
    """
    # Without copying to a list, if a generator is passed, this will
    # silently not copy any parameters.
    for master_param, (param_group, _) in zip(master_params, param_groups_and_shapes):
        for (_, param), unflat_master_param in zip(
            param_group, unflatten_master_params(param_group, master_param.view(-1))
        ):
            param.detach().copy_(unflat_master_param)


def unflatten_master_params(param_group, master_param):
    return _unflatten_dense_tensors(master_param, [param for (_, param) in param_group])


def get_param_groups_and_shapes(named_model_params):
    named_model_params = list(named_model_params)
    scalar_vector_named_params = (
        [(n, p) for (n, p) in named_model_params if p.ndim <= 1],
        (-1),
    )
    matrix_named_params = (
        [(n, p) for (n, p) in named_model_params if p.ndim > 1],
        (1, -1),
    )
    return [scalar_vector_named_params, matrix_named_params]


def master_params_to_state_dict(
    model, param_groups_and_shapes, master_params, use_fp16
):
    if use_fp16:
        state_dict = model.state_dict()
        for master_param, (param_group, _) in zip(
            master_params, param_groups_and_shapes
        ):
            for (name, _), unflat_master_param in zip(
                param_group, unflatten_master_params(param_group, master_param.view(-1))
            ):
                assert name in state_dict
                state_dict[name] = unflat_master_param
    else:
        state_dict = model.state_dict()
        for i, (name, _value) in enumerate(model.named_parameters()):
            assert name in state_dict
            state_dict[name] = master_params[i]
    return state_dict


def state_dict_to_master_params(model, state_dict, use_fp16):
    if use_fp16:
        named_model_params = [
            (name, state_dict[name]) for name, _ in model.named_parameters()
        ]
        param_groups_and_shapes = get_param_groups_and_shapes(named_model_params)
        master_params = make_master_params(param_groups_and_shapes)
    else:
        master_params = [state_dict[name] for name, _ in model.named_parameters()]
    return master_params


def zero_master_grads(master_params):
    for param in master_params:
        param.grad = None


def zero_grad(model_params):
    for param in model_params:
        # Taken from https://pytorch.org/docs/stable/_modules/torch/optim/optimizer.html#Optimizer.add_param_group
        if param.grad is not None:
            param.grad.detach_()
            param.grad.zero_()


def param_grad_or_zeros(param):
    if param.grad is not None:
        return param.grad.data.detach()
    else:
        return th.zeros_like(param)


class MixedPrecisionTrainer:
    def __init__(
        self,
        *,
        model,
        use_fp16=False,
        fp16_scale_growth=1e-3,
        initial_lg_loss_scale=INITIAL_LOG_LOSS_SCALE,
    ):
        self.model = model
        self.use_fp16 = use_fp16
        self.fp16_scale_growth = fp16_scale_growth

        self.model_params = list(self.model.parameters())
        self.master_params = self.model_params
        self.param_groups_and_shapes = None
        self.lg_loss_scale = initial_lg_loss_scale

        if self.use_fp16:
            self.param_groups_and_shapes = get_param_groups_and_shapes(
                self.model.named_parameters()
            )
            self.master_params = make_master_params(self.param_groups_and_shapes)
            self.model.convert_to_fp16()

    def zero_grad(self):
        zero_grad(self.model_params)

    def backward(self, loss: th.Tensor):
        if self.use_fp16:
            loss_scale = 2 ** self.lg_loss_scale
            (loss * loss_scale).backward()
        else:
            loss.backward()

    def optimize(self, opt: th.optim.Optimizer):
        if self.use_fp16:
            return self._optimize_fp16(opt)
        else:
            return self._optimize_normal(opt)

    def _optimize_fp16(self, opt: th.optim.Optimizer):
        logger.logkv_mean("lg_loss_scale", self.lg_loss_scale)
        model_grads_to_master_grads(self.param_groups_and_shapes, self.master_params)
        grad_norm, param_norm = self._compute_norms(grad_scale=2 ** self.lg_loss_scale)
        if check_overflow(grad_norm):
            self.lg_loss_scale -= 1
            logger.log(f"Found NaN, decreased lg_loss_scale to {self.lg_loss_scale}")
            zero_master_grads(self.master_params)
            return False

        logger.logkv_mean("grad_norm", grad_norm)
        logger.logkv_mean("param_norm", param_norm)

        for p in self.master_params:
            p.grad.mul_(1.0 / (2 ** self.lg_loss_scale))
        opt.step()
        zero_master_grads(self.master_params)
        master_params_to_model_params(self.param_groups_and_shapes, self.master_params)
        self.lg_loss_scale += self.fp16_scale_growth
        return True

    def _optimize_normal(self, opt: th.optim.Optimizer):
        grad_norm, param_norm = self._compute_norms()
        logger.logkv_mean("grad_norm", grad_norm)
        logger.logkv_mean("param_norm", param_norm)
        opt.step()
        return True

    def _compute_norms(self, grad_scale=1.0):
        grad_norm = 0.0
        param_norm = 0.0
        for p in self.master_params:
            with th.no_grad():
                param_norm += th.norm(p, p=2, dtype=th.float32).item() ** 2
                if p.grad is not None:
                    grad_norm += th.norm(p.grad, p=2, dtype=th.float32).item() ** 2
        return np.sqrt(grad_norm) / grad_scale, np.sqrt(param_norm)

    def master_params_to_state_dict(self, master_params):
        return master_params_to_state_dict(
            self.model, self.param_groups_and_shapes, master_params, self.use_fp16
        )

    def state_dict_to_master_params(self, state_dict):
        return state_dict_to_master_params(self.model, state_dict, self.use_fp16)


def check_overflow(value):
    return (value == float("inf")) or (value == -float("inf")) or (value != value)