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 | def get_optimizer_and_scheduler(
optimizer: AVAILABLE_OPTIMIZERS,
scheduler: AVAILABLE_SCHEDULERS,
steps: int,
learning_rate: float = 1e-5,
learning_rate_end: float = 1e-5,
gradient_accumulation_steps: int = 1,
extra_optimizer_kwargs: Optional[dict] = None,
weight_decay: float = 0.02,
warmup_steps: int = 0
):
"""
The get_optimizer_and_scheduler function is a helper function that returns an optimizer and scheduler
based on the parameters passed to it.
:param optimizer: AVAILABLE_OPTIMIZERS: Choose the optimizer
:param scheduler: AVAILABLE_SCHEDULERS: Determine the learning rate scheduler
:param steps: int: Specify the number of steps in the training process
:param learning_rate: float: Set the learning rate for the optimizer
:param learning_rate_end: float: Set the final learning rate
:param gradient_accumulation_steps: int: Accumulate the gradients before updating the weights
:param extra_optimizer_kwargs: dict | None: Pass extra arguments to the optimizer
:param weight_decay: float: Set the weight decay for adamw optimizer
:param warmup_steps: int: Specify the number of steps to warm up the learning rate
:return: A tuple of two objects: (Optimizer and scheduler)
"""
if extra_optimizer_kwargs is None:
extra_optimizer_kwargs = {}
if optimizer == EasyDeLOptimizers.ADAFACTOR:
if scheduler == EasyDeLSchedulers.LINEAR:
tx, sc = fjformer.optimizers.get_adafactor_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate_end,
gradient_accumulation_steps=gradient_accumulation_steps,
steps=steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.COSINE:
tx, sc = fjformer.optimizers.get_adafactor_with_cosine_scheduler(
learning_rate=learning_rate,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.NONE:
tx, sc = fjformer.optimizers.get_adafactor_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_COSINE:
tx, sc = fjformer.optimizers.get_adafactor_with_warm_up_cosine_scheduler(
learning_rate=learning_rate,
steps=steps,
weight_decay=weight_decay,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_LINEAR:
tx, sc = fjformer.optimizers.get_adafactor_with_warmup_linear_scheduler(
learning_rate_start=learning_rate,
steps=steps,
learning_rate_end=learning_rate_end,
gradient_accumulation_steps=gradient_accumulation_steps,
warmup_steps=warmup_steps,
**extra_optimizer_kwargs
)
else:
raise ValueError(
"seems like you have choose wrong type or unavailable scheduler"
)
elif optimizer == EasyDeLOptimizers.LION:
if scheduler == EasyDeLSchedulers.LINEAR:
tx, sc = fjformer.optimizers.get_lion_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate_end,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.COSINE:
tx, sc = fjformer.optimizers.get_lion_with_cosine_scheduler(
learning_rate=learning_rate,
gradient_accumulation_steps=gradient_accumulation_steps,
steps=steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.NONE:
tx, sc = fjformer.optimizers.get_lion_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_COSINE:
tx, sc = fjformer.optimizers.get_lion_with_warm_up_cosine_scheduler(
learning_rate=learning_rate,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_LINEAR:
tx, sc = fjformer.optimizers.get_lion_with_with_warmup_linear_scheduler(
learning_rate_start=learning_rate,
steps=steps,
learning_rate_end=learning_rate_end,
gradient_accumulation_steps=gradient_accumulation_steps,
warmup_steps=warmup_steps,
**extra_optimizer_kwargs
)
else:
raise ValueError(
"seems like you have choose wrong type or unavailable scheduler")
elif optimizer == EasyDeLOptimizers.ADAMW:
if scheduler == EasyDeLSchedulers.LINEAR:
tx, sc = fjformer.optimizers.get_adamw_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate_end,
steps=steps,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.COSINE:
tx, sc = fjformer.optimizers.get_adamw_with_cosine_scheduler(
learning_rate=learning_rate,
gradient_accumulation_steps=gradient_accumulation_steps,
steps=steps,
weight_decay=weight_decay,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.NONE:
tx, sc = fjformer.optimizers.get_adamw_with_linear_scheduler(
learning_rate_start=learning_rate,
learning_rate_end=learning_rate,
gradient_accumulation_steps=gradient_accumulation_steps,
steps=steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_COSINE:
tx, sc = fjformer.optimizers.get_adamw_with_warm_up_cosine_scheduler(
learning_rate=learning_rate,
steps=steps,
weight_decay=weight_decay,
gradient_accumulation_steps=gradient_accumulation_steps,
**extra_optimizer_kwargs
)
elif scheduler == EasyDeLSchedulers.WARM_UP_LINEAR:
tx, sc = fjformer.optimizers.get_adamw_with_warmup_linear_scheduler(
learning_rate_start=learning_rate,
steps=steps,
weight_decay=weight_decay,
learning_rate_end=learning_rate_end,
gradient_accumulation_steps=gradient_accumulation_steps,
warmup_steps=warmup_steps,
**extra_optimizer_kwargs
)
else:
raise ValueError(
"seems like you have choose wrong type or unavailable scheduler"
)
else:
raise ValueError(
f"seems like you have choose wrong type or unavailable optimizer {optimizer} and scheduler {scheduler}"
)
return tx, sc
|