-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy pathsupernet_engine_prompt.py
185 lines (157 loc) · 7.9 KB
/
supernet_engine_prompt.py
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
import math
import sys
from typing import Iterable, Optional
from timm.utils.model import unwrap_model
import torch
from timm.data import Mixup
from timm.utils import accuracy, ModelEma
from lib import utils
import random
import time
def sample_configs(choices, is_visual_prompt_tuning=False,is_adapter=False,is_LoRA=False,is_prefix=False):
config = {}
depth = choices['depth']
if is_visual_prompt_tuning == False and is_adapter == False and is_LoRA == False and is_prefix==False:
visual_prompt_depth = random.choice(choices['visual_prompt_depth'])
lora_depth = random.choice(choices['lora_depth'])
adapter_depth = random.choice(choices['adapter_depth'])
prefix_depth = random.choice(choices['prefix_depth'])
config['visual_prompt_dim'] = [random.choice(choices['visual_prompt_dim']) for _ in range(visual_prompt_depth)] + [0] * (depth - visual_prompt_depth)
config['lora_dim'] = [random.choice(choices['lora_dim']) for _ in range(lora_depth)] + [0] * (depth - lora_depth)
config['adapter_dim'] = [random.choice(choices['adapter_dim']) for _ in range(adapter_depth)] + [0] * (depth - adapter_depth)
config['prefix_dim'] = [random.choice(choices['prefix_dim']) for _ in range(prefix_depth)] + [0] * (depth - prefix_depth)
else:
if is_visual_prompt_tuning:
config['visual_prompt_dim'] = [choices['super_prompt_tuning_dim']] * (depth)
else:
config['visual_prompt_dim'] = [0] * (depth)
if is_adapter:
config['adapter_dim'] = [choices['super_adapter_dim']] * (depth)
else:
config['adapter_dim'] = [0] * (depth)
if is_LoRA:
config['lora_dim'] = [choices['super_LoRA_dim']] * (depth)
else:
config['lora_dim'] = [0] * (depth)
if is_prefix:
config['prefix_dim'] = [choices['super_prefix_dim']] * (depth)
else:
config['prefix_dim'] = [0] * (depth)
return config
def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
data_loader: Iterable, optimizer: torch.optim.Optimizer,
device: torch.device, epoch: int, loss_scaler, max_norm: float = 0,
model_ema: Optional[ModelEma] = None, mixup_fn: Optional[Mixup] = None,
amp: bool = True, teacher_model: torch.nn.Module = None,
teach_loss: torch.nn.Module = None, choices=None, mode='super', retrain_config=None,is_visual_prompt_tuning=False,is_adapter=False,is_LoRA=False,is_prefix=False):
model.train()
criterion.train()
# set random seed
random.seed(epoch)
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=1, fmt='{value:.6f}'))
header = 'Epoch: [{}]'.format(epoch)
print_freq = 10
if mode == 'retrain':
config = retrain_config
model_module = unwrap_model(model)
print(config)
model_module.set_sample_config(config=config)
print(model_module.get_sampled_params_numel(config))
for samples, targets in metric_logger.log_every(data_loader, print_freq, header):
samples = samples.to(device, non_blocking=True)
targets = targets.to(device, non_blocking=True)
# sample random config
if mode == 'super':
# sample
config = sample_configs(choices=choices,is_visual_prompt_tuning=is_visual_prompt_tuning,is_adapter=is_adapter,is_LoRA=is_LoRA,is_prefix=is_prefix)
# print("current iter config: {}".format(config))
model_module = unwrap_model(model)
model_module.set_sample_config(config=config)
elif mode == 'retrain':
config = retrain_config
model_module = unwrap_model(model)
model_module.set_sample_config(config=config)
if mixup_fn is not None:
samples, targets = mixup_fn(samples, targets)
if amp:
with torch.cuda.amp.autocast():
if teacher_model:
with torch.no_grad():
teach_output = teacher_model(samples)
_, teacher_label = teach_output.topk(1, 1, True, True)
outputs = model(samples)
loss = 1/2 * criterion(outputs, targets) + 1/2 * teach_loss(outputs, teacher_label.squeeze())
else:
outputs = model(samples)
loss = criterion(outputs, targets)
else:
outputs = model(samples)
if teacher_model:
with torch.no_grad():
teach_output = teacher_model(samples)
_, teacher_label = teach_output.topk(1, 1, True, True)
loss = 1 / 2 * criterion(outputs, targets) + 1 / 2 * teach_loss(outputs, teacher_label.squeeze())
else:
loss = criterion(outputs, targets)
loss_value = loss.item()
if not math.isfinite(loss_value):
print("Loss is {}, stopping training".format(loss_value))
sys.exit(1)
optimizer.zero_grad()
# this attribute is added by timm on one optimizer (adahessian)
if amp:
is_second_order = hasattr(optimizer, 'is_second_order') and optimizer.is_second_order
loss_scaler(loss, optimizer, clip_grad=max_norm,
parameters=model.parameters(), create_graph=is_second_order)
else:
loss.backward()
optimizer.step()
torch.cuda.synchronize()
if model_ema is not None:
model_ema.update(model)
metric_logger.update(loss=loss_value)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger)
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}
@torch.no_grad()
def evaluate(data_loader, model, device, amp=True, choices=None, mode='super', retrain_config=None,is_visual_prompt_tuning=False,is_adapter=False,is_LoRA=False,is_prefix=False):
criterion = torch.nn.CrossEntropyLoss()
metric_logger = utils.MetricLogger(delimiter=" ")
header = 'Test:'
# switch to evaluation mode
model.eval()
if mode == 'super':
config = sample_configs(choices=choices,is_visual_prompt_tuning=is_visual_prompt_tuning,is_adapter=is_adapter,is_LoRA=is_LoRA,is_prefix=False)
model_module = unwrap_model(model)
model_module.set_sample_config(config=config)
else:
config = retrain_config
model_module = unwrap_model(model)
model_module.set_sample_config(config=config)
print("sampled model config: {}".format(config))
parameters = model_module.get_sampled_params_numel(config)
print("sampled model parameters: {}".format(parameters))
for images, target in metric_logger.log_every(data_loader, 10, header):
images = images.to(device, non_blocking=True)
target = target.to(device, non_blocking=True)
# compute output
if amp:
with torch.cuda.amp.autocast():
output = model(images)
loss = criterion(output, target)
else:
output = model(images)
loss = criterion(output, target)
acc1, acc5 = accuracy(output, target, topk=(1, 5))
batch_size = images.shape[0]
metric_logger.update(loss=loss.item())
metric_logger.meters['acc1'].update(acc1.item(), n=batch_size)
metric_logger.meters['acc5'].update(acc5.item(), n=batch_size)
# gather the stats from all processes
metric_logger.synchronize_between_processes()
print('* Acc@1 {top1.global_avg:.3f} Acc@5 {top5.global_avg:.3f} loss {losses.global_avg:.3f}'
.format(top1=metric_logger.acc1, top5=metric_logger.acc5, losses=metric_logger.loss))
return {k: meter.global_avg for k, meter in metric_logger.meters.items()}