-
Notifications
You must be signed in to change notification settings - Fork 0
/
main_vision.py
205 lines (179 loc) · 9.89 KB
/
main_vision.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
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
import os, sys, platform
import pathlib, os
import time
import numpy as np
import torch
import torch.nn as nn
from torch.utils.tensorboard import SummaryWriter
from args import args
import trainers.adaptors as adaptors
from models.gemresnet import GEMResNet18
from models.resnet import ResNet50
from models.small import LeNet
import trainers
from utils import utils
from utils.metrics import get_forgetting_metric
from utils import my_utils
from data.aval_datasets import AvalancheToPytorchDataset
def main():
before = time.time()
if args.seed is None:
args.seed = np.random.randint(10000)
print(f"SETTING SEED TO {args.seed}")
my_utils.seed_everything(args.seed)
# Make the a directory corresponding to this run for saving results, checkpoints etc.
i = 0
while True:
run_base_dir = pathlib.Path(f"{args.log_dir}/{args.name}~try={str(i)}")
if not run_base_dir.exists():
os.makedirs(run_base_dir)
args.name = args.name + f"-{platform.node().split('.')[0]}~try={i}"
break
i += 1
args.project_name = args.log_dir.split('/')[-1]
f = open(f'{run_base_dir}/logs.txt', 'w')
sys.stdout = my_utils.Tee(sys.stdout, f)
(run_base_dir / "settings.txt").write_text(str(args))
args.run_base_dir = run_base_dir
print(f"=> Saving data in {run_base_dir}")
print(args)
data_loader = AvalancheToPytorchDataset(dataset=args.dataname)
epoch_evaluate = getattr(adaptors, 'epoch_evaluate')
# Track accuracy on all tasks.
best_mask_acc = [0.0 for _ in range(args.num_tasks)]
last_acc = [0.0 for _ in range(args.num_tasks)]
all_test_acc = np.zeros((args.num_tasks,args.num_tasks))
# Get the model.
# model = utils.get_model(args.text_exp, data_loader.max_classes)
model_classes = data_loader.total_classes if args.strategy=='joint' else data_loader.max_classes
if 'cifar' in args.dataname:
print(f"Inititalizing Resnet18 model with {model_classes} classes!")
model = GEMResNet18(num_classes=model_classes)
elif 'imagenet' in args.dataname:
print(f"Inititalizing Resnet50 model with {model_classes} classes!")
model = ResNet50(num_classes=model_classes)
elif 'mnist' in args.dataname:
print(f"Inititalizing LeNet model with {model_classes} classes!")
model = LeNet(num_classes=model_classes)
else:
raise NotImplementedError("")
# If necessary, set the sparsity of the model of the model using the ER sparsity budget (see paper).
if args.er_sparsity != 'no':
for n, m in model.named_modules():
if hasattr(m, "sparsity"):
if args.er_sparsity == 'normal':
sp = args.sparsity
elif args.er_sparsity == 'er':
sp = args.sparsity * (m.weight.size(0) + m.weight.size(1)) / ( m.weight.size(0) * m.weight.size(1) * m.weight.size(2) * m.weight.size(3) )
m.sparsity = sp
if args.verbose: print(f"Set sparsity of {n} to {m.sparsity}")
# Put the model on the GPU.
model = utils.set_gpu(model)
criterion = nn.CrossEntropyLoss().to(args.device)
writer = SummaryWriter(log_dir=run_base_dir)
num_tasks_learned = 0
# Iterate through all tasks for training.
for curr_idx in range(args.num_tasks or 0):
print("\nTRAINING FOR MASKS\n")
if curr_idx > 0:
# Load best Checkpoint of the Last Task.
model = my_utils.load_best_training(model, run_base_dir / "local_best.pt")
## Update Dataloader and Task ##
print(f"Task {args.set}: {curr_idx}")
model.apply(lambda m: setattr(m, "task", curr_idx))
assert hasattr(data_loader, "update_task" ), "[ERROR] Need to implement update task method for use with multitask experiments"
data_loader.update_task(curr_idx)
# Train for masks.
if args.epochs > 0:
train, batch_evaluate = my_utils.get_train_test_function("default")
if curr_idx != 0 and curr_idx != args.num_tasks and args.sim_init != "":
if "knn" in args.sim_init:
print('Performing KNN classification to find similar tasks!')
task_accs = my_utils.findSimilarTasks(model, data_loader, num_tasks_learned, type=args.sim_init, num_topk=args.num_topk)
print(f"task accs: {task_accs}")
best_indices = np.array([])
if args.sim_init == "knn_best":
if task_accs.max() > 1/data_loader.max_classes:
best_indices = np.array([task_accs.argmax()])
elif args.sim_init == "knn_always":
best_indices = np.array([task_accs.argmax()])
elif args.sim_init == "knn_all":
best_indices = np.where(task_accs > 1/data_loader.max_classes)[0]
if best_indices.size == 0:
print(f'No Good tasks found.')
else:
print(f"Type: {type}\tBest Index: {best_indices}")
my_utils.score_transfer(model, best_indices, curr_idx)
model.apply(lambda m: setattr(m, "task", curr_idx))
data_loader.update_task(curr_idx)
else:
print("No Similarity based initialization.")
optimizer, scheduler, train_epochs = my_utils.get_optim_and_scheduler(model, optim_type=args.mask_opt, idx=curr_idx)
# Train on the scores for current task.
for epoch in range(1, train_epochs + 1):
print('\n')
train(model, writer, data_loader, optimizer, criterion, epoch, curr_idx,)
utils.cache_weights(model, num_tasks_learned + 1)
last_acc[curr_idx] = batch_evaluate( model, writer, criterion, data_loader, epoch, curr_idx, split='Val')
if last_acc[curr_idx] > best_mask_acc[curr_idx]:
best_mask_acc[curr_idx] = last_acc[curr_idx]
torch.save( { "epoch": args.epochs, "arch": args.model, "state_dict": model.state_dict(), "best_mask_acc": best_mask_acc,
"last_acc": last_acc, "args": args, }, run_base_dir / "local_best.pt",)
if scheduler:
scheduler[1].step()
scheduler[0].step(last_acc[curr_idx])
if ( args.iter_lim > 0 and len(data_loader.train_loader) * epoch > args.iter_lim ):
break
# caching masks and deleting optimizer and schedulers
utils.cache_masks(model)
del optimizer, scheduler
# Train on the weights for current task.
if args.weight_epochs > 0:
print("\nTRAINING FOR WEIGHTS\n")
optimizer, scheduler, train_epochs = my_utils.get_optim_and_scheduler(model, optim_type=args.weight_opt, idx=-1)
train, batch_evaluate = my_utils.get_train_test_function('weights')
get_editable_weights_mask_dict = getattr(trainers, "weights").get_editable_weights_mask_dict
weight_mask_dict, curr_act_dict = get_editable_weights_mask_dict(model, type=args.weight_mask_type)
for weight_epoch in range(1, args.weight_epochs+1):
train( model, writer, data_loader.train_loader, optimizer, criterion, weight_epoch, curr_idx, weight_mask_dict, curr_act_dict)
print('\n')
last_acc[curr_idx] = batch_evaluate( model, writer, criterion, data_loader, weight_epoch, curr_idx, split='Val' )
if last_acc[curr_idx] > best_mask_acc[curr_idx]:
best_mask_acc[curr_idx] = last_acc[curr_idx]
torch.save( { "epoch": args.epochs, "arch": args.model, "state_dict": model.state_dict(), "best_mask_acc": best_mask_acc,
"last_acc": last_acc, "args": args, }, run_base_dir / "local_best.pt",)
if scheduler:
scheduler[1].step()
scheduler[0].step(last_acc[curr_idx])
del optimizer, scheduler
num_tasks_learned += 1
model.apply(lambda m: setattr(m, "num_tasks_learned", num_tasks_learned))
# EVALUTATION ON ALL TASKS!
print('EPOCH END EVALUATION')
if num_tasks_learned in args.eval_ckpts or num_tasks_learned == args.num_tasks or args.eval_all:
# Evaluate until current task + 1 if not the last task
eval_tasks = (num_tasks_learned + 1) if curr_idx < args.num_tasks-1 else num_tasks_learned
for test_idx in range(eval_tasks):
for p in model.parameters(): p.grad = None
for b in model.buffers(): b.grad = None
all_test_acc[curr_idx, test_idx] = epoch_evaluate(model, data_loader, in_task=test_idx, out_task=test_idx, split='Test')
writer.add_scalar(f"task_test/acc_{test_idx}", all_test_acc[curr_idx, test_idx], curr_idx)
print(f"Adapt Test Accuracies: {all_test_acc[curr_idx,:]}")
print(f"Average Test Accuracy: {all_test_acc[curr_idx, :num_tasks_learned].mean():.4f}")
utils.clear_masks(model)
torch.cuda.empty_cache()
# printing stuff to console
if args.num_tasks > 1:
test_forgetting = get_forgetting_metric(all_test_acc, bwt=True)
print(f'Test Forgetting: {test_forgetting.tolist()}')
print(f'Average Test Forgetting: {test_forgetting.mean():.4f}')
overlap_obj = my_utils.getOverlap(model, args.num_tasks)
print(f"Sparse Overalp: {overlap_obj.mask_sparse_overlap.tolist()}")
print(f"Total Overalp: {overlap_obj.mask_total_overlap.tolist()}")
print(f"Avg. Sparse Overalp: {overlap_obj.avg_sparse_overlap:.8f}")
print(f"Avg. Total Overalp: {overlap_obj.avg_sparse_overlap:.8f}")
print( f"Finished experiment in {str((time.time() - before) / 60.0)} minutes." )
return all_test_acc
if __name__ == "__main__":
main()
pass