forked from HSG-AIML/MaskedSST
-
Notifications
You must be signed in to change notification settings - Fork 0
/
finetune.py
223 lines (194 loc) · 7.58 KB
/
finetune.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
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
# %%
import os
# limit resource usage
os.environ["OMP_NUM_THREADS"] = "4" # export OMP_NUM_THREADS=4
os.environ["OPENBLAS_NUM_THREADS"] = "4" # export OPENBLAS_NUM_THREADS=4
os.environ["MKL_NUM_THREADS"] = "4" # export MKL_NUM_THREADS=6
os.environ["VECLIB_MAXIMUM_THREADS"] = "4" # export VECLIB_MAXIMUM_THREADS=4
os.environ["NUMEXPR_NUM_THREADS"] = "4" # export NUMEXPR_NUM_THREADS=6
os.environ["GDAL_NUM_THREADS"] = "4"
from pathlib import Path
import random
import sys
from accelerate import Accelerator
from aim import Run
import numpy as np
from rich.progress import track
import torch
from torchmetrics import Accuracy
from tqdm import tqdm
from DeepHyperX.models import get_model
from src.utils import (
get_finetune_config,
get_supervised_data,
get_val_epochs,
load_checkpoint,
train_step,
)
from src.utils import validate_downstream as validate
from src.vit_original import ViTRGB
from src.vit_spatial_spectral import ViTSpatialSpectral
import wandb
SEED = 5
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True
# %%
if __name__ == "__main__":
dataset_name = sys.argv[1]
valid_datasets = ["enmap", "houston2018"]
assert (
dataset_name in valid_datasets
), f"Please provide a valid dataset name from {valid_datasets}, provided: {dataset_name=}"
config = get_finetune_config(
f"configs/finetune_config_{dataset_name}.yaml",
"configs/config.yaml",
SEED,
)
run = wandb.init(config=config, project="downstream")
config.run_id = run.id
if config.method_name == "li":
model, optimizer, criterion, model_params = get_model(
name=config.method_name,
n_classes=config.n_classes,
n_bands=config.n_bands,
ignored_labels=[config.ignored_label],
patch_size=config.image_size - config.patch_sub, # one prediction per patch
)
elif config.method_name == "ViTSpatialSpectral":
model = ViTSpatialSpectral(
image_size=config.image_size - config.patch_sub,
spatial_patch_size=config.patch_size,
spectral_patch_size=config.band_patch_size,
num_classes=config.n_classes,
dim=config.transformer_dim,
depth=config.transformer_depth,
heads=config.transformer_n_heads,
mlp_dim=config.transformer_mlp_dim,
dropout=config.transformer_dropout,
emb_dropout=config.transformer_emb_dropout,
channels=config.n_bands,
spectral_pos=config.spectral_pos,
spectral_pos_embed=config.spectral_pos_embed,
blockwise_patch_embed=config.blockwise_patch_embed,
spectral_only=config.spectral_only,
pixelwise=config.pixelwise,
pos_embed_len=config.pos_embed_len,
)
elif config.method_name == "ViTRGB":
model = ViTRGB(
image_size=config.image_size,
patch_size=config.patch_size,
num_classes=config.n_classes,
dim=config.transformer_dim,
depth=config.transformer_depth,
heads=config.transformer_n_heads,
mlp_dim=config.transformer_mlp_dim,
dropout=config.transformer_dropout,
emb_dropout=config.transformer_emb_dropout,
channels=config.n_bands,
pixelwise=True, # one prediction per pixel, not per patch
)
else:
msg = f"method {config.method_name} not available"
raise NotImplementedError(msg)
classifier_name = "fc" if config.method_name == "li" else "mlp_head"
if config.checkpoint_path is not None:
model = load_checkpoint(config, model, classifier_name)
if config.linear_eval:
print("Linear evaluation... only training mlp_head")
for n, p in model.named_parameters():
if classifier_name not in n:
p.requires_grad = False
params = list(getattr(model, classifier_name).parameters())
else:
# fine-tuning
params = list(model.parameters())
# set different LR for transformer and MLP head
if config.lr != config.mlp_head_lr:
mlp_param_list = [
p for n, p in model.named_parameters() if classifier_name in n
]
rest_param_list = [
p for n, p in model.named_parameters() if classifier_name not in n
]
params = [
{"params": mlp_param_list, "lr": config.mlp_head_lr},
{"params": rest_param_list},
]
if config.method_name != "li" or config.overwrite_li_optim:
optimizer = torch.optim.Adam(
params, lr=config.lr, weight_decay=config.weight_decay
)
criterion = torch.nn.CrossEntropyLoss(ignore_index=config.ignored_label)
scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(
optimizer, factor=0.9, patience=5, verbose=True
)
acc_criterion = Accuracy(
"multiclass", num_classes=config.n_classes, average="macro"
)
model_params = sum([p.numel() for p in model.parameters()])
config.num_params = model_params
print(f"Model name: {config.method_name}")
print(f"Model parameters: {model_params:,}")
dataloader, val_dataloader = get_supervised_data(config)
# os.mkdir(f"models/{config.run_id}/")
Path(f"models/{config.run_id}/").mkdir(parents=True, exist_ok=True)
losses = []
accs = []
macro_accs = []
acc_per_epoch = []
current_val_acc = 0
best_val_acc = 0
step = 0
epoch = 0
validation_epochs = get_val_epochs(config, dataloader)
wandb.config.update(config)
epochs_pbar = tqdm(range(config.epoch + 1))
while epoch < config.epoch + 1 or step < config.max_steps + 1:
epochs_pbar.set_description(f"Epoch {epoch}")
model.train()
train_pbar = tqdm(enumerate(dataloader), total=len(dataloader), leave=False)
for idx, batch in train_pbar:
train_pbar.set_description(f"Training {step:,}")
img = batch["img"]
label = batch["label"]
loss, acc, macro_acc = train_step(
img, label, model, config, criterion, optimizer, acc_criterion
)
step += 1
losses.append(loss.detach().item())
accs.append(acc.detach().item())
macro_accs.append(macro_acc.detach().item())
if step % config.logging_freq == 0:
wandb.log(
{
"epoch": epoch,
"acc": np.array(accs[-1 * config.logging_freq :]).mean(),
"macro_acc": np.array(
macro_accs[-1 * config.logging_freq :]
).mean(),
"loss": np.array(losses[-1 * config.logging_freq :]).mean(),
"lr": optimizer.param_groups[0]["lr"],
},
step=step,
)
# log at end of training epoch (to same step as validation stats below)
wandb.log({"epoch": epoch, "acc": acc.item(), "loss": loss.item()}, step=step)
if epoch in validation_epochs:
val_losses, best_val_acc = validate(
config,
epoch,
model,
val_dataloader,
criterion,
acc_criterion,
step,
best_val_acc,
optimizer.param_groups[0]["lr"],
pixelwise=config.pixelwise,
)
scheduler.step(torch.tensor(val_losses).mean().item())
epoch += 1