-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_fcvit_probs.py
173 lines (143 loc) · 6.3 KB
/
train_fcvit_probs.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
from transformers import ViTFeatureExtractor, ViTForImageClassification
from model.fcvit.modeling_fcvit import FCViTForImageClassification, FCViTForImageClassificationProbs
from transformers import TrainingArguments, Trainer
import numpy as np
from sklearn.model_selection import train_test_split
from torch.utils.data import Subset, DataLoader
import torch
import torchmetrics
import albumentations as A
from albumentations.core.transforms_interface import ImageOnlyTransform
from albumentations.pytorch import ToTensorV2
from data.dataset import ISICDataset
from datasets import load_metric
import os, sys
os.chdir(sys.path[0])
feature_extractor = ViTFeatureExtractor.from_pretrained('google/vit-base-patch16-224-in21k')
class VITPreprocess(ImageOnlyTransform):
def __init__(self, feature_extractor, always_apply: bool = True, p: float = 1.0):
super().__init__(always_apply, p)
self.feature_extractor = feature_extractor
def apply(self, img, **params):
return self.feature_extractor(img, data_format="channels_last")['pixel_values'][0]
train_transform = A.Compose(
[
A.SmallestMaxSize(max_size=450),
A.ShiftScaleRotate(shift_limit=0.05, scale_limit=0.05, rotate_limit=90, p=0.75),
A.RandomCrop(height=400, width=400),
A.RGBShift(r_shift_limit=15, g_shift_limit=15, b_shift_limit=15, p=0.5),
A.RandomBrightnessContrast(p=0.5),
VITPreprocess(feature_extractor),
ToTensorV2()
]
)
valid_transform = A.Compose(
[
VITPreprocess(feature_extractor),
ToTensorV2()
]
)
full_dataset = ISICDataset("../ISIC2019/ISIC_2019_Training_GroundTruth.csv", "../ISIC2019/TrainInput", transform=train_transform, val_transform=valid_transform)
train_size = int(0.8 * len(full_dataset))
valid_size = len(full_dataset) - train_size
SPLIT_SEED = 42
BATCH_SIZE = 64
EPOCHS = 20
EPOCHS_WARM_UP = 5
EPOCHS_FINAL = 5
MODEL_TYPE = "FCViTProbs"
if MODEL_TYPE == "ViT":
ARCHITECTURE = ViTForImageClassification
elif MODEL_TYPE == "FCViT":
ARCHITECTURE = FCViTForImageClassification
elif MODEL_TYPE == "FCViTProbs":
ARCHITECTURE = FCViTForImageClassificationProbs
train_indices, valid_indices, _, _ = train_test_split(
range(len(full_dataset)),
full_dataset.labels,
stratify=full_dataset.labels,
test_size=valid_size,
random_state=42)
full_dataset.set_indices(train_indices, valid_indices)
train_dataset = Subset(full_dataset, train_indices)
valid_dataset = Subset(full_dataset, valid_indices)
metric_acc = torchmetrics.Accuracy(task="multiclass", num_classes=full_dataset.classes, average="weighted", top_k=1)
metric_precision = torchmetrics.Precision(task = "multiclass", num_classes=full_dataset.classes, average="weighted", top_k=1)
metric_recall = torchmetrics.Recall(task = "multiclass", num_classes=full_dataset.classes, average="weighted", top_k=1)
metric_f1 = torchmetrics.F1Score(task = "multiclass", num_classes=full_dataset.classes, average="weighted", top_k=1)
metric_bacc = torchmetrics.Recall(task = "multiclass", num_classes=full_dataset.classes, average="macro", top_k=1)
def compute_metrics(p):
logits, labels = p
predictions=np.argmax(logits, axis=1)
predictions = torch.tensor(predictions)
labels = torch.tensor(labels)
acc = metric_acc(predictions, labels)
prec = metric_precision(predictions, labels)
recall = metric_recall(predictions, labels)
f1 = metric_f1(predictions, labels)
bacc = metric_bacc(predictions, labels)
#acc = metric_acc.compute(predictions=predictions, references=labels)
#prec = metric_prec.compute(predictions=predictions, references=labels, average="weighted")
#recall = metric_recall.compute(predictions=predictions, references=labels, average="weighted")
#f1 = metric_f1.compute(predictions=predictions, references=labels, average="weighted")
return {"accuracy": acc, "precision": prec, "recall": recall, "f1": f1, "bacc": bacc}
model = ARCHITECTURE.from_pretrained(
'google/vit-base-patch16-224-in21k',
num_labels = len(full_dataset.classes_names),
id2label = {str(i): c for i, c in enumerate(full_dataset.classes_names)},
label2id = {c: str(i) for i, c in enumerate(full_dataset.classes_names)}
)
max_steps = int( (EPOCHS_WARM_UP + EPOCHS + EPOCHS_FINAL)*len(train_dataset)/BATCH_SIZE/1000)*1000+1
warm_up_steps = (EPOCHS_WARM_UP)*len(train_dataset)/BATCH_SIZE
num_layers = model.config.num_hidden_layers
num_layers_finetune = 0
train_collate_step = 0
step_each = int(max_steps / num_layers)
def collate_fn(batch):
global train_collate_step, num_layers_finetune
images = torch.stack( [x[0] for x in batch])
labels = torch.tensor([x[1] for x in batch])
train_collate_step += 1
if train_collate_step > warm_up_steps and (train_collate_step-warm_up_steps) % step_each == 0:
num_layers_finetune += 1
if num_layers_finetune == num_layers: # we skip one layer
num_layers_finetune -= 1
return {
"pixel_values": images,
"labels": labels,
"num_layers_finetune": num_layers_finetune
}
training_args = TrainingArguments(
output_dir = f"../experiments/{MODEL_TYPE}",
per_device_train_batch_size=BATCH_SIZE,
evaluation_strategy="steps",
num_train_epochs=EPOCHS,
max_steps = max_steps,
save_steps=1000,
eval_steps=1000,
logging_steps=100,
learning_rate=2e-4,
save_total_limit=2,
remove_unused_columns=True,
push_to_hub=False,
report_to="none",
load_best_model_at_end=True,
metric_for_best_model="eval_bacc"
)
trainer = Trainer(
model=model,
args=training_args,
compute_metrics=compute_metrics,
data_collator=collate_fn,
train_dataset=train_dataset,
eval_dataset=valid_dataset,
tokenizer=feature_extractor
)
train_results = trainer.train(ignore_keys_for_eval=["all_logits"])
trainer.save_model()
trainer.log_metrics("train", train_results.metrics)
trainer.save_metrics("train", train_results.metrics)
trainer.save_state()
metrics = trainer.evaluate(valid_dataset, ignore_keys_for_eval=["all_logits"])
trainer.log_metrics("eval", metrics)
trainer.save_metrics("eval", metrics)