File size: 11,523 Bytes
db6ee6a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
# os.environ["CUDA_VISIBLE_DEVICES"] = "6"
import argparse
import json
from collections import defaultdict

import numpy as np
import pytorch_lightning as pl
import torch
import wandb
from pytorch_lightning.callbacks import ModelCheckpoint
from sklearn.metrics import accuracy_score, classification_report, jaccard_score, roc_auc_score
from torch.nn import BCEWithLogitsLoss
from torch.utils.data import DataLoader
from torchinfo import summary
from tqdm import tqdm
from transformers import AdamW

from findings_classifier.chexpert_dataset import Chexpert_Dataset
from findings_classifier.chexpert_model import ChexpertClassifier
from local_config import WANDB_ENTITY

class ExpandChannels:
    """
    Transforms an image with one channel to an image with three channels by copying
    pixel intensities of the image along the 1st dimension.
    """

    def __call__(self, data: torch.Tensor) -> torch.Tensor:
        """
        :param data: Tensor of shape [1, H, W].
        :return: Tensor with channel copied three times, shape [3, H, W].
        """
        if data.shape[0] != 1:
            raise ValueError(f"Expected input of shape [1, H, W], found {data.shape}")
        return torch.repeat_interleave(data, 3, dim=0)

class LitIGClassifier(pl.LightningModule):
    def __init__(self, num_classes, class_names, class_weights=None, learning_rate=1e-5):
        super().__init__()

        # Model
        self.model = ChexpertClassifier(num_classes)

        # Loss with class weights
        if class_weights is None:
            self.criterion = BCEWithLogitsLoss()
        else:
            self.criterion = BCEWithLogitsLoss(pos_weight=class_weights)

        # Learning rate
        self.learning_rate = learning_rate
        self.class_names = class_names

    def forward(self, x):
        return self.model(x)

    def step(self, batch, batch_idx):
        x, y = batch['image'].to(self.device), batch['labels'].to(self.device)
        logits = self(x)
        loss = self.criterion(logits, y)

        # Apply sigmoid to get probabilities
        preds_probs = torch.sigmoid(logits)

        # Get predictions as boolean values
        preds = preds_probs > 0.5

        # calculate jaccard index
        jaccard = jaccard_score(y.cpu().numpy(), preds.detach().cpu().numpy(), average='samples')

        class_report = classification_report(y.cpu().numpy(), preds.detach().cpu().numpy(), output_dict=True)
        # scores = class_report['micro avg']
        scores = class_report['macro avg']
        metrics_per_label = {label: metrics for label, metrics in class_report.items() if label.isdigit()}

        f1 = scores['f1-score']
        rec = scores['recall']
        prec = scores['precision']
        acc = accuracy_score(y.cpu().numpy().flatten(), preds.detach().cpu().numpy().flatten())
        try:
            auc = roc_auc_score(y.cpu().numpy().flatten(), preds_probs.detach().cpu().numpy().flatten())
        except Exception as e:
            auc = 0.

        return loss, acc, f1, rec, prec, jaccard, auc, metrics_per_label

    def training_step(self, batch, batch_idx):
        loss, acc, f1, rec, prec, jaccard, auc, _ = self.step(batch, batch_idx)
        train_stats = {'loss': loss, 'train_acc': acc, 'train_f1': f1, 'train_rec': rec, 'train_prec': prec, 'train_jaccard': jaccard,
                       'train_auc': auc}
        wandb_run.log(train_stats)
        return train_stats

    def training_epoch_end(self, outputs):
        avg_loss = torch.stack([x['loss'] for x in outputs]).mean()
        avg_acc = np.mean([x['train_acc'] for x in outputs])
        avg_f1 = np.mean([x['train_f1'] for x in outputs])
        avg_rec = np.mean([x['train_rec'] for x in outputs])
        avg_prec = np.mean([x['train_prec'] for x in outputs])
        avg_jaccard = np.mean([x['train_jaccard'] for x in outputs])
        avg_auc = np.mean([x['train_auc'] for x in outputs])
        wandb_run.log({'epoch_train_loss': avg_loss, 'epoch_train_acc': avg_acc, 'epoch_train_f1': avg_f1, 'epoch_train_rec': avg_rec,
                       'epoch_train_prec': avg_prec, 'epoch_train_jaccard': avg_jaccard, 'epoch_train_auc': avg_auc})

    def validation_step(self, batch, batch_idx):
        loss, acc, f1, rec, prec, jaccard, auc, metrics_per_label = self.step(batch, batch_idx)
        # log f1 for checkpoint callback
        self.log('val_f1', f1)
        return {'val_loss': loss, 'val_acc': acc, 'val_f1': f1, 'val_rec': rec, 'val_prec': prec, 'val_jaccard': jaccard,
                'val_auc': auc}, metrics_per_label

    def validation_epoch_end(self, outputs):
        outputs, per_label_metrics_outputs = zip(*outputs)
        avg_loss = torch.stack([x['val_loss'] for x in outputs]).mean()
        avg_acc = np.mean([x['val_acc'] for x in outputs])
        avg_f1 = np.mean([x['val_f1'] for x in outputs])
        avg_rec = np.mean([x['val_rec'] for x in outputs])
        avg_prec = np.mean([x['val_prec'] for x in outputs])
        avg_jaccard = np.mean([x['val_jaccard'] for x in outputs])
        avg_auc = np.mean([x['val_auc'] for x in outputs])

        per_label_metrics = defaultdict(lambda: defaultdict(float))
        label_counts = defaultdict(int)
        for metrics_per_label in per_label_metrics_outputs:
            for label, metrics in metrics_per_label.items():
                label_name = self.class_names[int(label)]
                per_label_metrics[label_name]['precision'] += metrics['precision']
                per_label_metrics[label_name]['recall'] += metrics['recall']
                per_label_metrics[label_name]['f1-score'] += metrics['f1-score']
                per_label_metrics[label_name]['support'] += metrics['support']
                label_counts[label_name] += 1

        # Average the metrics
        for label, metrics in per_label_metrics.items():
            for metric_name in ['precision', 'recall', 'f1-score']:
                if metrics['support'] > 0:
                    per_label_metrics[label][metric_name] /= label_counts[label]

        val_stats = {'val_loss': avg_loss, 'val_acc': avg_acc, 'val_f1': avg_f1, 'val_rec': avg_rec, 'val_prec': avg_prec, 'val_jaccard': avg_jaccard,
                     'val_auc': avg_auc}
        wandb_run.log(val_stats)

    def test_step(self, batch, batch_idx):
        loss, acc, f1, rec, prec, jaccard, auc, _ = self.step(batch, batch_idx)
        return {'test_loss': loss, 'test_acc': acc, 'test_f1': f1, 'test_rec': rec, 'test_prec': prec, 'test_jaccard': jaccard, 'test_auc': auc}

    def test_epoch_end(self, outputs):
        avg_loss = torch.stack([x['test_loss'] for x in outputs]).mean()
        avg_acc = np.mean([x['test_acc'] for x in outputs])
        avg_f1 = np.mean([x['test_f1'] for x in outputs])
        avg_rec = np.mean([x['test_rec'] for x in outputs])
        avg_prec = np.mean([x['test_prec'] for x in outputs])
        avg_jaccard = np.mean([x['test_jaccard'] for x in outputs])
        avg_auc = np.mean([x['test_auc'] for x in outputs])

        test_stats = {'test_loss': avg_loss, 'test_acc': avg_acc, 'test_f1': avg_f1, 'test_rec': avg_rec, 'test_prec': avg_prec,
                      'test_jaccard': avg_jaccard, 'test_auc': avg_auc}
        wandb_run.log(test_stats)

    def configure_optimizers(self):
        optimizer = AdamW(self.parameters(), lr=self.learning_rate)
        return optimizer


def save_preds(dataloader, split):
    # load checkpoint
    ckpt_path = f"findings_classifier/checkpoints/chexpert_train/ChexpertClassifier-epoch=06-val_f1=0.36.ckpt"
    model = LitIGClassifier.load_from_checkpoint(ckpt_path, num_classes=num_classes, class_weights=val_dataset.get_class_weights(),
                                                 class_names=class_names, learning_rate=args.lr)
    model.eval()
    model.cuda()
    model.half()
    class_names_np = np.asarray(class_names)

    # get predictions for all study ids
    structured_preds = {}
    for batch in tqdm(dataloader):
        dicom_ids = batch['dicom_id']
        logits = model(batch['image'].half().cuda())
        preds_probs = torch.sigmoid(logits)
        preds = preds_probs > 0.5

        # iterate over each study id in the batch
        for i, (dicom_id, pred) in enumerate(zip(dicom_ids, preds.detach().cpu())):
            # get all positive labels
            findings = class_names_np[pred].tolist()
            structured_preds[dicom_id] = findings

    # save predictions
    with open(f"findings_classifier/predictions/structured_preds_chexpert_log_weighting_macro_{split}.json", "w") as f:
        json.dump(structured_preds, f, indent=4)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--run_name", type=str, default="debug")
    parser.add_argument("--lr", type=float, default=5e-5)
    parser.add_argument("--epochs", type=int, default=6)
    parser.add_argument("--loss_weighting", type=str, default="log", choices=["lin", "log", "none"])
    parser.add_argument("--truncate", type=int, default=None)
    parser.add_argument("--batch_size", type=int, default=64)
    parser.add_argument("--num_workers", type=int, default=12)
    parser.add_argument("--use_augs", action="store_true", default=False)
    parser.add_argument("--train", action="store_true", default=False)
    args = parser.parse_args()

    TRAIN = args.train

    # fix all seeds
    pl.seed_everything(42, workers=True)

    # Create DataLoaders
    train_dataset = Chexpert_Dataset(split='train', truncate=args.truncate, loss_weighting=args.loss_weighting, use_augs=args.use_augs)
    val_dataset = Chexpert_Dataset(split='validate', truncate=args.truncate)
    test_dataset = Chexpert_Dataset(split='test')

    train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)
    val_dataloader = DataLoader(val_dataset, batch_size=args.batch_size, num_workers=args.num_workers)
    test_dataloader = DataLoader(test_dataset, batch_size=args.batch_size, num_workers=args.num_workers)

    # Number of classes for IGClassifier
    num_classes = len(train_dataset.chexpert_cols)
    class_names = train_dataset.chexpert_cols

    if TRAIN:
        class_weights = torch.tensor(train_dataset.get_class_weights(), dtype=torch.float32)
        # Define the model
        lit_model = LitIGClassifier(num_classes, class_weights, class_names, learning_rate=args.lr)
        print(summary(lit_model))

        # WandB logger
        wandb_run = wandb.init(
            project="ChexpertClassifier",
            entity= WANDB_ENTITY,
            name=args.run_name
        )

        # checkpoint callback
        checkpoint_callback = ModelCheckpoint(
            monitor='val_f1',
            dirpath=f'findings_classifier/checkpoints/{args.run_name}',
            filename='ChexpertClassifier-{epoch:02d}-{val_f1:.2f}',
            save_top_k=1,
            save_last=True,
            mode='max',
        )
        # Train the model
        trainer = pl.Trainer(max_epochs=args.epochs, gpus=1, callbacks=[checkpoint_callback], benchmark=False, deterministic=True, precision=16)
        trainer.fit(lit_model, train_dataloader, val_dataloader)

        # Test the model
        # trainer.validate(lit_model, val_dataloader, ckpt_path="checkpoints_IGCLassifier/lr_5e-5_to0_log_weighting_patches_augs_imgemb/IGClassifier-epoch=09-val_f1=0.65.ckpt")
    else:
        save_preds(train_dataloader, "train")
        save_preds(val_dataloader, "val")
        save_preds(test_dataloader, "test")