File size: 8,050 Bytes
158f4dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import shutil
import time
import numpy as np
from tqdm import tqdm
from transformers import ViTModel, ViTFeatureExtractor
from transformers.modeling_outputs import SequenceClassifierOutput
import torch.nn as nn
import torch
from PIL import Image
import logging
import os
from sklearn.preprocessing import LabelEncoder
from train import (
    re_training, metric, f1_score,
    classification_report
)

data_path = os.environ.get('DATA_PATH', "./data")

logging.basicConfig(level=os.getenv("LOGGER_LEVEL", logging.WARNING))
logger = logging.getLogger(__name__)

class ViTForImageClassification(nn.Module):
    def __init__(self, model_name, num_labels=24, dropout=0.25, image_size=224):
        logger.info("Loading model")
        super(ViTForImageClassification, self).__init__()
        self.vit = ViTModel.from_pretrained(model_name)
        self.feature_extractor = ViTFeatureExtractor.from_pretrained(model_name)
        self.feature_extractor.do_resize = True
        self.feature_extractor.size = image_size
        self.dropout = nn.Dropout(dropout)
        self.classifier = nn.Linear(self.vit.config.hidden_size, num_labels)
        self.num_labels = num_labels
        self.label_encoder = LabelEncoder()
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model_name = model_name
        # To device
        self.vit.to(self.device)
        self.to(self.device)
        self.classifier.to(self.device)
        logger.info("Model loaded")

    def forward(self, pixel_values, labels):
        logger.info("Forwarding")
        pixel_values = pixel_values.to(self.device)
        outputs = self.vit(pixel_values=pixel_values)
        output = self.dropout(outputs.last_hidden_state[:,0])
        logits = self.classifier(output)

        loss = None
        if labels is not None:
            loss_fct = nn.CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        return SequenceClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    def preprocess_image(self, images):
        logger.info("Preprocessing images")
        return self.feature_extractor(images, return_tensors='pt')

    def predict(self, images, batch_size=32, classes_names=True, return_probabilities=False):
        logger.info("Predicting")
        if not isinstance(images, list):
            images = [images]
        classes_list = []
        confidence_list = []
        for bs in tqdm(range(0, len(images), batch_size), desc="Preprocessing training images"):
            images_batch = [image for image in images[bs:bs+batch_size]]
            images_batch = self.preprocess_image(images_batch)['pixel_values']
            sequence_classifier_output = self.forward(images_batch, None)
            # Get max prob
            probs = sequence_classifier_output.logits.softmax(dim=-1).tolist()
            classes = np.argmax(probs, axis=1)
            confidences = np.max(probs, axis=1)
            classes_list.extend(classes)
            confidence_list.extend(confidences)
        if classes_names:
            classes_list = self.label_encoder.inverse_transform(classes_list)
        if return_probabilities:
            return classes_list, confidence_list, probs
        return classes_list, confidence_list

    def save(self, path):
        logger.info("Saving model")
        os.makedirs(path, exist_ok=True)
        torch.save(self.state_dict(), path + "/model.pt")
        # Save label encoder
        np.save(path + "/label_encoder.npy", self.label_encoder.classes_)

    def load(self, path):
        logger.info("Loading model")
        # Load label encoder
        # Check if label encoder and model exists
        if not os.path.exists(path + "/label_encoder.npy") or not os.path.exists(path + "/model.pt"):
            logger.warning("Label encoder or model not found")
            return
        self.label_encoder.classes_ = np.load(path + "/label_encoder.npy")
        # Reload classifier layer
        self.classifier = nn.Linear(self.vit.config.hidden_size, len(self.label_encoder.classes_))
        
        self.load_state_dict(torch.load(path + "/model.pt", map_location=self.device))
        self.vit.to(self.device)
        self.vit.eval()
        self.to(self.device)
        self.eval()
        
    def evaluate(self, images, labels):
        logger.info("Evaluating")
        labels = self.label_encoder.transform(labels)
        # Predict
        y_pred, _ = self.predict(images, classes_names=False)
        # Evaluate
        metrics = metric.compute(predictions=y_pred, references=labels)
        f1 = f1_score.compute(predictions=y_pred, references=labels, average="macro")
        print(classification_report(labels, y_pred, labels=[i for i in range(len(self.label_encoder.classes_))], target_names=self.label_encoder.classes_))
        print(f"Accuracy: {metrics['accuracy']}")
        print(f"F1: {f1}")
        
    def partial_fit(self, images, labels, save_model_path='new_model', num_epochs=10):
        logger.info("Partial fitting")
        # Freeze ViT model but last layer
        # params = [param for param in self.vit.parameters()]
        # for param in params[:-1]:
        #     param.requires_grad = False
        # Model in training mode
        self.vit.train()
        self.train()
        re_training(images, labels, self, save_model_path, num_epochs)
        self.load(save_model_path)
        self.vit.eval()
        self.eval()
        self.evaluate(images, labels)
        
    def __load_from_path(self, path, num_per_label=None):
        images = []
        labels = []
        for label in os.listdir(path):
            count = 0
            label_folder_path = os.path.join(path, label)
            for image_file in tqdm(os.listdir(label_folder_path), desc="Resizing images for label {}".format(label)):
                file_path = os.path.join(label_folder_path, image_file)
                try:
                    image = Image.open(file_path)
                    image_shape = (self.feature_extractor.size, self.feature_extractor.size)
                    if image.size != image_shape:
                        image = image.resize(image_shape)
                    images.append(image.convert('RGB'))
                    labels.append(label)
                    count += 1
                except Exception as e:
                    print(f"ERROR - Could not resize image {file_path} - {e}")
                if num_per_label is not None and count >= num_per_label:
                    break
        return images, labels
        
    def retrain_from_path(self,
                          path='./data/feedback',
                          num_per_label=None,
                          save_model_path='new_model',
                          remove_path=False,
                          num_epochs=10,
                          save_new_data=data_path + '/new_data'):
        logger.info("Retraining from path")
        # Load path
        images, labels = self.__load_from_path(path, num_per_label)
        # Retrain
        self.partial_fit(images, labels, save_model_path, num_epochs)
        # Save new data
        if save_new_data is not None:
            logger.info("Saving new data")
            for i ,(image, label) in enumerate(zip(images, labels)):
                label_path = os.path.join(save_new_data, label)
                os.makedirs(label_path, exist_ok=True)
                image.save(os.path.join(label_path, str(int(time.time())) + f"_{i}.jpg"))
        # Remove path folder
        if remove_path:
            logger.info("Removing feedback path")
            shutil.rmtree(path)
        
    def evaluate_from_path(self, path, num_per_label=None):
        logger.info("Evaluating from path")
        # Load images
        images, labels = self.__load_from_path(path, num_per_label)
        # Evaluate
        self.evaluate(images, labels)