File size: 17,977 Bytes
1eced3c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349

import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import math
from datetime import datetime
from sklearn.utils import shuffle
from sklearn.model_selection import train_test_split
from numpy import save, load, asarray
import csv
from skimage.io import imread
import pickle
from sklearn.metrics import accuracy_score
import os
import time

from AffectNetClass import AffectNet
from RafdbClass import RafDB
from FerPlusClass import FerPlus

from config import DatasetName, AffectnetConf, InputDataSize, LearningConfig, DatasetType, RafDBConf, FerPlusConf
from cnn_model import CNNModel
from custom_loss import CustomLosses
from data_helper import DataHelper
from dataset_class import CustomDataset


class TrainModel:
    def __init__(self, dataset_name, ds_type, weights='imagenet', lr=1e-3, aug=True):
        self.dataset_name = dataset_name
        self.ds_type = ds_type
        self.weights = weights
        self.lr = lr

        self.base_lr = 1e-5
        self.max_lr = 5e-4
        if dataset_name == DatasetName.fer2013:
            self.drop = 0.1
            self.epochs_drop = 5
            if aug:
                self.img_path = FerPlusConf.aug_train_img_path
                self.annotation_path = FerPlusConf.aug_train_annotation_path
                self.masked_img_path = FerPlusConf.aug_train_masked_img_path
            else:
                self.img_path = FerPlusConf.no_aug_train_img_path
                self.annotation_path = FerPlusConf.no_aug_train_annotation_path

            self.val_img_path = FerPlusConf.test_img_path
            self.val_annotation_path = FerPlusConf.test_annotation_path
            self.eval_masked_img_path = FerPlusConf.test_masked_img_path
            self.num_of_classes = 7
            self.num_of_samples = None

        elif dataset_name == DatasetName.rafdb:
            self.drop = 0.1
            self.epochs_drop = 5

            if aug:
                self.img_path = RafDBConf.aug_train_img_path
                self.annotation_path = RafDBConf.aug_train_annotation_path
                self.masked_img_path = RafDBConf.aug_train_masked_img_path
            else:
                self.img_path = RafDBConf.no_aug_train_img_path
                self.annotation_path = RafDBConf.no_aug_train_annotation_path

            self.val_img_path = RafDBConf.test_img_path
            self.val_annotation_path = RafDBConf.test_annotation_path
            self.eval_masked_img_path = RafDBConf.test_masked_img_path
            self.num_of_classes = 7
            self.num_of_samples = None

        elif dataset_name == DatasetName.affectnet:
            self.drop = 0.1
            self.epochs_drop = 5

            if ds_type == DatasetType.train:
                self.img_path = AffectnetConf.aug_train_img_path
                self.annotation_path = AffectnetConf.aug_train_annotation_path
                self.masked_img_path = AffectnetConf.aug_train_masked_img_path
                self.val_img_path = AffectnetConf.eval_img_path
                self.val_annotation_path = AffectnetConf.eval_annotation_path
                self.eval_masked_img_path = AffectnetConf.eval_masked_img_path
                self.num_of_classes = 8
                self.num_of_samples = AffectnetConf.num_of_samples_train
            elif ds_type == DatasetType.train_7:
                if aug:
                    self.img_path = AffectnetConf.aug_train_img_path_7
                    self.annotation_path = AffectnetConf.aug_train_annotation_path_7
                    self.masked_img_path = AffectnetConf.aug_train_masked_img_path_7
                else:
                    self.img_path = AffectnetConf.no_aug_train_img_path_7
                    self.annotation_path = AffectnetConf.no_aug_train_annotation_path_7

                self.val_img_path = AffectnetConf.eval_img_path_7
                self.val_annotation_path = AffectnetConf.eval_annotation_path_7
                self.eval_masked_img_path = AffectnetConf.eval_masked_img_path_7
                self.num_of_classes = 7
                self.num_of_samples = AffectnetConf.num_of_samples_train_7

    def train(self, arch, weight_path):
        """"""

        '''create loss'''
        c_loss = CustomLosses()

        '''create summary writer'''
        summary_writer = tf.summary.create_file_writer(
            "./train_logs/fit/" + datetime.now().strftime("%Y%m%d-%H%M%S"))
        start_train_date = datetime.now().strftime("%Y%m%d-%H%M%S")

        '''making models'''
        model = self.make_model(arch=arch, w_path=weight_path)
        '''create save path'''
        if self.dataset_name == DatasetName.affectnet:
            save_path = AffectnetConf.weight_save_path + start_train_date + '/'
        elif self.dataset_name == DatasetName.rafdb:
            save_path = RafDBConf.weight_save_path + start_train_date + '/'
        elif self.dataset_name == DatasetName.fer2013:
            save_path = FerPlusConf.weight_save_path + start_train_date + '/'
        if not os.path.exists(save_path):
            os.makedirs(save_path)

        '''create sample generator'''
        dhp = DataHelper()

        '''     Train   Generator'''
        img_filenames, exp_filenames = dhp.create_generator_full_path(img_path=self.img_path,
                                                                      annotation_path=self.annotation_path)
        '''create dataset'''
        cds = CustomDataset()
        ds = cds.create_dataset(img_filenames=img_filenames,
                                anno_names=exp_filenames,
                                is_validation=False)

        '''create train configuration'''
        step_per_epoch = len(img_filenames) // LearningConfig.batch_size
        gradients = None
        virtual_step_per_epoch = LearningConfig.virtual_batch_size // LearningConfig.batch_size

        '''create optimizer'''
        optimizer = tf.keras.optimizers.Adam(self.lr, decay=1e-5)

        '''start train:'''
        all_gt_exp = []
        all_pr_exp = []

        for epoch in range(LearningConfig.epochs):
            ce_weight = 2
            batch_index = 0

            for img_batch, exp_batch in ds:
                '''since the calculation of the confusion matrix will be time-consuming,
                    we only save 1000 labels each time. Moreover, this help us to be more qiuck on updates
                    '''
                all_gt_exp, all_pr_exp = self._update_all_labels_arrays(all_gt_exp, all_pr_exp)
                '''load annotation and images'''
                '''squeeze'''
                exp_batch = exp_batch[:, -1]
                img_batch = img_batch[:, -1, :, :]

                '''train step'''
                step_gradients, all_gt_exp, all_pr_exp = self.train_step(epoch=epoch, step=batch_index,
                                                                         total_steps=step_per_epoch,
                                                                         img_batch=img_batch,
                                                                         anno_exp=exp_batch,
                                                                         model=model, optimizer=optimizer,
                                                                         c_loss=c_loss,
                                                                         ce_weight=ce_weight,
                                                                         summary_writer=summary_writer,
                                                                         all_gt_exp=all_gt_exp,
                                                                         all_pr_exp=all_pr_exp)
                batch_index += 1

            '''evaluating part'''
            global_accuracy, conf_mat, avg_acc = self._eval_model(model=model)
            '''save weights'''
            save_name = save_path + '_' + str(epoch) + '_' + self.dataset_name + '_AC_' + str(global_accuracy)
            model.save(save_name + '.h5')
            self._save_confusion_matrix(conf_mat, save_name + '.txt')

    def train_step(self, epoch, step, total_steps, model, ce_weight,
                   img_batch, anno_exp, optimizer, summary_writer, c_loss, all_gt_exp, all_pr_exp):
        with tf.GradientTape() as tape:
            pr_data = model([img_batch], training=True)
            exp_pr_vec = pr_data[0]
            embeddings = pr_data[1:]

            bs_size = tf.shape(exp_pr_vec, out_type=tf.dtypes.int64)[0]

            loss_exp, accuracy = c_loss.cross_entropy_loss(y_pr=exp_pr_vec, y_gt=anno_exp,
                                                           num_classes=self.num_of_classes,
                                                           ds_name=self.dataset_name)

            '''Feature difference loss'''
            # embedding_similarity_loss = 0
            embedding_similarity_loss = c_loss.embedding_loss_distance(embeddings=embeddings)

            '''update confusion matrix'''
            exp_pr = tf.constant([np.argmax(exp_pr_vec[i]) for i in range(bs_size)], dtype=tf.dtypes.int64)
            tr_conf_matrix, all_gt_exp, all_pr_exp = c_loss.update_confusion_matrix(anno_exp,  # real labels
                                                                                    exp_pr,  # real labels
                                                                                    all_gt_exp,
                                                                                    all_pr_exp)
            ''' correlation between the embeddings'''
            correlation_loss = c_loss.correlation_loss_multi(embeddings=embeddings,
                                                             exp_gt_vec=anno_exp,
                                                             exp_pr_vec=exp_pr_vec,
                                                             tr_conf_matrix=tr_conf_matrix)
            '''mean loss'''
            mean_correlation_loss = c_loss.mean_embedding_loss_distance(embeddings=embeddings,
                                                                        exp_gt_vec=anno_exp,
                                                                        exp_pr_vec=exp_pr_vec,
                                                                        num_of_classes=self.num_of_classes)

            lamda_param = 50
            loss_total = lamda_param * loss_exp + \
                         embedding_similarity_loss + \
                         correlation_loss + \
                         mean_correlation_loss

            # '''calculate gradient'''
        gradients_of_model = tape.gradient(loss_total, model.trainable_variables)
        # '''apply Gradients:'''
        optimizer.apply_gradients(zip(gradients_of_model, model.trainable_variables))
        # '''printing loss Values: '''
        tf.print("->EPOCH: ", str(epoch), "->STEP: ", str(step) + '/' + str(total_steps),
                 ' -> : accuracy: ', accuracy,
                 ' -> : loss_total: ', loss_total,
                 ' -> : loss_exp: ', loss_exp,
                 ' -> : embedding_similarity_loss: ', embedding_similarity_loss,
                 ' -> : correlation_loss: ', correlation_loss,
                 ' -> : mean_correlation_loss: ', mean_correlation_loss)
        with summary_writer.as_default():
            tf.summary.scalar('loss_total', loss_total, step=epoch)
            tf.summary.scalar('loss_exp', loss_exp, step=epoch)
            tf.summary.scalar('correlation_loss', correlation_loss, step=epoch)
            tf.summary.scalar('mean_correlation_loss', mean_correlation_loss, step=epoch)
            tf.summary.scalar('embedding_similarity_loss', embedding_similarity_loss, step=epoch)
        return gradients_of_model, all_gt_exp, all_pr_exp

    def train_step_old(self, epoch, step, total_steps, model, ce_weight,
                       img_batch, anno_exp, optimizer, summary_writer, c_loss, all_gt_exp, all_pr_exp):
        with tf.GradientTape() as tape:
            # '''create annotation_predicted'''
            # exp_pr, embedding = model([img_batch], training=True)
            exp_pr_vec, embedding_class, embedding_mean, embedding_var = model([img_batch], training=True)

            bs_size = tf.shape(exp_pr_vec, out_type=tf.dtypes.int64)[0]
            # # '''CE loss'''
            loss_exp, accuracy = c_loss.cross_entropy_loss(y_pr=exp_pr_vec, y_gt=anno_exp,
                                                           num_classes=self.num_of_classes,
                                                           ds_name=self.dataset_name)
            #
            loss_cls_mean, loss_cls_var, loss_mean_var = c_loss.embedding_loss_distance(
                embedding_class=embedding_class,
                embedding_mean=embedding_mean,
                embedding_var=embedding_var,
                bs_size=bs_size)
            feature_diff_loss = loss_cls_mean + loss_cls_var + loss_mean_var

            # correlation between the class_embeddings
            cor_loss, all_gt_exp, all_pr_exp = c_loss.correlation_loss(embedding=embedding_class,  # distribution
                                                                       exp_gt_vec=anno_exp,
                                                                       exp_pr_vec=exp_pr_vec,
                                                                       num_of_classes=self.num_of_classes,
                                                                       all_gt_exp=all_gt_exp,
                                                                       all_pr_exp=all_pr_exp)
            # correlation between the mean_emb_cor_loss
            mean_emb_cor_loss, mean_emb_kl_loss = c_loss.mean_embedding_loss(embedding=embedding_mean,
                                                                             exp_gt_vec=anno_exp,
                                                                             exp_pr_vec=exp_pr_vec,
                                                                             num_of_classes=self.num_of_classes)
            mean_loss = mean_emb_cor_loss + 10 * mean_emb_kl_loss

            var_emb_cor_loss, var_emb_kl_loss = c_loss.variance_embedding_loss(embedding=embedding_var,
                                                                               exp_gt_vec=anno_exp,
                                                                               exp_pr_vec=exp_pr_vec,
                                                                               num_of_classes=self.num_of_classes)
            var_loss = var_emb_cor_loss + 10 * var_emb_kl_loss
            # '''total:'''
            loss_total = 100 * loss_exp + cor_loss + 10 * feature_diff_loss + mean_loss + var_loss

        # '''calculate gradient'''
        gradients_of_model = tape.gradient(loss_total, model.trainable_variables)
        # '''apply Gradients:'''
        optimizer.apply_gradients(zip(gradients_of_model, model.trainable_variables))
        # '''printing loss Values: '''
        tf.print("->EPOCH: ", str(epoch), "->STEP: ", str(step) + '/' + str(total_steps),
                 ' -> : accuracy: ', accuracy,
                 ' -> : loss_total: ', loss_total,
                 ' -> : loss_exp: ', loss_exp,
                 ' -> : cor_loss: ', cor_loss,
                 ' -> : feature_loss: ', feature_diff_loss,
                 ' -> : mean_loss: ', mean_loss,
                 ' -> : var_loss: ', var_loss)

        with summary_writer.as_default():
            tf.summary.scalar('loss_total', loss_total, step=epoch)
            tf.summary.scalar('loss_exp', loss_exp, step=epoch)
            tf.summary.scalar('loss_correlation', cor_loss, step=epoch)
        return gradients_of_model, all_gt_exp, all_pr_exp

    def _eval_model(self, model):
        """"""
        '''first we need to create the 4 bunch here: '''

        '''for Affectnet, we need to calculate accuracy of each label and then total avg accuracy:'''
        global_accuracy = 0
        avg_acc = 0
        conf_mat = []
        if self.dataset_name == DatasetName.affectnet:
            if self.ds_type == DatasetType.train:
                affn = AffectNet(ds_type=DatasetType.eval)
            else:
                affn = AffectNet(ds_type=DatasetType.eval_7)
            global_accuracy, conf_mat, avg_acc, precision, recall, fscore, support = \
                affn.test_accuracy(model=model)
        elif self.dataset_name == DatasetName.rafdb:
            rafdb = RafDB(ds_type=DatasetType.test)
            global_accuracy, conf_mat, avg_acc, precision, recall, fscore, support = rafdb.test_accuracy(model=model)
        elif self.dataset_name == DatasetName.fer2013:
            ferplus = FerPlus(ds_type=DatasetType.test)
            global_accuracy, conf_mat, avg_acc, precision, recall, fscore, support = ferplus.test_accuracy(model=model)
        print("================== global_accuracy =====================")
        print(global_accuracy)
        print("================== Average Accuracy =====================")
        print(avg_acc)
        print("================== Confusion Matrix =====================")
        print(conf_mat)
        return global_accuracy, conf_mat, avg_acc

    def make_model(self, arch, w_path):
        cnn = CNNModel()
        model = cnn.get_model(arch=arch, num_of_classes=LearningConfig.num_classes, weights=self.weights)
        if w_path is not None:
            model.load_weights(w_path)
        return model

    def _save_confusion_matrix(self, conf_mat, save_name):
        f = open(save_name, "a")
        print(save_name)
        f.write(np.array_str(conf_mat))
        f.close()

    def _update_all_labels_arrays(self, all_gt_exp, all_pr_exp):
        if len(all_gt_exp) < LearningConfig.labels_history_frame:
            return all_gt_exp, all_pr_exp
        else:  # remove the first batch:
            return all_gt_exp[LearningConfig.batch_size:], all_pr_exp[LearningConfig.batch_size:]