VerificationCode/dev/click_captcha/train_yolo_char.py

# -*- coding: utf-8 -*-
"""
Created on Tue Jun 24 10:53:51 2022
train
@author: fangjiasheng
"""
import copy
import json
import os
import random
import sys
sys.path.append(os.path.dirname(os.path.abspath(__file__)) + "/../")
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
from click_captcha.utils import get_classes, get_anchors
from click_captcha.model import yolo_net

os.environ["CUDA_VISIBLE_DEVICES"] = "0"
# train need keras low version
# from tensorflow.keras.optimizers import Adam
# from keras.optimizer_v2.adam import Adam
from keras.optimizers import Adam
# from tensorflow.keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
from keras.callbacks import ModelCheckpoint, ReduceLROnPlateau
from glob import glob
from click_captcha.pre_process import gen_yolo_char
from keras.metrics import mse
# from keras.losses import categorical_crossentropy as CategoricalCrossentropy


PRETRAINED = False
CHECKPOINT = False
random.seed(42)


if __name__ == '__main__':
    annotation_path = '../data/detect/map.txt'
    weight_path = 'models/e20-loss116.38.h5'
    log_dir = 'yolo_data/logs/000/'
    classes_path = 'yolo_data/my_classes.txt'
    anchors_path = 'yolo_data/my_anchors.txt'
    class_names = get_classes(classes_path)
    num_classes = len(class_names)
    anchors = get_anchors(anchors_path)

    # multiple of 32, hw
    input_shape = (160, 256)

    # default setting
    is_tiny_version = len(anchors) == 6
    model = yolo_net(input_shape, anchors, num_classes,
                     load_pretrained=PRETRAINED,
                     weights_path=weight_path)

    val_split = 0.1
    with open(annotation_path) as f:
        lines = f.readlines()
    random.shuffle(lines)
    # lines = lines[:100]
    num_val = int(len(lines)*val_split)
    num_train = len(lines) - num_val
    print('total:', len(lines), 'train:', num_train, 'test:', num_val)

    file_path = 'models/e{epoch:02d}-loss{val_loss:.2f}.h5'
    checkpoint = ModelCheckpoint(file_path, monitor='val_loss',
                                 save_weights_only=True, save_best_only=True, period=2)
    reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=10, verbose=1)
    # model.compile(optimizer=Adam(lr=0.003), loss={'yolo_loss': lambda y_true, y_pred: y_pred},
    #               metrics=['acc', mse])

    # batch num
    batch_size = 32
    steps_per_epoch = max(1, num_train//batch_size)

    # data loader
    train_loader = gen_yolo_char(lines[:num_train], batch_size, input_shape, anchors, num_classes)
    test_loader = gen_yolo_char(lines[num_train:], batch_size, input_shape, anchors, num_classes)

    # Train with frozen layers first, to get a stable loss.
    # model.fit_generator(train_loader,
    #                     steps_per_epoch=steps_per_epoch,
    #                     callbacks=[reduce_lr],
    #                     validation_data=test_loader,
    #                     validation_steps=max(1, num_val // batch_size),
    #                     epochs=20,
    #                     max_queue_size=50)

    # Unfreeze and continue training, to fine-tune.
    for i in range(len(model.layers)):
        model.layers[i].trainable = True
    print('Unfreeze all of the layers.')

    model.compile(optimizer=Adam(lr=0.003), loss={'yolo_loss': lambda y_true, y_pred: y_pred},
                  metrics=[mse])
    model.fit_generator(train_loader,
                        steps_per_epoch=steps_per_epoch,
                        validation_data=test_loader,
                        validation_steps=max(1, num_val//batch_size),
                        epochs=500,
                        callbacks=[checkpoint, reduce_lr])