Yolov5 train.py main()函数解析

main()函数作用

main() 的作用是：把训练安排好，然后启动训练。

它主要做三件事：

  1. 准备文件和设置：检查数据、模型文件；如果要继续上次训练，就找回存档和旧设置。
  2. 选择设备：决定用 CPU 还是 GPU，并准备好多张 GPU 的协作。
  3. 启动任务：正常训练就调用 train()；开启超参数搜索，就尝试多套设置并比较效果。

代码解析

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
def main(opt, callbacks=None):
    如果调用 main() 时没有传入 callbacks，就创建一个默认对象。它负责在训练的不同时间点触发日志记录等操作。
    if callbacks is None:
        callbacks = Callbacks()
        
    只让主进程执行下面三行：-1 表示普通单进程运行，0 表示多 GPU 训练中的主进程。这样多 GPU 时就不会重复打印和检查。   
    依次是：打印本次训练参数、检查 Git 状态、检查所需的 Python 依赖。其中 vars(opt) 是把 opt 中的参数转成字典，方便打印。
    if RANK in {-1, 0}:
        print_args(vars(opt))	
        check_git_status()
        check_requirements(ROOT / "requirements.txt")
        
    # 在训练开始前，准备好训练设置和文件位置。
    # 接着上次训练：找到 last.pt，读回上次的设置，让后面的程序继续训练。
    # 其他情况：检查这次提供的数据和模型文件，并确定训练结果保存在哪里。
    接着上次训练、从本地last.pt恢复、不开启超参数搜索 则接着上次训练继续
    opt.resume这次要不要接着上次训练。 
    check_comet_resume(opt)这次是不是要从 Comet  上保存的训练记录继续？ not...表示没有通过 Comet 找回训练记录，需要从电脑里的文件恢复。
    opt.evolve要不要让程序自动尝试不同的训练设置，找出效果更好的设置。 not...表示这次没有开启超参数搜索，也就是按一套指定的设置训练。
    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
        找到上次保存的 last.pt。如果 --resume 后写了路径，就用指定文件；否则找最近的一次训练。
        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
        定位上次训练保存的参数文件。
        opt_yaml = last.parent.parent / "opt.yaml"  # train options yaml
        先记住这次传入的数据集地址。
        opt_data = opt.data  # original dataset
        优先从 opt.yaml 读取旧参数；没有这个文件，就从 last.pt 读取。
        if opt_yaml.is_file():
            with open(opt_yaml, errors="ignore") as f:
                d = yaml.safe_load(f)
        else:
            d = torch_load(last, map_location="cpu")["opt"]
        用上次训练的参数替换当前参数。    
        opt = argparse.Namespace(**d)  # replace
        指定从 last.pt 加载权重，并保持“继续训练”状态。
        opt.cfg, opt.weights, opt.resume = "", str(last), True  # reinstate
        如果这次的数据集地址是 URL，重新解析这个地址。
        if is_url(opt_data):
            opt.data = check_file(opt_data)  # re-resolve a URL dataset locally on resume
    准备新训练
    else:
        检查并定位数据集、模型和超参数文件。
        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = (
            check_file(opt.data),
            check_yaml(opt.cfg),
            check_yaml(opt.hyp),
            str(opt.weights),
            str(opt.project),
        )  # checks
        模型配置 --cfg 和初始权重 --weights，至少要提供一个。
        assert len(opt.cfg) or len(opt.weights), "either --cfg or --weights must be specified"
        如果要搜索超参数，把默认输出位置改为 runs/evolve，并关闭训练断点恢复。
        if opt.evolve:
            if opt.project == str(ROOT / "runs/train"):  # if default project name, rename to runs/evolve
                opt.project = str(ROOT / "runs/evolve")
            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
        如果实验名写的是 cfg，就改用模型配置文件的名称。
        if opt.name == "cfg":
            opt.name = Path(opt.cfg).stem  # use model.yaml as name
        确定本次结果保存目录。已有同名目录时，通常会生成新名字，例如 exp2。
        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))

        
    # DDP 选择用 CPU 还是 GPU 训练；如果用多张 GPU，就让它们准备好一起工作。DDP 就是一种让多张 GPU 一起训练的方式。
    device = select_device(opt.device, batch_size=opt.batch_size)
    if LOCAL_RANK != -1:
        msg = "is not compatible with YOLOv5 Multi-GPU DDP training"
        assert not opt.image_weights, f"--image-weights {msg}"
        assert not opt.evolve, f"--evolve {msg}"
        assert opt.batch_size != -1, f"AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size"
        assert opt.batch_size % WORLD_SIZE == 0, f"--batch-size {opt.batch_size} must be multiple of WORLD_SIZE"
        assert torch.cuda.device_count() > LOCAL_RANK, "insufficient CUDA devices for DDP command"
        torch.cuda.set_device(LOCAL_RANK)
        device = torch.device("cuda", LOCAL_RANK)
        dist.init_process_group(
            backend="nccl" if dist.is_nccl_available() else "gloo", timeout=timedelta(seconds=10800)
        )

    # Train 如果没有开启超参数搜索，就开始正常训练。
    if not opt.evolve:
        train(opt.hyp, opt, device, callbacks)

    # Evolve hyperparameters (optional) 开启超参数搜索后，让程序自动尝试多套训练设置，比较哪套效果好。
    else:
        # Hyperparameter evolution metadata (including this hyperparameter True-False, lower_limit, upper_limit)
        meta = {
            "lr0": (False, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
            "lrf": (False, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
            "momentum": (False, 0.6, 0.98),  # SGD momentum/Adam beta1
            "weight_decay": (False, 0.0, 0.001),  # optimizer weight decay
            "warmup_epochs": (False, 0.0, 5.0),  # warmup epochs (fractions ok)
            "warmup_momentum": (False, 0.0, 0.95),  # warmup initial momentum
            "warmup_bias_lr": (False, 0.0, 0.2),  # warmup initial bias lr
            "box": (False, 0.02, 0.2),  # box loss gain
            "cls": (False, 0.2, 4.0),  # cls loss gain
            "cls_pw": (False, 0.5, 2.0),  # cls BCELoss positive_weight
            "obj": (False, 0.2, 4.0),  # obj loss gain (scale with pixels)
            "obj_pw": (False, 0.5, 2.0),  # obj BCELoss positive_weight
            "iou_t": (False, 0.1, 0.7),  # IoU training threshold
            "anchor_t": (False, 2.0, 8.0),  # anchor-multiple threshold
            "anchors": (False, 2.0, 10.0),  # anchors per output grid (0 to ignore)
            "fl_gamma": (False, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
            "hsv_h": (True, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
            "hsv_s": (True, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
            "hsv_v": (True, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
            "degrees": (True, 0.0, 45.0),  # image rotation (+/- deg)
            "translate": (True, 0.0, 0.9),  # image translation (+/- fraction)
            "scale": (True, 0.0, 0.9),  # image scale (+/- gain)
            "shear": (True, 0.0, 10.0),  # image shear (+/- deg)
            "perspective": (True, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
            "flipud": (True, 0.0, 1.0),  # image flip up-down (probability)
            "fliplr": (True, 0.0, 1.0),  # image flip left-right (probability)
            "mosaic": (True, 0.0, 1.0),  # image mosaic (probability)
            "mixup": (True, 0.0, 1.0),  # image mixup (probability)
            "copy_paste": (True, 0.0, 1.0),  # segment copy-paste (probability)
        }

        # GA configs
        pop_size = 50
        mutation_rate_min = 0.01
        mutation_rate_max = 0.5
        crossover_rate_min = 0.5
        crossover_rate_max = 1
        min_elite_size = 2
        max_elite_size = 5
        tournament_size_min = 2
        tournament_size_max = 10

        with open(opt.hyp, errors="ignore") as f:
            hyp = yaml.safe_load(f)  # load hyps dict
            if "anchors" not in hyp:  # anchors commented in hyp.yaml
                hyp["anchors"] = 3
        if opt.noautoanchor:
            del hyp["anchors"], meta["anchors"]
        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
        evolve_yaml, evolve_csv = save_dir / "hyp_evolve.yaml", save_dir / "evolve.csv"
        if opt.bucket:
            # download evolve.csv if exists
            subprocess.run(
                [
                    "gsutil",
                    "cp",
                    f"gs://{opt.bucket}/evolve.csv",
                    str(evolve_csv),
                ],
                check=False,
            )

        # Delete the items in meta dictionary whose first value is False
        del_ = [item for item, value_ in meta.items() if value_[0] is False]
        hyp_GA = hyp.copy()  # Make a copy of hyp dictionary
        for item in del_:
            del meta[item]  # Remove the item from meta dictionary
            del hyp_GA[item]  # Remove the item from hyp_GA dictionary

        # Set lower_limit and upper_limit arrays to hold the search space boundaries
        lower_limit = np.array([meta[k][1] for k in hyp_GA])
        upper_limit = np.array([meta[k][2] for k in hyp_GA])

        # Create gene_ranges list to hold the range of values for each gene in the population
        gene_ranges = [(lower_limit[i], upper_limit[i]) for i in range(len(upper_limit))]

        # Initialize the population with initial_values or random values
        initial_values = []

        # If resuming evolution from a previous checkpoint
        if opt.resume_evolve is not None:
            assert os.path.isfile(ROOT / opt.resume_evolve), "evolve population path is wrong!"
            with open(ROOT / opt.resume_evolve, errors="ignore") as f:
                evolve_population = yaml.safe_load(f)
                for value in evolve_population.values():
                    value = np.array([value[k] for k in hyp_GA])
                    initial_values.append(list(value))

        # If not resuming from a previous checkpoint, generate initial values from .yaml files in opt.evolve_population
        else:
            yaml_files = [f for f in os.listdir(opt.evolve_population) if f.endswith(".yaml")]
            for file_name in yaml_files:
                with open(os.path.join(opt.evolve_population, file_name)) as yaml_file:
                    value = yaml.safe_load(yaml_file)
                    value = np.array([value[k] for k in hyp_GA])
                    initial_values.append(list(value))

        # Generate random values within the search space for the rest of the population
        if not initial_values:
            population = [generate_individual(gene_ranges, len(hyp_GA)) for _ in range(pop_size)]
        elif pop_size > 1:
            population = [generate_individual(gene_ranges, len(hyp_GA)) for _ in range(pop_size - len(initial_values))]
            for initial_value in initial_values:
                population = [initial_value, *population]

        # Run the genetic algorithm for a fixed number of generations
        list_keys = list(hyp_GA.keys())
        best_fitness, best_individual = float("-inf"), None
        for generation in range(opt.evolve):
            if generation >= 1:
                save_dict = {}
                for i in range(len(population)):
                    little_dict = {list_keys[j]: float(population[i][j]) for j in range(len(population[i]))}
                    save_dict[f"gen{generation!s}number{i!s}"] = little_dict

                with open(save_dir / "evolve_population.yaml", "w") as outfile:
                    yaml.dump(save_dict, outfile, default_flow_style=False)

            # Adaptive elite size
            elite_size = min_elite_size + int((max_elite_size - min_elite_size) * (generation / opt.evolve))
            # Evaluate the fitness of each individual in the population
            fitness_scores = []
            for individual in population:
                for key, value in zip(hyp_GA.keys(), individual):
                    hyp_GA[key] = value
                hyp.update(hyp_GA)
                results = train(hyp.copy(), opt, device, callbacks)
                callbacks = Callbacks()
                # Write mutation results
                keys = (
                    "metrics/precision",
                    "metrics/recall",
                    "metrics/mAP_0.5",
                    "metrics/mAP_0.5:0.95",
                    "val/box_loss",
                    "val/obj_loss",
                    "val/cls_loss",
                )
                print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
                fitness_scores.append(results[2])

            # Select the fittest individuals for reproduction using adaptive tournament selection
            selected_indices = []
            for _ in range(pop_size - elite_size):
                # Adaptive tournament size
                tournament_size = max(
                    max(2, tournament_size_min),
                    int(min(tournament_size_max, pop_size) - (generation / (opt.evolve / 10))),
                )
                # Perform tournament selection to choose the best individual
                tournament_indices = random.sample(range(pop_size), tournament_size)
                tournament_fitness = [fitness_scores[j] for j in tournament_indices]
                winner_index = tournament_indices[tournament_fitness.index(max(tournament_fitness))]
                selected_indices.append(winner_index)

            # Add the elite individuals to the selected indices
            elite_indices = [i for i in range(pop_size) if fitness_scores[i] in sorted(fitness_scores)[-elite_size:]]
            selected_indices.extend(elite_indices)
            # Create the next generation through crossover and mutation
            next_generation = []
            for _ in range(pop_size):
                parent1_index = selected_indices[random.randint(0, pop_size - 1)]
                parent2_index = selected_indices[random.randint(0, pop_size - 1)]
                # Adaptive crossover rate
                crossover_rate = max(
                    crossover_rate_min, min(crossover_rate_max, crossover_rate_max - (generation / opt.evolve))
                )
                if random.uniform(0, 1) < crossover_rate:
                    crossover_point = random.randint(1, len(hyp_GA) - 1)
                    child = population[parent1_index][:crossover_point] + population[parent2_index][crossover_point:]
                else:
                    child = population[parent1_index].copy()  # copy, else mutation edits the parent in place
                # Adaptive mutation rate
                mutation_rate = max(
                    mutation_rate_min, min(mutation_rate_max, mutation_rate_max - (generation / opt.evolve))
                )
                for j in range(len(hyp_GA)):
                    if random.uniform(0, 1) < mutation_rate:
                        child[j] += random.uniform(-0.1, 0.1)
                        child[j] = min(max(child[j], gene_ranges[j][0]), gene_ranges[j][1])
                next_generation.append(child)
            # Track the best evaluated individual across all generations
            if max(fitness_scores) > best_fitness:
                best_fitness = max(fitness_scores)
                best_individual = population[fitness_scores.index(best_fitness)]
            # Replace the old population with the new generation
            population = next_generation
        # Print the best solution found
        LOGGER.info(f"Best solution found: {best_individual}")
        # Plot results
        plot_evolve(evolve_csv)
        LOGGER.info(
            f"Hyperparameter evolution finished {opt.evolve} generations\n"
            f"Results saved to {colorstr('bold', save_dir)}\n"
            f"Usage example: $ python train.py --hyp {evolve_yaml}"
        )

源代码

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
def main(opt, callbacks=None):
    """Runs the main entry point for training or hyperparameter evolution with specified options and optional callbacks.

    Args:
        opt (argparse.Namespace): The command-line arguments parsed for YOLOv5 training and evolution.
        callbacks (ultralytics.utils.callbacks.Callbacks, optional): Callback functions for various training stages.
            Defaults to Callbacks().

    Returns:
        None

    Notes:
        For detailed usage, refer to:
        https://github.com/ultralytics/yolov5/tree/master/models
    """
    if callbacks is None:
        callbacks = Callbacks()
    if RANK in {-1, 0}:
        print_args(vars(opt))
        check_git_status()
        check_requirements(ROOT / "requirements.txt")

    # Resume (from specified or most recent last.pt)
    if opt.resume and not check_comet_resume(opt) and not opt.evolve:
        last = Path(check_file(opt.resume) if isinstance(opt.resume, str) else get_latest_run())
        opt_yaml = last.parent.parent / "opt.yaml"  # train options yaml
        opt_data = opt.data  # original dataset
        if opt_yaml.is_file():
            with open(opt_yaml, errors="ignore") as f:
                d = yaml.safe_load(f)
        else:
            d = torch_load(last, map_location="cpu")["opt"]
        opt = argparse.Namespace(**d)  # replace
        opt.cfg, opt.weights, opt.resume = "", str(last), True  # reinstate
        if is_url(opt_data):
            opt.data = check_file(opt_data)  # re-resolve a URL dataset locally on resume
    else:
        opt.data, opt.cfg, opt.hyp, opt.weights, opt.project = (
            check_file(opt.data),
            check_yaml(opt.cfg),
            check_yaml(opt.hyp),
            str(opt.weights),
            str(opt.project),
        )  # checks
        assert len(opt.cfg) or len(opt.weights), "either --cfg or --weights must be specified"
        if opt.evolve:
            if opt.project == str(ROOT / "runs/train"):  # if default project name, rename to runs/evolve
                opt.project = str(ROOT / "runs/evolve")
            opt.exist_ok, opt.resume = opt.resume, False  # pass resume to exist_ok and disable resume
        if opt.name == "cfg":
            opt.name = Path(opt.cfg).stem  # use model.yaml as name
        opt.save_dir = str(increment_path(Path(opt.project) / opt.name, exist_ok=opt.exist_ok))

    # DDP mode
    device = select_device(opt.device, batch_size=opt.batch_size)
    if LOCAL_RANK != -1:
        msg = "is not compatible with YOLOv5 Multi-GPU DDP training"
        assert not opt.image_weights, f"--image-weights {msg}"
        assert not opt.evolve, f"--evolve {msg}"
        assert opt.batch_size != -1, f"AutoBatch with --batch-size -1 {msg}, please pass a valid --batch-size"
        assert opt.batch_size % WORLD_SIZE == 0, f"--batch-size {opt.batch_size} must be multiple of WORLD_SIZE"
        assert torch.cuda.device_count() > LOCAL_RANK, "insufficient CUDA devices for DDP command"
        torch.cuda.set_device(LOCAL_RANK)
        device = torch.device("cuda", LOCAL_RANK)
        dist.init_process_group(
            backend="nccl" if dist.is_nccl_available() else "gloo", timeout=timedelta(seconds=10800)
        )

    # Train
    if not opt.evolve:
        train(opt.hyp, opt, device, callbacks)

    # Evolve hyperparameters (optional)
    else:
        # Hyperparameter evolution metadata (including this hyperparameter True-False, lower_limit, upper_limit)
        meta = {
            "lr0": (False, 1e-5, 1e-1),  # initial learning rate (SGD=1E-2, Adam=1E-3)
            "lrf": (False, 0.01, 1.0),  # final OneCycleLR learning rate (lr0 * lrf)
            "momentum": (False, 0.6, 0.98),  # SGD momentum/Adam beta1
            "weight_decay": (False, 0.0, 0.001),  # optimizer weight decay
            "warmup_epochs": (False, 0.0, 5.0),  # warmup epochs (fractions ok)
            "warmup_momentum": (False, 0.0, 0.95),  # warmup initial momentum
            "warmup_bias_lr": (False, 0.0, 0.2),  # warmup initial bias lr
            "box": (False, 0.02, 0.2),  # box loss gain
            "cls": (False, 0.2, 4.0),  # cls loss gain
            "cls_pw": (False, 0.5, 2.0),  # cls BCELoss positive_weight
            "obj": (False, 0.2, 4.0),  # obj loss gain (scale with pixels)
            "obj_pw": (False, 0.5, 2.0),  # obj BCELoss positive_weight
            "iou_t": (False, 0.1, 0.7),  # IoU training threshold
            "anchor_t": (False, 2.0, 8.0),  # anchor-multiple threshold
            "anchors": (False, 2.0, 10.0),  # anchors per output grid (0 to ignore)
            "fl_gamma": (False, 0.0, 2.0),  # focal loss gamma (efficientDet default gamma=1.5)
            "hsv_h": (True, 0.0, 0.1),  # image HSV-Hue augmentation (fraction)
            "hsv_s": (True, 0.0, 0.9),  # image HSV-Saturation augmentation (fraction)
            "hsv_v": (True, 0.0, 0.9),  # image HSV-Value augmentation (fraction)
            "degrees": (True, 0.0, 45.0),  # image rotation (+/- deg)
            "translate": (True, 0.0, 0.9),  # image translation (+/- fraction)
            "scale": (True, 0.0, 0.9),  # image scale (+/- gain)
            "shear": (True, 0.0, 10.0),  # image shear (+/- deg)
            "perspective": (True, 0.0, 0.001),  # image perspective (+/- fraction), range 0-0.001
            "flipud": (True, 0.0, 1.0),  # image flip up-down (probability)
            "fliplr": (True, 0.0, 1.0),  # image flip left-right (probability)
            "mosaic": (True, 0.0, 1.0),  # image mosaic (probability)
            "mixup": (True, 0.0, 1.0),  # image mixup (probability)
            "copy_paste": (True, 0.0, 1.0),  # segment copy-paste (probability)
        }

        # GA configs
        pop_size = 50
        mutation_rate_min = 0.01
        mutation_rate_max = 0.5
        crossover_rate_min = 0.5
        crossover_rate_max = 1
        min_elite_size = 2
        max_elite_size = 5
        tournament_size_min = 2
        tournament_size_max = 10

        with open(opt.hyp, errors="ignore") as f:
            hyp = yaml.safe_load(f)  # load hyps dict
            if "anchors" not in hyp:  # anchors commented in hyp.yaml
                hyp["anchors"] = 3
        if opt.noautoanchor:
            del hyp["anchors"], meta["anchors"]
        opt.noval, opt.nosave, save_dir = True, True, Path(opt.save_dir)  # only val/save final epoch
        # ei = [isinstance(x, (int, float)) for x in hyp.values()]  # evolvable indices
        evolve_yaml, evolve_csv = save_dir / "hyp_evolve.yaml", save_dir / "evolve.csv"
        if opt.bucket:
            # download evolve.csv if exists
            subprocess.run(
                [
                    "gsutil",
                    "cp",
                    f"gs://{opt.bucket}/evolve.csv",
                    str(evolve_csv),
                ],
                check=False,
            )

        # Delete the items in meta dictionary whose first value is False
        del_ = [item for item, value_ in meta.items() if value_[0] is False]
        hyp_GA = hyp.copy()  # Make a copy of hyp dictionary
        for item in del_:
            del meta[item]  # Remove the item from meta dictionary
            del hyp_GA[item]  # Remove the item from hyp_GA dictionary

        # Set lower_limit and upper_limit arrays to hold the search space boundaries
        lower_limit = np.array([meta[k][1] for k in hyp_GA])
        upper_limit = np.array([meta[k][2] for k in hyp_GA])

        # Create gene_ranges list to hold the range of values for each gene in the population
        gene_ranges = [(lower_limit[i], upper_limit[i]) for i in range(len(upper_limit))]

        # Initialize the population with initial_values or random values
        initial_values = []

        # If resuming evolution from a previous checkpoint
        if opt.resume_evolve is not None:
            assert os.path.isfile(ROOT / opt.resume_evolve), "evolve population path is wrong!"
            with open(ROOT / opt.resume_evolve, errors="ignore") as f:
                evolve_population = yaml.safe_load(f)
                for value in evolve_population.values():
                    value = np.array([value[k] for k in hyp_GA])
                    initial_values.append(list(value))

        # If not resuming from a previous checkpoint, generate initial values from .yaml files in opt.evolve_population
        else:
            yaml_files = [f for f in os.listdir(opt.evolve_population) if f.endswith(".yaml")]
            for file_name in yaml_files:
                with open(os.path.join(opt.evolve_population, file_name)) as yaml_file:
                    value = yaml.safe_load(yaml_file)
                    value = np.array([value[k] for k in hyp_GA])
                    initial_values.append(list(value))

        # Generate random values within the search space for the rest of the population
        if not initial_values:
            population = [generate_individual(gene_ranges, len(hyp_GA)) for _ in range(pop_size)]
        elif pop_size > 1:
            population = [generate_individual(gene_ranges, len(hyp_GA)) for _ in range(pop_size - len(initial_values))]
            for initial_value in initial_values:
                population = [initial_value, *population]

        # Run the genetic algorithm for a fixed number of generations
        list_keys = list(hyp_GA.keys())
        best_fitness, best_individual = float("-inf"), None
        for generation in range(opt.evolve):
            if generation >= 1:
                save_dict = {}
                for i in range(len(population)):
                    little_dict = {list_keys[j]: float(population[i][j]) for j in range(len(population[i]))}
                    save_dict[f"gen{generation!s}number{i!s}"] = little_dict

                with open(save_dir / "evolve_population.yaml", "w") as outfile:
                    yaml.dump(save_dict, outfile, default_flow_style=False)

            # Adaptive elite size
            elite_size = min_elite_size + int((max_elite_size - min_elite_size) * (generation / opt.evolve))
            # Evaluate the fitness of each individual in the population
            fitness_scores = []
            for individual in population:
                for key, value in zip(hyp_GA.keys(), individual):
                    hyp_GA[key] = value
                hyp.update(hyp_GA)
                results = train(hyp.copy(), opt, device, callbacks)
                callbacks = Callbacks()
                # Write mutation results
                keys = (
                    "metrics/precision",
                    "metrics/recall",
                    "metrics/mAP_0.5",
                    "metrics/mAP_0.5:0.95",
                    "val/box_loss",
                    "val/obj_loss",
                    "val/cls_loss",
                )
                print_mutation(keys, results, hyp.copy(), save_dir, opt.bucket)
                fitness_scores.append(results[2])

            # Select the fittest individuals for reproduction using adaptive tournament selection
            selected_indices = []
            for _ in range(pop_size - elite_size):
                # Adaptive tournament size
                tournament_size = max(
                    max(2, tournament_size_min),
                    int(min(tournament_size_max, pop_size) - (generation / (opt.evolve / 10))),
                )
                # Perform tournament selection to choose the best individual
                tournament_indices = random.sample(range(pop_size), tournament_size)
                tournament_fitness = [fitness_scores[j] for j in tournament_indices]
                winner_index = tournament_indices[tournament_fitness.index(max(tournament_fitness))]
                selected_indices.append(winner_index)

            # Add the elite individuals to the selected indices
            elite_indices = [i for i in range(pop_size) if fitness_scores[i] in sorted(fitness_scores)[-elite_size:]]
            selected_indices.extend(elite_indices)
            # Create the next generation through crossover and mutation
            next_generation = []
            for _ in range(pop_size):
                parent1_index = selected_indices[random.randint(0, pop_size - 1)]
                parent2_index = selected_indices[random.randint(0, pop_size - 1)]
                # Adaptive crossover rate
                crossover_rate = max(
                    crossover_rate_min, min(crossover_rate_max, crossover_rate_max - (generation / opt.evolve))
                )
                if random.uniform(0, 1) < crossover_rate:
                    crossover_point = random.randint(1, len(hyp_GA) - 1)
                    child = population[parent1_index][:crossover_point] + population[parent2_index][crossover_point:]
                else:
                    child = population[parent1_index].copy()  # copy, else mutation edits the parent in place
                # Adaptive mutation rate
                mutation_rate = max(
                    mutation_rate_min, min(mutation_rate_max, mutation_rate_max - (generation / opt.evolve))
                )
                for j in range(len(hyp_GA)):
                    if random.uniform(0, 1) < mutation_rate:
                        child[j] += random.uniform(-0.1, 0.1)
                        child[j] = min(max(child[j], gene_ranges[j][0]), gene_ranges[j][1])
                next_generation.append(child)
            # Track the best evaluated individual across all generations
            if max(fitness_scores) > best_fitness:
                best_fitness = max(fitness_scores)
                best_individual = population[fitness_scores.index(best_fitness)]
            # Replace the old population with the new generation
            population = next_generation
        # Print the best solution found
        LOGGER.info(f"Best solution found: {best_individual}")
        # Plot results
        plot_evolve(evolve_csv)
        LOGGER.info(
            f"Hyperparameter evolution finished {opt.evolve} generations\n"
            f"Results saved to {colorstr('bold', save_dir)}\n"
            f"Usage example: $ python train.py --hyp {evolve_yaml}"
        )