DALMLPII/main.py at master · Fluffel/DALMLPII · GitHub

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
import os
import torch
from torch.utils.data import Subset, random_split

from model import *
from strategies import *
from utils import *

import wandb


def load_mnist(batch_size = 64, normalize=True):
    from torchvision.datasets import FashionMNIST
    from torchvision.transforms import ToTensor, Normalize, Compose


    path_to_mean_std = "data/mnist_mean_std.pt"
    if normalize and os.path.exists(path_to_mean_std):
        loaded_tensors = torch.load(path_to_mean_std)
        train_mean = loaded_tensors['train_mean']
        train_std = loaded_tensors['train_std']
        test_mean = loaded_tensors['test_mean']
        test_std = loaded_tensors['test_std']

        transform = Compose([ToTensor(), Normalize(train_mean, train_std)])
        train_set = FashionMNIST('./data', train=True, download=True, transform=transform)

        transform = Compose([ToTensor(), Normalize(test_mean, test_std)])
        test_set = FashionMNIST('./data', train=False, download=True, transform=transform)

        return train_set, test_set

    transform = ToTensor()
    train_set = FashionMNIST('./data', train=True, download=True, transform=transform)
    test_set = FashionMNIST('./data', train=False, download=True, transform=transform)

    # Normalize
    if normalize:
        train_mean, train_std = get_mean_std(train_set)
        test_mean, test_std = get_mean_std(test_set)

        torch.save({'train_mean': train_mean,
                    'train_std': train_std,
                    'test_mean': test_mean,
                    'test_std': test_std}, path_to_mean_std)

        transform = Compose([ToTensor(), Normalize(train_mean, train_std)])
        train_set = FashionMNIST('./data', train=True, download=True, transform=transform)

        transform = Compose([ToTensor(), Normalize(test_mean, test_std)])
        test_set = FashionMNIST('./data', train=False, download=True, transform=transform)

    return train_set, test_set


def train_model(model_path, train_set, validation_set, iteration, device, lr=1e-3, epochs=200, early_stop=False):
    model = LeNet().to(device)
    model.load_state_dict(torch.load(model_path))
    model.train()

    train_loader = get_data_loader(train_set)
    validation_loader = get_data_loader(validation_set)
    # optimizer = get_SGD_optimizer(model, lr)
    optimizer = get_adam_optimizer(model, lr)
    loss_function = nn.CrossEntropyLoss()
    if early_stop:
        early_stopper = EarlyStop(15, 0.0001)

    epoch_count = 0
    # print("Eval...")
    # loss, acc = eval_model(model, validation_loader, loss_function, device)
    # wandb.log({"validation/accuracy": acc, validation"/loss": loss})
    for e in range(epochs):
        # print("Train...")
        train_step(model, train_loader, optimizer, loss_function, device)
        # print("Eval...")
        # wandb.log({"validation/accuracy": acc, "validation/loss": loss})
        if early_stop:
            loss, acc = eval_model(model, validation_loader, loss_function, device)
            if early_stopper.early_stop(loss):
                break
        epoch_count += 1
    return model


def main(initial_size, strat, sampling_batch_size=100, iterations=1, sub_iterations_dal=4, epochs=200, discriminator_epochs=550, lr=1e-3, train_size=100, seed=None, no_gpu=True):
    wandb.login()
    if strat == "random":
        config = {"initial_size": initial_size, "sampling_batch_size": sampling_batch_size, "iterations":iterations, "epochs": epochs, "learning_rate":lr, "seed": None}
        wandb.init(project="DiscriminativeActiveLearning", name="RandomStrategy", config=config)
    elif strat == "dal":
        config = {"initial_size": initial_size, "sampling_batch_size": sampling_batch_size, "iterations":iterations, "sub_iterations_dal":sub_iterations_dal, "epochs": epochs, "discriminator_epochs":discriminator_epochs, "learning_rate":lr, "seed": None}
        wandb.init(project="DiscriminativeActiveLearning", name="DiscriminativeActiveLearning", config=config)
    else:
        config = {"epochs": epochs, "learning_rate":lr, "seed": None}
        wandb.init(project="DiscriminativeActiveLearning", name="LeNetPlain", config=config)


    if seed:
        np.random.seed(seed)
        torch.manual_seed(seed)

    use_gpu = not no_gpu and torch.cuda.is_available()
    if use_gpu:
        device = torch.device("cuda")
    else:
        device = torch.device("cpu")

    print("Initialize model...")
    model_path = 'data/initial_model'
    model = LeNet().to(device)
    torch.save(model.state_dict(), model_path)

    # Load Dataset
    print("Load Dataset...")
    train_set, test_set = load_mnist()
    train_set_size = int(0.8 * len(train_set))
    test_loader = get_data_loader(test_set)
    train_set, validation_set = random_split(train_set, [train_set_size, len(train_set) - train_set_size])
    labeled = np.random.permutation(len(train_set))[:initial_size]

    # Initialize strategy
    if strat == 'random':
        strategy = RandomSampling(sampling_batch_size)
    elif strat == 'dal':
        strategy = DiscrimativeRepresentationSampling(sampling_batch_size, None, num_sub_batches=sub_iterations_dal, discriminator_epochs=discriminator_epochs, device=device)
    else:
        # If no strategy is selected, the model is just trained on a custom sized train set, reporting validation scores

        small_train_set, small_validation_set, _ = random_split(train_set, [train_size, int(0.5*train_size), int(len(train_set) - (1.5 * train_size))])
        current_model = train_model(model_path, small_train_set, small_validation_set, "Full Dataset", device, lr=lr, epochs=epochs, early_stop=True)
        loss, acc = eval_model(current_model, test_loader, nn.CrossEntropyLoss(), device)
        print("Test loss: {}".format(loss))
        print("Test acc: {}".format(acc))
        return


    print("-----------Start Iterations-----------")
    accuracies = []
    losses = []
    number_labeled = []

    print("Train LeNet...")
    current_model = train_model(model_path, Subset(train_set, labeled), validation_set, 0, device, lr=lr, epochs=epochs)
    loss, acc = eval_model(current_model, test_loader, nn.CrossEntropyLoss(), device)
    losses.append(loss)
    accuracies.append(acc)
    number_labeled.append(len(labeled))

    wandb.log({"test/acc": acc, "test/loss": loss})
    print("-----------Iteration "+ str(0) + "-----------")
    print("Labeled sample size: " + str(len(labeled)))
    print("Accuracy: " + str(acc))
    print("Loss: " + str(loss))
    for i in range(iterations):
        # configure strat
        strategy.update_model(current_model)

        # Update labeled data
        labeled = strategy.get_next_batch(train_set, labeled)
        labeled_train_set = Subset(train_set, labeled)
        number_labeled.append(len(labeled))

        # Train model
        print("Train LeNet...")
        current_model = train_model(model_path, labeled_train_set, validation_set, i+1, device, lr=lr, epochs=epochs)

        # Evaluate current iteration
        loss, acc = eval_model(current_model, test_loader, nn.CrossEntropyLoss(), device)
        losses.append(loss)
        accuracies.append(acc)
        wandb.log({"test/acc": acc, "test/loss": loss})
        print("-----------Iteration "+ str(i + 1) + "-----------")
        print("Labeled sample size: " + str(len(labeled)))
        print("Accuracy: " + str(acc))
        print("Loss: " + str(loss))

        strategy.next_iteration()


if __name__ == '__main__':
    main(initial_size=100, strat='dal', sampling_batch_size=100, iterations=20, sub_iterations_dal=4, epochs=200, discriminator_epochs=700, lr=1e-4, seed=None, no_gpu=True)