task-amenability/pneumoniamnist_experiment_test_final_models.py at main · s-sd/task-amenability · 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

from interface import DDPGInterface

from keras import layers
import keras

import numpy as np
import os

import matplotlib.pyplot as plt


def build_task_predictor(input_shape):
    inputs = keras.Input(shape=input_shape)
    x = layers.Conv2D(32, (3,3), activation='relu')(inputs)
    x = layers.MaxPool2D((2,2))(x)
    x = layers.Conv2D(64, (3,3), activation='relu')(x)
    x = layers.MaxPool2D((2,2))(x)
    x = layers.Conv2D(64, (3,3), activation='relu')(x)
    x = layers.Flatten()(x)
    x = layers.Dense(64, activation='relu')(x)
    x = layers.Dropout(0.3)(x)
    x = layers.Dense(32, activation='relu')(x)
    x = layers.Dropout(0.2)(x)
    x = layers.Dense(16, activation='relu')(x)
    outputs = layers.Dense(1, activation='sigmoid')(x)
    return keras.Model(inputs=inputs, outputs=outputs)

img_shape = (28, 28, 1)

numpy_zip = np.load(r'pneumoniamnist_experiment/pneumoniamnist_corrupted.npz')

x_train, y_train = numpy_zip['x_train'], numpy_zip['y_train']
x_val, y_val = numpy_zip['x_val'], numpy_zip['y_val']
x_holdout, y_holdout = numpy_zip['x_holdout'], numpy_zip['y_holdout']

num_train_samples = len(y_train)
num_val_samples = len(y_val)
num_holdout_samples = len(y_holdout)


task_predictor = build_task_predictor(img_shape)
task_predictor.compile(optimizer='adam', loss='binary_crossentropy', metrics=['accuracy']) # speciffy the loss and metric used to train target net and controller respectively


def build_actor_critic(img_shape, action_shape=(1,)):

    n_actions = action_shape[0]

    act_in = layers.Input((1,) + img_shape)
    act_in_reshape = layers.Reshape((img_shape))(act_in)
    act_x = layers.Conv2D(32, (3,3), activation='relu')(act_in_reshape)
    act_x = layers.MaxPool2D((2,2))(act_x)
    act_x = layers.Conv2D(64, (3,3), activation='relu')(act_x)
    act_x = layers.MaxPool2D((2,2))(act_x)
    act_x = layers.Conv2D(64, (3,3), activation='relu')(act_x)
    act_x = layers.Flatten()(act_x)
    act_x = layers.Dense(64, activation='relu')(act_x)
    act_x = layers.Dense(32, activation='relu')(act_x)
    act_x = layers.Dense(16, activation='relu')(act_x)
    act_out = layers.Dense(n_actions, activation='sigmoid')(act_x)
    actor = keras.Model(inputs=act_in, outputs=act_out)

    action_input = layers.Input(shape=(n_actions,), name='action_input')
    observation_input = layers.Input((1,) + img_shape, name='observation_input')
    observation_input_reshape = layers.Reshape((img_shape))(observation_input)
    observation_x = layers.Conv2D(32, (3,3), activation='relu')(observation_input_reshape)
    observation_x = layers.MaxPool2D((2,2))(observation_x)
    observation_x = layers.Conv2D(64, (3,3), activation='relu')(observation_x)
    observation_x = layers.MaxPool2D((2,2))(observation_x)
    observation_x = layers.Conv2D(64, (3,3), activation='relu')(observation_x)
    flattened_observation = layers.Flatten()(observation_x)
    x = layers.Concatenate()([action_input, flattened_observation])
    x = layers.Dense(64, activation='relu')(x)
    x = layers.Dense(32, activation='relu')(x)
    x = layers.Dense(16, activation='relu')(x)
    x = layers.Dense(1)(x)
    critic = keras.Model(inputs=[action_input, observation_input], outputs=x)
    return actor, critic, action_input


controller_batch_size = 512
task_predictor_batch_size = 256

num_train_episodes = 512

save_path = r'pneumoniamnist_experiment/final_models/pneumoniamnist_experiment_train_session'

controller_weights_save_path = save_path + 'controller_episode_' + str(num_train_episodes)
task_predictor_save_path = save_path + 'task_predictor_episode_' + str(num_train_episodes)

actor, critic, action_input = build_actor_critic(img_shape)

interface = DDPGInterface(x_train, y_train, x_val, y_val, x_holdout, y_holdout, task_predictor, img_shape,
                          load_models=True, controller_weights_save_path=controller_weights_save_path, task_predictor_save_path=task_predictor_save_path,
                          custom_controller=True, actor=actor, critic=critic, action_input=action_input,
                          modify_env_params=True, modified_env_params_list=[controller_batch_size, task_predictor_batch_size])

holdout_controller_preds = interface.get_controller_preds_on_holdout()

def reject_lowest_controller_valued_samples(rejection_ratio, holdout_controller_preds, x_holdout, y_holdout):
    sorted_inds = np.argsort(holdout_controller_preds)
    num_rejected = int(rejection_ratio * len(sorted_inds))
    selected_x_holdout, selected_y_holdout = x_holdout[sorted_inds[num_rejected:], :, :, :], y_holdout[sorted_inds[num_rejected:]]
    return selected_x_holdout, selected_y_holdout

def compute_mean_performance(x, y, interface):
    mean_performance_metric = interface.task_predictor.evaluate(x, y)
    return mean_performance_metric[-1]

def compute_std(x, y, interface, samples, sample_size):
    means = []
    for _ in range(samples):
        inds = np.random.choice(range(len(x)), sample_size, replace=False)
        mean = compute_mean_performance(x[inds], y[inds], interface)
        means.append(mean)
    return np.std(means)

performances = []
variances = []
for rejection_ratio in np.arange(0.0, 0.5, 0.1):
    selected_x_holdout, selected_y_holdout = reject_lowest_controller_valued_samples(rejection_ratio, holdout_controller_preds, x_holdout, y_holdout)
    performance = compute_mean_performance(selected_x_holdout, selected_y_holdout, interface)
    performances.append(performance)
    variance = compute_std(selected_x_holdout, selected_y_holdout, interface, samples=100, sample_size=100)
    variances.append(variance)

plt.plot(performances)