train-procgen-pytorch/decoder_vis.py at master · tbuckworth/train-procgen-pytorch · 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
189
190
191
192
193
194
import os
import re

import numpy as np
from matplotlib import pyplot as plt

from helper_local import print_values_actions, add_encoder_to_env, last_folder, save_gif, latest_model_path
from inspect_agent import load_policy, predict
import torch
from common.model import Decoder
from train_decoder import retrieve_coinrun_data, send_reconstruction_update, get_decoder_details


def main(send_reconstructions=False):
    device = torch.device('cpu')
    encoder_path = "logs/train/coinrun/coinrun/2023-10-31__10-49-30__seed_6033/"

    decoder_3, impala_latents_3 = load_decoder_at(device, "block3")
    decoder_2, impala_latents_2 = load_decoder_at(device, "block2")
    decoder_fc, impala_latents_fc = load_decoder_at(device, "fc")

    decoding_info = {"decoder_3": decoder_3,
                     "decoder_2": decoder_2,
                     "decoder_fc": decoder_fc,
                     }
    # load encoder
    action_names, done, env, hidden_state, obs, policy, storage = load_policy(render=True, logdir=encoder_path, n_envs=2,
                                                                     decoding_info=decoding_info, start_level=431,
                                                                     repeat_level=False, num_levels=10)
    encoder = policy.embedder
    add_encoder_to_env(env, encoder)

    # if send_reconstructions:
    #     epoch = int(re.search(r"model_(\d*)\.pth", last_model).group(1))
    #     train_data, valid_data, _ = retrieve_coinrun_data()
    #     send_reconstruction_update(decoder, encoder, epoch, "data/plots", train_data, valid_data, device, impala_latents)

    frames = np.expand_dims(obs[0], 0)
    while True:
        act, log_prob_act, value, next_hidden_state, pi = predict(policy, obs, hidden_state, done)
        # print_values_actions(action_names, pi, value)
        next_obs, rew, done, info = env.step(act)
        frames = np.concatenate((frames[-100:], np.expand_dims(next_obs[0], 0)), 0)
        obs = next_obs
        hidden_state = next_hidden_state
        if done[0]:
            print(f"Level seed: {info[0]['level_seed']}")
    # obs = np.expand_dims(frames[-15],0)
    # np.save(os.path.join(encoder_path, "right_jump_obs.npy"), obs)


def value_saliency_gif():
    device = torch.device('cpu')
    encoder_path = "logs/train/coinrun/coinrun/2023-10-31__10-49-30__seed_6033/"
    action_names, done, env, hidden_state, obs, policy, storage = load_policy(render=True, logdir=encoder_path, n_envs=2,
                                                                     decoding_info={}, start_level=0,
                                                                     repeat_level=False)
    from agents.ppo import PPO as AGENT
    agent = AGENT(env, policy, None, None, device, 0)

    ep_len = 0
    frames = np.expand_dims(obs[0], 0)
    while True:
        obs = np.expand_dims(obs[0], 0)
        act, log_prob_act, value, next_hidden_state, vs_obs = agent.predict_w_value_saliency(obs,
                                                                                             hidden_state,
                                                                                             done)
        sds = (vs_obs - np.mean(vs_obs)) / np.std(vs_obs)
        flt = np.abs(sds) > 1
        flt = np.logical_or.reduce(flt, 1)
        flt = np.tile(flt, (1, 3, 1, 1))
        obs[np.bitwise_not(flt)] -= .2
        frames = np.concatenate((frames[-1000:], np.expand_dims(obs[0], 0)), 0)

        obs, rew, done, info = env.step(np.tile(act, 2))
        ep_len += 1
        hidden_state = next_hidden_state
        if done[0]:
            old_seed = info[0]['level_seed']
            print(f"Level seed: {info[0]['level_seed']}")
            save_gif(frames[-ep_len:], os.path.join(encoder_path, f"saliency_{old_seed}.gif"))
            ep_len = 0


def extreme_action_probs_inspection():
    device = torch.device('cpu')
    encoder_path = "logs/train/coinrun/coinrun/2023-10-31__10-49-30__seed_6033/"
    image_path = os.path.join(encoder_path, "right_jump_obs.npy")

    obs = np.load(image_path, allow_pickle='TRUE')
    action_names, done, env, hidden_state, _, policy, storage = load_policy(render=True, logdir=encoder_path, n_envs=2,
                                                                   decoding_info={}, start_level=0,
                                                                   repeat_level=False)
    from agents.ppo import PPO as AGENT
    agent = AGENT(env, policy, None, None, device, 0)

    plt.imshow(obs[0].transpose(1, 2, 0))
    plt.show()
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    n_obs = obs.copy()

    background = obs[0, :, 30, 20]

    n_obs[0, :, 0:28, 0:28] = np.tile(background, (28, 28, 1)).transpose(2, 0, 1)
    plt.imshow(n_obs[0].transpose(1, 2, 0))
    plt.show()
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, n_obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    n_obs[:, :, 0:28, :] = 0  # np.tile(background,(28,64,1)).transpose(2,0,1)
    plt.imshow(n_obs[0].transpose(1, 2, 0))
    plt.show()
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, n_obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    n_obs[:, :, 38:, :] = 0
    plt.imshow(n_obs[0].transpose(1, 2, 0))
    plt.show()
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, n_obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    act, log_prob_act, value, next_hidden_state, value_saliency_obs = agent.predict_w_value_saliency(obs, hidden_state,
                                                                                                     done)
    # n_obs[:, :, 38:, :] = 0

    min_v = np.min(value_saliency_obs)
    max_v = np.max(value_saliency_obs)
    vs_obs = (value_saliency_obs - min_v) / (max_v - min_v)

    plt.imshow(vs_obs[0].transpose(1, 2, 0))
    plt.show()
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, vs_obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    sds = (vs_obs - np.mean(vs_obs)) / np.std(vs_obs)
    flt = np.abs(sds) > 1
    flt = np.logical_or.reduce(flt, 1)
    flt = np.tile(flt, (1, 3, 1, 1))
    vs_obs[flt] = 1
    plt.imshow(vs_obs[0].transpose(1, 2, 0))
    plt.show()

    n_obs = obs.copy()
    n_obs[np.bitwise_not(flt)] -= 0.25
    plt.imshow(n_obs[0].transpose(1, 2, 0))
    plt.savefig(os.path.join(encoder_path, "right_jump_w_saliency.png"))
    act, log_prob_act, value, next_hidden_state, pi = predict(policy, n_obs, hidden_state, done)
    print_values_actions(action_names, pi, value)

    # l = impala_latents_2(policy.embedder, torch.Tensor(obs))
    # obs2 = decoder_2(l).detach().cpu().numpy()
    # plt.imshow(obs2[0].transpose(1, 2, 0))
    # plt.show()
    #
    # l = impala_latents_3(policy.embedder, torch.Tensor(obs))
    # obs3 = decoder_3(l).detach().cpu().numpy()
    # plt.imshow(obs3[0].transpose(1, 2, 0))
    # plt.show()


def load_decoder_at(device, latent_layer):
    decoder_path = last_folder(f"logs/decode/coinrun/decode_{latent_layer}", 1)
    decoder_params, impala_latents = get_decoder_details(latent_layer)
    decoder = Decoder(**decoder_params)
    last_model = latest_model_path(decoder_path)
    print(last_model)
    decoder.load_state_dict(torch.load(last_model, map_location=device)["model_state_dict"])
    return decoder, impala_latents


if __name__ == "__main__":
    main()

    # #This one shows crazy blue smudge behaviour - the blue smudge seems to correlate with enemies. (just MSE loss)
    # "logs/decode/coinrun/decode/2024-01-17__15-04-40__seed_6033"
    # # level 315 (or 352?) has intense blue code behaviour
    # # level 152 (or 87?) it gets stuck and blue code seems to be the cause.
    # also 49! 366! 64

    # logs/decode/coinrun/decode/2024-01-18__11-44-53__seed_6033
    # 404 - mistook background for enemy - colours indicate enemy?
    # 295 - a lot of colour when landing on first obstacle - indicate enemy/danger?
    # action was DOWN mostly, during colour, but then changed to right as colour went away
    #

    # Tri-decode: 498 - gets stuck and there's loads of intense colour in the image
    # 210: same as 498

    # Tri-decode fully trained:
    # 299 - gets stuck, but not for long
    # 59 - jumps over background image as if enemy
    # 97 - goes back left and then gets stuck