文章
文章 標籤 分類
夥伴
關於
關於作者 隱私權政策
更多
有趣 工具 實用網站 留言板 時間旅行

目錄

1140311 meeting

$SSSD^{S4}$ 程式碼探討與修改

 收錄於 Meeting
 發佈於   更新於    約 6945 字   預計閱讀 14 分鐘   次閱讀

以下為原文用於推論、填補的程式碼。

  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

def generate(output_directory,
             num_samples,
             ckpt_path,
             data_path,
             ckpt_iter,
             use_model,
             masking,
             missing_k,
             only_generate_missing):
    
    """
    Generate data based on ground truth 

    Parameters:
    output_directory (str):           save generated speeches to this path
    num_samples (int):                number of samples to generate, default is 4
    ckpt_path (str):                  checkpoint path
    ckpt_iter (int or 'max'):         the pretrained checkpoint to be loaded; 
                                      automitically selects the maximum iteration if 'max' is selected
    data_path (str):                  path to dataset, numpy array.
    use_model (int):                  0:DiffWave. 1:SSSDSA. 2:SSSDS4.
    masking (str):                    'mnr': missing not at random, 'bm': black-out, 'rm': random missing
    only_generate_missing (int):      0:all sample diffusion.  1:only apply diffusion to missing portions of the signal
    missing_k (int)                   k missing time points for each channel across the length.
    """

    # generate experiment (local) path
    local_path = "T{}_beta0{}_betaT{}".format(diffusion_config["T"],
                                              diffusion_config["beta_0"],
                                              diffusion_config["beta_T"])

    # Get shared output_directory ready
    output_directory = os.path.join(output_directory, local_path)
    if not os.path.isdir(output_directory):
        os.makedirs(output_directory)
        os.chmod(output_directory, 0o775)
    print("output directory", output_directory, flush=True)

    # map diffusion hyperparameters to gpu
    for key in diffusion_hyperparams:
        if key != "T":
            diffusion_hyperparams[key] = diffusion_hyperparams[key].cuda()

            
    # predefine model
    if use_model == 0:
        net = DiffWaveImputer(**model_config).cuda()
    elif use_model == 1:
        net = SSSDSAImputer(**model_config).cuda()
    elif use_model == 2:
        net = SSSDS4Imputer(**model_config).cuda()
    else:
        print('Model chosen not available.')
    print_size(net)

    
    # load checkpoint
    ckpt_path = os.path.join(ckpt_path, local_path)
    if ckpt_iter == 'max':
        ckpt_iter = find_max_epoch(ckpt_path)
    model_path = os.path.join(ckpt_path, '{}.pkl'.format(ckpt_iter))
    try:
        checkpoint = torch.load(model_path, map_location='cpu')
        net.load_state_dict(checkpoint['model_state_dict'])
        print('Successfully loaded model at iteration {}'.format(ckpt_iter))
    except:
        raise Exception('No valid model found')

        
        
    ### Custom data loading and reshaping ###
    
    testing_data = np.load(trainset_config['test_data_path'])
    testing_data = np.split(testing_data, 4, 0)
    testing_data = np.array(testing_data)
    testing_data = torch.from_numpy(testing_data).float().cuda()
    print('Data loaded')

    all_mse = []

    
    for i, batch in enumerate(testing_data):

        if masking == 'mnr':
            mask_T = get_mask_mnr(batch[0], missing_k)
            mask = mask_T.permute(1, 0)
            mask = mask.repeat(batch.size()[0], 1, 1)
            mask = mask.type(torch.float).cuda()

        elif masking == 'bm':
            mask_T = get_mask_bm(batch[0], missing_k)
            mask = mask_T.permute(1, 0)
            mask = mask.repeat(batch.size()[0], 1, 1)
            mask = mask.type(torch.float).cuda()

        elif masking == 'rm':
            mask_T = get_mask_rm(batch[0], missing_k)
            mask = mask_T.permute(1, 0)
            mask = mask.repeat(batch.size()[0], 1, 1).float().cuda()

            
            
        batch = batch.permute(0,2,1)
        
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        start.record()

        sample_length = batch.size(2)
        sample_channels = batch.size(1)
        generated_audio = sampling(net, (num_samples, sample_channels, sample_length),
                                   diffusion_hyperparams,
                                   cond=batch,
                                   mask=mask,
                                   only_generate_missing=only_generate_missing)

        end.record()
        torch.cuda.synchronize()

        print('generated {} utterances of random_digit at iteration {} in {} seconds'.format(num_samples,
                                                                                             ckpt_iter,
                                                                                             int(start.elapsed_time(
                                                                                                 end) / 1000)))

        
        generated_audio = generated_audio.detach().cpu().numpy()
        batch = batch.detach().cpu().numpy()
        mask = mask.detach().cpu().numpy() 
        
        
        outfile = f'imputation{i}.npy'
        new_out = os.path.join(ckpt_path, outfile)
        np.save(new_out, generated_audio)

        outfile = f'original{i}.npy'
        new_out = os.path.join(ckpt_path, outfile)
        np.save(new_out, batch)

        outfile = f'mask{i}.npy'
        new_out = os.path.join(ckpt_path, outfile)
        np.save(new_out, mask)

        print('saved generated samples at iteration %s' % ckpt_iter)
        
        mse = mean_squared_error(generated_audio[~mask.astype(bool)], batch[~mask.astype(bool)])
        all_mse.append(mse)
    
    print('Total MSE:', mean(all_mse))

從以上程式碼，由註解中我們可以發現此程式碼是將完整的測試集資料輸入，然後再藉由 config.json 決定是何種缺失 (mnr, bm, rm) ，再使用 get_mask_mnr 、 get_mask_bm 或 get_mask_rm 依照 missing_k 設定的數量模擬缺失部分生成遮罩 (mask) 。

由於在原文所提供的 Github 儲存庫中未找到相關預測的 Python 腳本，此處修改 inference.py 腳本，使程式碼可依照輸入的資料集自動找出缺失的遮罩，並加以填補。

參考了 https://github.com/AI4HealthUOL/SSSD/issues/15

  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

import os
import argparse
import json
import numpy as np
import torch

# from utils.util import get_mask_mnr, get_mask_bm, get_mask_rm
from utils.util import find_max_epoch, print_size, sampling, calc_diffusion_hyperparams

from imputers.DiffWaveImputer import DiffWaveImputer
from imputers.SSSDSAImputer import SSSDSAImputer
from imputers.SSSDS4Imputer import SSSDS4Imputer

import random

def get_mask_rm(sample):
    """Get mask of random points (missing at random) across channels based on k,
    where k == number of data points. Mask of sample's shape where 0's to be imputed, and 1's to preserved
    as per ts imputers"""

    # mask = torch.ones(sample.shape)  # 建立全為 1 的遮罩
    # length_index = torch.tensor(range(mask.shape[0]))  # 建立時間索引 T
    # for channel in range(mask.shape[1]):  # 逐一處理每個通道
    #     perm = torch.randperm(len(length_index))  # 產生隨機排列索引
    #     idx = perm[0:k]  # 取前 k 個索引作為缺失點
    #     mask[:, channel][idx] = 0  # 設定缺失值

    # return mask

    mask = ~torch.isnan(sample) # 不是 NaN 的位置設為 True，NaN 的位置為 False
    return mask.float() # 轉換為 0/1 格式 (0: 缺失, 1: 保留)


def get_mask_mnr(sample):
    """Get mask of random segments (non-missing at random) across channels based on k,
    where k == number of segments. Mask of sample's shape where 0's to be imputed, and 1's to preserved
    as per ts imputers"""

    # mask = torch.ones(sample.shape)
    # length_index = torch.tensor(range(mask.shape[0]))
    # list_of_segments_index = torch.split(length_index, k)
    # for channel in range(mask.shape[1]):
    #     s_nan = random.choice(list_of_segments_index)
    #     mask[:, channel][s_nan[0]:s_nan[-1] + 1] = 0

    # return mask

    mask = ~torch.isnan(sample) # 不是 NaN 的位置設為 True，NaN 的位置為 False
    return mask.float() # 轉換為 0/1 格式 (0: 缺失, 1: 保留)


def get_mask_bm(sample):
    """Get mask of same segments (black-out missing) across channels based on k,
    where k == number of segments. Mask of sample's shape where 0's to be imputed, and 1's to be preserved
    as per ts imputers"""

    # mask = torch.ones(sample.shape)
    # length_index = torch.tensor(range(mask.shape[0]))
    # list_of_segments_index = torch.split(length_index, k)
    # s_nan = random.choice(list_of_segments_index)
    # for channel in range(mask.shape[1]):
    #     mask[:, channel][s_nan[0]:s_nan[-1] + 1] = 0

    # return mask

    mask = ~torch.isnan(sample) # 不是 NaN 的位置設為 True，NaN 的位置為 False
    return mask.float() # 轉換為 0/1 格式 (0: 缺失, 1: 保留)

def predict(output_directory,
             num_samples,
             ckpt_path,
             data_path,
             ckpt_iter,
             config_path):
    
    """
    predict data based on ground truth 

    Parameters:
    output_directory (str):           save generated speeches to this path
    num_samples (int):                number of samples to generate, default is 4
    ckpt_path (str):                  checkpoint path
    ckpt_iter (int or 'max'):         the pretrained checkpoint to be loaded; 
                                      automitically selects the maximum iteration if 'max' is selected
    data_path (str):                  path to dataset, numpy array.
    use_model (int):                  0:DiffWave. 1:SSSDSA. 2:SSSDS4.
    masking (str):                    'mnr': missing not at random, 'bm': black-out, 'rm': random missing
    config_path (str):                str, it's the path to the predict config
    only_generate_missing (int):      0:all sample diffusion.  1:only apply diffusion to missing portions of the signal
    """

    # load config
    with open(config_path) as f:
        data = f.read()
    config = json.loads(data)
    print(f'load config.')
    print(config)
    
    # init parameters
    train_config = config["train_config"]
    use_model = train_config["use_model"]
    masking = train_config["masking"]
    only_generate_missing = train_config["only_generate_missing"]
    
    if use_model == 0:
        model_config = config['wavenet_config']
    elif use_model == 1:
        model_config = config['sashimi_config']
    elif use_model == 2:
        model_config = config['wavenet_config']

    

    # generate experiment (local) path
    diffusion_config = config["diffusion_config"] 
    local_path = "T{}_beta0{}_betaT{}".format(diffusion_config["T"],
                                              diffusion_config["beta_0"],
                                              diffusion_config["beta_T"])

    # Get shared output_directory ready
    output_directory = os.path.join(output_directory, local_path)
    if not os.path.isdir(output_directory):
        os.makedirs(output_directory)
        if os.name != "nt":  # 只在非 Windows 系統執行 chmod
            os.chmod(output_directory, 0o775)
    print("output directory", output_directory, flush=True)

    # calculate diffusion hyperparams
    diffusion_hyperparams = calc_diffusion_hyperparams(
    **diffusion_config)  # dictionary of all diffusion hyperparameters

    # map diffusion hyperparameters to gpu
    for key in diffusion_hyperparams:
        if key != "T":
            diffusion_hyperparams[key] = diffusion_hyperparams[key].cuda()

            
    # predefine model
    if use_model == 0:
        net = DiffWaveImputer(**model_config).cuda()
    elif use_model == 1:
        net = SSSDSAImputer(**model_config).cuda()
    elif use_model == 2:
        net = SSSDS4Imputer(**model_config).cuda()
    else:
        print('Model chosen not available.')
    print_size(net)

    
    # load checkpoint
    ckpt_path = os.path.join(ckpt_path, local_path)
    if ckpt_iter == 'max':
        ckpt_iter = find_max_epoch(ckpt_path)
    model_path = os.path.join(ckpt_path, '{}.pkl'.format(ckpt_iter))
    try:
        checkpoint = torch.load(model_path, map_location='cpu')
        net.load_state_dict(checkpoint['model_state_dict'])
        print('Successfully loaded model at iteration {}'.format(ckpt_iter))
    except:
        raise Exception('No valid model found')

        
        
    ### Custom data loading and reshaping ###
    # 這裡改過
    testing_data = np.load(data_path)
    testing_data = np.array(np.array_split(testing_data, 4, axis=0))
    testing_data = torch.from_numpy(testing_data).float().cuda()
    print('Data loaded')

    # all_mse = []

    
    for i, batch in enumerate(testing_data):

        # debug
        print(f'i = {i}, batch = {batch}')

        if masking == 'mnr':
            mask_T = get_mask_mnr(batch[0])
            mask = mask_T.permute(1, 0) # 交換維度
            mask = mask.repeat(batch.size()[0], 1, 1)
            mask = mask.type(torch.float).cuda()

        elif masking == 'bm':
            mask_T = get_mask_bm(batch[0])
            mask = mask_T.permute(1, 0)
            mask = mask.repeat(batch.size()[0], 1, 1)
            mask = mask.type(torch.float).cuda()

        elif masking == 'rm':
            mask_T = get_mask_rm(batch[0])
            mask = mask_T.permute(1, 0)
            mask = mask.repeat(batch.size()[0], 1, 1).float().cuda()

        # 將 batch 是 nan 的地方替換成 0
        batch = torch.nan_to_num(batch, nan=1.0)


        # debug
        print(f'mask')
        print(mask)
            
        batch = batch.permute(0,2,1)
        
        start = torch.cuda.Event(enable_timing=True)
        end = torch.cuda.Event(enable_timing=True)
        start.record()

        sample_length = batch.size(2)
        sample_channels = batch.size(1)
        generated_audio = sampling(net, (num_samples, sample_channels, sample_length),
                                   diffusion_hyperparams,
                                   cond=batch,
                                   mask=mask,
                                   only_generate_missing=only_generate_missing)
        # debug        
        print('sampling')
        print(generated_audio)

        end.record()
        torch.cuda.synchronize() # 確保 GPU 上的操作完成

        print('generated {} utterances of random_digit at iteration {} in {} seconds'.format(num_samples,
                                                                                             ckpt_iter,
                                                                                             int(start.elapsed_time(
                                                                                                 end) / 1000)))

        
        generated_audio = generated_audio.detach().cpu().numpy()
        batch = batch.detach().cpu().numpy()
        mask = mask.detach().cpu().numpy() 
        
        
        outfile = f'imputation{i}.npy'
        new_out = os.path.join(output_directory, outfile)
        np.save(new_out, generated_audio)

        outfile = f'original{i}.npy'
        new_out = os.path.join(output_directory, outfile)
        np.save(new_out, batch)

        outfile = f'mask{i}.npy'
        new_out = os.path.join(output_directory, outfile)
        np.save(new_out, mask)

        print('saved generated samples at iteration %s' % ckpt_iter)
        
        # mse = mean_squared_error(generated_audio[~mask.astype(bool)], batch[~mask.astype(bool)])
        # all_mse.append(mse)
    
    # print('Total MSE:', mean(all_mse))


if __name__ == "__main__":    

# predict
    predict(output_directory=r'D:\Code\sssd_cp_learning_and_testing\learning_and_testing\SSSD\results\mujoco\90_result',
             ckpt_path=r'D:\Code\sssd_cp_learning_and_testing\learning_and_testing\SSSD\results\mujoco\90',
             ckpt_iter='max',
             num_samples=500,
             data_path=r'D:\Code\sssd_cp_learning_and_testing\learning_and_testing\SSSD\results\mujoco\test_rm\test_rm.npy',
             config_path=r'D:\Code\sssd_cp_learning_and_testing\learning_and_testing\SSSD\config\predict_config_SSSDS4.json'
             )

此處修改了 get_mask_XX() 函數，使輸入不再選擇隨機挑選要填充的部分，而是依實際資料缺失 (nan) 選擇填補的位置。其中，mask 為 0 時表示需要填補，為 1 時表示已有資料。

目前遇到的問題是，當輸入值為 nan 時，所有的預測結果都會變成 nan 。造成此結果的兇手如下。

 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

def sampling(net, size, diffusion_hyperparams, cond, mask, only_generate_missing=0, guidance_weight=0):
    """
    Perform the complete sampling step according to p(x_0|x_T) = \prod_{t=1}^T p_{\theta}(x_{t-1}|x_t)

    Parameters:
    net (torch network):            the wavenet model
    size (tuple):                   size of tensor to be generated, 
                                    usually is (number of audios to generate, channels=1, length of audio)
    diffusion_hyperparams (dict):   dictionary of diffusion hyperparameters returned by calc_diffusion_hyperparams
                                    note, the tensors need to be cuda tensors 
    
    Returns:
    the generated audio(s) in torch.tensor, shape=size
    """

    _dh = diffusion_hyperparams
    T, Alpha, Alpha_bar, Sigma = _dh["T"], _dh["Alpha"], _dh["Alpha_bar"], _dh["Sigma"]
    assert len(Alpha) == T
    assert len(Alpha_bar) == T
    assert len(Sigma) == T
    assert len(size) == 3

    print('begin sampling, total number of reverse steps = %s' % T)

    x = std_normal(size)

    with torch.no_grad():
        for t in range(T - 1, -1, -1):
            if only_generate_missing == 1:
                x = x * (1 - mask).float() + cond * mask.float()  # np.nan * 0 >>> nan
            diffusion_steps = (t * torch.ones((size[0], 1))).cuda()  # use the corresponding reverse step
            epsilon_theta = net((x, cond, mask, diffusion_steps,))  # predict \epsilon according to \epsilon_\theta
            # update x_{t-1} to \mu_\theta(x_t)
            x = (x - (1 - Alpha[t]) / torch.sqrt(1 - Alpha_bar[t]) * epsilon_theta) / torch.sqrt(Alpha[t])
            if t > 0:
                x = x + Sigma[t] * std_normal(size)  # add the variance term to x_{t-1}

    return x

因為我們會將每 batch 的資料輸入 cond ，而在經過 x = x * (1 - mask).float() + cond * mask.float() 時，因 nan 乘以任意數會回傳 nan ，進而導致預測值全為 nan ，如下所示。

 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

sampling
tensor([[[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]],

        [[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]],

        [[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]],

        ...,

        [[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]],

        [[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]],

        [[nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         ...,
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan],
         [nan, nan, nan,  ..., nan, nan, nan]]], device='cuda:0')

這裡在預測的腳本中先採用了 batch = torch.nan_to_num(batch, nan=1.0) 進行補救。雖有值輸出但仍不理想。

  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
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627

PS D:\Code\sssd_cp_learning_and_testing> & "D:/Program Files/Python312/python.exe" d:/Code/sssd_cp_learning_and_testing/learning_and_testing/SSSD/predict.py
CUDA extension for cauchy multiplication not found. Install by going to extensions/cauchy/ and running `python setup.py install`. This should speed up end-to-end training by 10-50%
Falling back on slow Cauchy kernel. Install at least one of pykeops or the CUDA extension for efficiency.
load config.
{'diffusion_config': {'T': 200, 'beta_0': 0.0001, 'beta_T': 0.02}, 'wavenet_config': {'in_channels': 14, 'out_channels': 14, 'num_res_layers': 36, 'res_channels': 256, 'skip_channels': 256, 'diffusion_step_embed_dim_in': 128, 'diffusion_step_embed_dim_mid': 512, 'diffusion_step_embed_dim_out': 512, 's4_lmax': 100, 's4_d_state': 64, 's4_dropout': 0.0, 's4_bidirectional': 1, 's4_layernorm': 1}, 'train_config': {'only_generate_missing': 1, 'use_model': 2, 'masking': 'rm'}}
output directory D:\Code\sssd_cp_learning_and_testing\learning_and_testing\SSSD\results\mujoco\90_result\T200_beta00.0001_betaT0.02
C:\Users\a1021\AppData\Roaming\Python\Python312\site-packages\torch\nn\utils\weight_norm.py:143: FutureWarning: `torch.nn.utils.weight_norm` is deprecated in favor of `torch.nn.utils.parametrizations.weight_norm`.
  WeightNorm.apply(module, name, dim)
SSSDS4Imputer Parameters: 48.371726M
Successfully loaded model at iteration 13100
Data loaded
i = 0, batch = tensor([[[   nan, 0.4450, 1.1080,  ..., 2.0913, 0.7361, 1.4328],
         [1.1419, 0.4405,    nan,  ..., 2.0890, 0.7470, 1.4328],
         [1.1435, 0.4359, 1.1095,  ..., 2.0888, 0.7525, 1.4329],
         ...,
         [   nan, 0.0990, 1.2304,  ...,    nan, 0.8020, 1.4459],
         [1.2356, 0.0989, 1.2330,  ..., 1.9985, 0.8004, 1.4463],
         [1.2367, 0.0988, 1.2356,  ..., 1.9977, 0.7988, 1.4466]],

        [[0.7846, 0.1823, 1.2447,  ..., 2.0702,    nan, 1.4399],
         [0.7815, 0.1759, 1.2457,  ..., 2.0702, 0.6619, 1.4401],
         [0.7784, 0.1694, 1.2466,  ..., 2.0703, 0.6615, 1.4402],
         ...,
         [0.7394,    nan, 1.2939,  ..., 1.9836, 0.7462, 1.4307],
         [0.7399, 0.0412, 1.2957,  ..., 1.9815, 0.7458, 1.4308],
         [0.7404, 0.0397, 1.2976,  ..., 1.9794, 0.7455, 1.4310]],

        [[1.1802, 0.4251, 1.0593,  ..., 2.0391, 0.7275, 1.4547],
         [1.1829, 0.4211, 1.0606,  ...,    nan, 0.7268, 1.4548],
         [1.1856, 0.4170, 1.0618,  ..., 2.0419, 0.7262, 1.4548],
         ...,
         [1.3531, 0.0802, 1.1660,  ..., 2.0199, 0.7511, 1.4173],
         [1.3533, 0.0801, 1.1661,  ..., 2.0142, 0.7506, 1.4175],
         [1.3535, 0.0802, 1.1663,  ..., 2.0114, 0.7504, 1.4176]],

        ...,

        [[1.0410, 0.3429, 0.8728,  ..., 2.2161, 1.0149, 1.3141],
         [1.0441, 0.3391, 0.8711,  ...,    nan, 1.0114, 1.3149],
         [1.0472, 0.3351, 0.8694,  ..., 2.2318, 1.0082, 1.3156],
         ...,
         [1.2279, 0.0237, 0.8619,  ..., 1.7460, 0.7155, 1.4611],
         [1.2282, 0.0242, 0.8624,  ..., 1.7672, 0.7156, 1.4616],
         [1.2284, 0.0250, 0.8628,  ..., 1.7801, 0.7158,    nan]],

        [[0.8293, 0.0369, 0.7955,  ..., 2.2419,    nan, 1.3916],
         [0.8288, 0.0398, 0.7929,  ..., 2.2403,    nan,    nan],
         [0.8283, 0.0427,    nan,  ...,    nan, 0.7819, 1.3927],
         ...,
         [0.8174, 0.0516, 0.7810,  ..., 1.9140, 0.7994, 1.4366],
         [0.8179, 0.0504, 0.7825,  ..., 1.9133, 0.7998, 1.4376],
         [0.8184, 0.0491, 0.7840,  ..., 1.9099, 0.8000, 1.4389]],

        [[0.8919, 0.1616,    nan,  ..., 1.6940, 0.7771, 1.2511],
         [0.8903, 0.1537, 1.1706,  ..., 1.6989, 0.7763, 1.2511],
         [0.8887, 0.1457, 1.1778,  ..., 1.7038, 0.7756, 1.2511],
         ...,
         [0.8643, 0.0856, 1.2142,  ..., 2.0461, 0.7352, 1.4291],
         [0.8643, 0.0856, 1.2143,  ..., 2.0468,    nan, 1.4290],
         [0.8643, 0.0856, 1.2144,  ..., 2.0475, 0.7357, 1.4289]]],
       device='cuda:0')
mask
tensor([[[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        ...,

        [[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[0., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 0., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]]], device='cuda:0')
begin sampling, total number of reverse steps = 200
sampling
tensor([[[1.1412, 1.1559, 1.1576,  ..., 1.2361, 1.2461, 1.2500],
         [0.4368, 0.4410, 0.4420,  ..., 0.0964, 0.0978, 0.0971],
         [1.1162, 1.0996, 1.1074,  ..., 1.2289, 1.2308, 1.2424],
         ...,
         [2.0977, 2.0847, 2.0945,  ..., 1.9863, 1.9933, 2.0007],
         [0.7450, 0.7480, 0.7530,  ..., 0.8022, 0.7931, 0.8012],
         [1.4310, 1.4356, 1.4332,  ..., 1.4420, 1.4441, 1.4543]],

        [[0.8221, 0.7947, 0.7874,  ..., 0.7693, 0.7511, 0.7595],
         [0.1916, 0.1898, 0.1827,  ..., 0.9388, 0.0674, 0.0676],
         [1.2389, 1.2206, 1.2389,  ..., 1.2996, 1.2885, 1.2852],
         ...,
         [2.0686, 2.0736, 2.0697,  ..., 1.8188, 1.9743, 1.9823],
         [0.9665, 0.6771, 0.6687,  ..., 0.7446, 0.7418, 0.7412],
         [1.4374, 1.4392, 1.4408,  ..., 1.4306, 1.4332, 1.4317]],

        [[1.1703, 1.1822, 1.1828,  ..., 1.2879, 1.3491, 1.3423],
         [0.4293, 0.4218, 0.4275,  ..., 0.0899, 0.0849, 0.0916],
         [1.0657, 1.0015, 1.0549,  ..., 1.1630, 1.1656, 1.1677],
         ...,
         [2.0004, 1.0708, 2.0006,  ..., 1.9059, 2.0149, 2.0151],
         [0.7336, 0.7314, 0.7318,  ..., 0.7541, 0.7516, 0.7501],
         [1.4504, 1.4587, 1.4481,  ..., 1.4187, 1.4217, 1.4181]],

        ...,

        [[1.0682, 1.0448, 1.0484,  ..., 1.2401, 1.2242, 1.2217],
         [0.3477, 0.3396, 0.3379,  ..., 0.0224, 0.0318, 0.0292],
         [0.8865, 0.8353, 0.8820,  ..., 0.8761, 0.8784, 0.8738],
         ...,
         [2.1767, 1.0707, 2.1742,  ..., 1.6243, 1.7667, 1.7870],
         [1.0083, 1.0129, 1.0052,  ..., 0.7140, 0.7164, 0.7195],
         [1.3178, 1.3218, 1.3169,  ..., 1.4603, 1.4465, 1.0566]],

        [[0.8327, 0.8341, 0.8325,  ..., 0.8366, 0.8220, 0.8292],
         [0.0592, 0.0489, 0.0488,  ..., 0.0646, 0.0659, 0.0770],
         [0.8127, 0.7968, 0.9783,  ..., 0.7859, 0.7908, 0.7951],
         ...,
         [2.2340, 2.2087, 1.0673,  ..., 1.7593, 1.9084, 1.9173],
         [0.9822, 0.9877, 0.7963,  ..., 0.8026, 0.7995, 0.7947],
         [1.3796, 1.0383, 1.3912,  ..., 1.4411, 1.4388, 1.4432]],

        [[0.9006, 0.8987, 0.8979,  ..., 0.8822, 0.8697, 0.8759],
         [0.1779, 0.1778, 0.1725,  ..., 0.0979, 0.0950, 0.1054],
         [1.0417, 1.1341, 1.1627,  ..., 1.2154, 1.2080, 1.2150],
         ...,
         [1.6935, 1.6955, 1.7064,  ..., 1.7631, 2.0269, 2.0519],
         [0.7751, 0.7715, 0.7711,  ..., 0.7518, 0.9632, 0.7511],
         [1.2614, 1.2620, 1.2590,  ..., 1.4270, 1.4280, 1.4320]]],
       device='cuda:0')
generated 500 utterances of random_digit at iteration 13100 in 520 seconds
saved generated samples at iteration 13100
i = 1, batch = tensor([[[1.0553,    nan, 1.1449,  ..., 1.9301, 0.6993, 1.4439],
         [1.0567, 0.4301, 1.1453,  ..., 1.9307, 0.6992, 1.4440],
         [1.0580, 0.4269, 1.1457,  ..., 1.9313,    nan, 1.4440],
         ...,
         [1.1368, 0.0835, 1.1747,  ...,    nan,    nan, 1.4438],
         [1.1368, 0.0835, 1.1748,  ..., 2.0368, 0.7542, 1.4437],
         [1.1369, 0.0835, 1.1748,  ...,    nan, 0.7529, 1.4436]],

        [[0.8229, 0.0687, 1.0990,  ..., 2.0212,    nan, 1.4289],
         [0.8227, 0.0686,    nan,  ..., 2.0211, 0.7765, 1.4292],
         [0.8226, 0.0686, 1.0983,  ..., 2.0210, 0.7776, 1.4294],
         ...,
         [0.8183, 0.0678, 1.0954,  ..., 2.0309, 0.7477, 1.4280],
         [0.8183, 0.0678, 1.0954,  ..., 2.0309, 0.7477, 1.4280],
         [   nan, 0.0678, 1.0954,  ..., 2.0309, 0.7477, 1.4280]],

        [[1.0557, 0.2825, 1.1325,  ..., 1.9074, 0.7224, 1.4425],
         [1.0574, 0.2778, 1.1375,  ..., 1.9384,    nan, 1.4424],
         [   nan, 0.2730, 1.1425,  ..., 1.9523, 0.7155, 1.4428],
         ...,
         [1.1367,    nan, 1.3243,  ..., 2.3118, 0.7675, 1.4615],
         [1.1376, 0.0494, 1.3237,  ..., 2.3134, 0.7674, 1.4619],
         [1.1385, 0.0483, 1.3231,  ..., 2.3149, 0.7673, 1.4622]],

        ...,

        [[0.9977, 0.0757,    nan,  ..., 1.9773, 0.6946, 1.4264],
         [   nan, 0.0758, 1.1306,  ..., 1.9781, 0.6945, 1.4265],
         [0.9983, 0.0759, 1.1312,  ...,    nan, 0.6944,    nan],
         ...,
         [1.0040, 0.0751, 1.1466,  ..., 2.0080, 0.7469, 1.4274],
         [1.0039, 0.0752, 1.1464,  ..., 2.0087, 0.7458, 1.4273],
         [1.0038, 0.0752, 1.1461,  ..., 2.0092, 0.7448, 1.4273]],

        [[0.9887, 0.1455, 0.8031,  ..., 2.3996, 0.6468, 1.4524],
         [0.9899, 0.1399, 0.8027,  ..., 2.3960, 0.6442,    nan],
         [0.9910, 0.1343, 0.8022,  ..., 2.3923, 0.6416, 1.4529],
         ...,
         [1.0159, 0.0159,    nan,  ..., 2.0298, 0.7477, 1.4283],
         [   nan, 0.0159, 0.8228,  ..., 2.0298, 0.7477,    nan],
         [1.0159, 0.0159, 0.8228,  ..., 2.0298, 0.7477, 1.4283]],

        [[0.8170, 0.5136, 1.0637,  ..., 2.0002, 0.6289, 1.4583],
         [0.8134, 0.5153,    nan,  ..., 1.9988, 0.6278, 1.4587],
         [0.8097, 0.5171, 1.0703,  ..., 1.9967, 0.6266, 1.4591],
         ...,
         [0.4414, 0.5051,    nan,  ..., 2.0390, 0.7529, 1.4748],
         [   nan, 0.5023, 1.3675,  ..., 2.0412, 0.7527, 1.4749],
         [0.4344, 0.4995, 1.3698,  ..., 2.0435, 0.7524, 1.4751]]],
       device='cuda:0')
mask
tensor([[[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        ...,

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 0., 1., 0.],
         [1., 1., 0.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]]], device='cuda:0')
begin sampling, total number of reverse steps = 200
sampling
tensor([[[1.0754, 1.0635, 1.0657,  ..., 1.1387, 1.1374, 1.1386],
         [0.4240, 0.4282, 0.4303,  ..., 0.0816, 0.0825, 0.0830],
         [1.1519, 1.1492, 1.1453,  ..., 1.1673, 1.1782, 1.1816],
         ...,
         [1.9298, 1.9359, 1.9252,  ..., 2.0157, 2.0380, 2.0125],
         [0.7011, 0.7024, 0.6909,  ..., 0.7509, 0.7565, 0.7577],
         [1.4368, 1.4432, 1.4499,  ..., 1.4443, 1.4359, 1.4474]],

        [[0.8342, 0.8348, 0.8320,  ..., 0.8329, 0.8378, 0.9697],
         [0.1945, 0.0941, 0.0959,  ..., 0.0948, 0.0872, 0.0822],
         [1.0955, 1.0227, 1.0754,  ..., 1.0944, 1.0943, 1.0971],
         ...,
         [2.0175, 2.0214, 2.0161,  ..., 1.9233, 2.0121, 1.8905],
         [0.9695, 0.7920, 0.8190,  ..., 0.7026, 0.7463, 0.7439],
         [1.4328, 1.4308, 1.4284,  ..., 1.4321, 1.4255, 1.4319]],

        [[1.0553, 1.0548, 1.0158,  ..., 1.1391, 1.1383, 1.1371],
         [0.2617, 0.2801, 0.2797,  ..., 0.9347, 0.0725, 0.0655],
         [1.1338, 1.1422, 1.1456,  ..., 1.3246, 1.3296, 1.3240],
         ...,
         [1.9059, 1.9324, 1.9458,  ..., 2.1282, 2.2909, 2.0779],
         [0.7485, 0.9582, 0.7292,  ..., 0.7350, 0.7677, 0.7646],
         [1.4388, 1.4413, 1.4418,  ..., 1.4648, 1.4626, 1.4589]],

        ...,

        [[1.0004, 0.9986, 0.9974,  ..., 1.0070, 1.0073, 1.0065],
         [0.2270, 0.0979, 0.0877,  ..., 0.0825, 0.0786, 0.0852],
         [1.0330, 1.1204, 1.1256,  ..., 1.1418, 1.1485, 1.1506],
         ...,
         [1.9625, 1.9509, 1.0672,  ..., 1.9432, 2.0121, 1.9516],
         [0.6975, 0.6922, 0.5811,  ..., 0.7017, 0.7431, 0.7445],
         [1.4239, 1.4194, 1.0303,  ..., 1.4285, 1.4305, 1.4277]],

        [[0.9951, 0.9948, 0.9993,  ..., 1.0196, 1.0042, 1.0224],
         [0.2225, 0.1473, 0.1408,  ..., 0.0287, 0.0272, 0.0351],
         [0.8112, 0.8035, 0.8065,  ..., 0.9793, 0.8324, 0.8361],
         ...,
         [2.3863, 2.3983, 2.3525,  ..., 1.9104, 2.0291, 1.9532],
         [0.6576, 0.6581, 0.7665,  ..., 0.7297, 0.7473, 0.7453],
         [1.4332, 1.0481, 1.4434,  ..., 1.4225, 1.0530, 1.4152]],

        [[0.8278, 0.8201, 0.8191,  ..., 0.4611, 0.9467, 0.4570],
         [0.5644, 0.5279, 0.5382,  ..., 0.5044, 0.5039, 0.4991],
         [1.0518, 1.0196, 1.0616,  ..., 1.0538, 1.3582, 1.3605],
         ...,
         [2.0003, 1.9945, 1.9878,  ..., 1.6412, 2.0270, 1.8444],
         [0.6374, 0.6307, 0.5985,  ..., 0.6314, 0.7486, 0.7537],
         [1.4489, 1.4535, 1.4568,  ..., 1.4711, 1.4732, 1.4712]]],
       device='cuda:0')
generated 500 utterances of random_digit at iteration 13100 in 520 seconds
saved generated samples at iteration 13100
i = 2, batch = tensor([[[0.8245, 0.0314, 1.0366,  ..., 1.9505, 0.7791, 1.4425],
         [0.8246, 0.0313, 1.0360,  ..., 1.9499, 0.7763, 1.4422],
         [0.8246, 0.0312, 1.0354,  ..., 1.9493, 0.7743, 1.4420],
         ...,
         [   nan, 0.0245,    nan,  ..., 2.0308, 0.7305, 1.4305],
         [0.8259, 0.0245, 1.0324,  ..., 2.0308, 0.7300, 1.4303],
         [0.8259,    nan, 1.0324,  ...,    nan, 0.7295, 1.4302]],

        [[0.9510,    nan, 1.2523,  ..., 1.9817, 0.7593,    nan],
         [   nan, 0.0713, 1.2535,  ..., 1.9801, 0.7592,    nan],
         [0.9520, 0.0706, 1.2547,  ..., 1.9786, 0.7591, 1.4584],
         ...,
         [0.9741, 0.0188,    nan,  ..., 2.0130, 0.7568, 1.4280],
         [0.9741, 0.0189, 1.3206,  ..., 2.0121, 0.7571, 1.4260],
         [0.9741, 0.0189, 1.3207,  ..., 2.0111, 0.7573, 1.4251]],

        [[1.3443, 0.1158, 0.7285,  ..., 2.1105, 0.6802, 1.4532],
         [1.3488, 0.1097, 0.7270,  ..., 2.1118, 0.6802, 1.4534],
         [1.3534, 0.1037, 0.7256,  ..., 2.1131, 0.6802, 1.4536],
         ...,
         [1.5258, 0.1238, 0.9449,  ...,    nan, 0.7541, 1.5244],
         [1.5269, 0.1229, 0.9452,  ..., 1.9532, 0.7564, 1.4721],
         [1.5279, 0.1219, 0.9456,  ..., 1.9554, 0.7578, 1.4431]],

        ...,

        [[0.8755, 0.4597, 0.9047,  ..., 1.7845, 1.2424, 1.4637],
         [0.8737, 0.4632, 0.9002,  ..., 1.8082, 1.2332,    nan],
         [0.8720, 0.4667, 0.8956,  ..., 1.8299, 1.2239, 1.4669],
         ...,
         [0.6221, 0.4476, 0.7993,  ..., 2.0325, 0.6983, 1.4336],
         [0.6195, 0.4433, 0.7992,  ..., 2.0332, 0.6981, 1.4336],
         [0.6168, 0.4388, 0.7991,  ..., 2.0339,    nan, 1.4336]],

        [[0.8157, 0.5797, 0.9274,  ..., 1.8391, 0.7453, 1.4168],
         [0.8142, 0.5780, 0.9270,  ..., 1.8406, 0.7449, 1.4169],
         [0.8128, 0.5763,    nan,  ..., 1.8422, 0.7445, 1.4170],
         ...,
         [0.6968, 0.0324, 0.8111,  ..., 2.0827, 0.7188, 1.4279],
         [0.6956, 0.0229, 0.8092,  ..., 2.0842, 0.7188, 1.4280],
         [0.6950, 0.0185, 0.8086,  ..., 1.9603, 0.7287, 1.4235]],

        [[   nan, 0.3998, 1.0323,  ..., 2.1279, 0.7370, 1.7171],
         [0.9899,    nan, 1.0351,  ..., 2.0696, 0.7415, 1.7167],
         [0.9932,    nan,    nan,  ..., 2.0381, 0.7437, 1.7164],
         ...,
         [1.1799, 0.0773, 1.1582,  ..., 1.8047, 0.8232, 1.4840],
         [   nan, 0.0771, 1.1576,  ..., 1.8048, 0.8230, 1.4841],
         [1.1794, 0.0768, 1.1570,  ..., 1.8050,    nan, 1.4841]]],
       device='cuda:0')
mask
tensor([[[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        ...,

        [[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]],

        [[1., 1., 1.,  ..., 0., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 0., 1., 1.],
         ...,
         [1., 1., 1.,  ..., 1., 1., 0.],
         [1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 1., 1.]]], device='cuda:0')
begin sampling, total number of reverse steps = 200
sampling
tensor([[[0.8389, 0.8303, 0.8300,  ..., 0.8187, 0.8265, 0.8353],
         [0.0364, 0.0342, 0.0334,  ..., 0.0238, 0.0241, 0.0142],
         [1.0381, 1.0487, 1.0389,  ..., 1.0204, 1.0371, 1.0413],
         ...,
         [1.9576, 1.9542, 1.9496,  ..., 2.0331, 2.0246, 2.0075],
         [0.7759, 0.7819, 0.7787,  ..., 0.7354, 0.7250, 0.7389],
         [1.4427, 1.4387, 1.4423,  ..., 1.4284, 1.4284, 1.4397]],

        [[0.9582, 0.9996, 0.9614,  ..., 0.9941, 0.9793, 0.9810],
         [0.9424, 0.0922, 0.0953,  ..., 0.0312, 0.0265, 0.0474],
         [1.2559, 1.2547, 1.2453,  ..., 1.2926, 1.3198, 1.3168],
         ...,
         [1.9803, 1.9823, 1.9658,  ..., 2.0094, 2.0083, 1.9028],
         [0.7615, 0.7619, 0.7688,  ..., 0.7716, 0.7547, 0.7592],
         [1.0172, 1.0345, 1.4445,  ..., 1.4040, 1.4168, 1.4198]],

        [[1.3308, 1.3404, 1.3474,  ..., 1.4778, 1.5062, 1.5002],
         [0.1403, 0.1314, 0.1213,  ..., 0.1251, 0.1332, 0.2215],
         [0.7453, 0.7427, 0.7354,  ..., 0.9406, 0.9545, 0.9574],
         ...,
         [2.1107, 2.1170, 2.1055,  ..., 1.0620, 1.9256, 1.8730],
         [0.6846, 0.6850, 0.6809,  ..., 0.7561, 0.7522, 0.7628],
         [1.4524, 1.4561, 1.4566,  ..., 1.5302, 1.4671, 1.4504]],

        ...,

        [[0.8806, 0.8747, 0.8786,  ..., 0.6581, 0.6313, 0.6260],
         [0.4668, 0.4686, 0.4698,  ..., 0.4473, 0.4432, 0.4328],
         [0.9131, 0.8971, 0.8966,  ..., 0.8283, 0.8120, 0.8113],
         ...,
         [1.7874, 1.8095, 1.8246,  ..., 2.0357, 2.0321, 1.9546],
         [1.2123, 1.2300, 1.2180,  ..., 0.7058, 0.7100, 0.9560],
         [1.4447, 1.0437, 1.4640,  ..., 1.4301, 1.4325, 1.4357]],

        [[0.8243, 0.8215, 0.8211,  ..., 0.7043, 0.7050, 0.7107],
         [0.5843, 0.5825, 0.5806,  ..., 0.0533, 0.0420, 0.1041],
         [0.9332, 0.9370, 0.9769,  ..., 0.7939, 0.8225, 0.8228],
         ...,
         [1.8488, 1.8388, 1.8361,  ..., 2.0901, 2.0830, 1.9920],
         [0.7486, 0.7521, 0.7541,  ..., 0.7186, 0.7191, 0.7322],
         [1.4229, 1.4181, 1.4234,  ..., 1.4307, 1.4313, 1.4258]],

        [[1.0087, 0.9997, 1.0037,  ..., 1.1793, 1.0444, 1.1777],
         [0.4267, 0.9522, 0.9461,  ..., 0.0878, 0.0851, 0.1254],
         [1.0400, 1.0346, 1.0068,  ..., 1.1113, 1.1569, 1.1581],
         ...,
         [2.1230, 2.0742, 2.0403,  ..., 1.8014, 1.8088, 1.6902],
         [0.7375, 0.7443, 0.7465,  ..., 0.8223, 0.8265, 0.9753],
         [1.6986, 1.7115, 1.7094,  ..., 1.4607, 1.4785, 1.4869]]],
       device='cuda:0')
generated 500 utterances of random_digit at iteration 13100 in 520 seconds
saved generated samples at iteration 13100
i = 3, batch = tensor([[[0.9612, 0.0289, 0.8076,  ...,    nan, 0.7661, 1.4034],
         [0.9613, 0.0286, 0.8079,  ..., 1.9762, 0.7675, 1.4022],
         [0.9614, 0.0283,    nan,  ..., 1.9731,    nan, 1.4011],
         ...,
         [0.9668, 0.0156,    nan,  ..., 2.0508, 0.7445, 1.4274],
         [0.9668,    nan,    nan,  ..., 2.0474, 0.7452,    nan],
         [0.9668, 0.0156, 0.8230,  ..., 2.0449, 0.7458, 1.4272]],

        [[   nan, 0.4041,    nan,  ..., 1.7553, 0.8838, 1.6504],
         [0.9692, 0.4057,    nan,  ..., 1.7561, 0.8837, 1.6502],
         [0.9678, 0.4072,    nan,  ..., 1.7569,    nan, 1.6499],
         ...,
         [0.8153, 0.0972, 1.0612,  ..., 1.6002, 0.8202, 1.4285],
         [0.8140, 0.0926, 1.0648,  ..., 1.5975, 0.8290, 1.4272],
         [0.8127, 0.0880, 1.0685,  ..., 1.5972, 0.8330, 1.4265]],

        [[0.9200, 0.5036, 0.9379,  ..., 1.8845,    nan, 1.4690],
         [0.9190, 0.5023, 0.9342,  ..., 1.8893, 0.7720, 1.4695],
         [   nan, 0.5009, 0.9305,  ...,    nan, 0.7707, 1.4700],
         ...,
         [   nan, 0.0966, 0.9081,  ..., 1.9101, 0.7383, 1.4421],
         [0.8207, 0.0901, 0.9065,  ..., 1.9143, 0.7351, 1.4425],
         [0.8197, 0.0834, 0.9050,  ..., 1.9187, 0.7332, 1.4429]],

        ...,

        [[1.0392, 0.0858, 0.7099,  ..., 2.0302, 0.7902,    nan],
         [1.0393, 0.0855, 0.7099,  ..., 2.0083, 0.7879, 1.3768],
         [1.0394, 0.0854,    nan,  ..., 1.9963, 0.7867, 1.3773],
         ...,
         [1.0310, 0.0867, 0.6916,  ..., 2.0219, 0.7508, 1.4282],
         [1.0310, 0.0867, 0.6915,  ..., 2.0217, 0.7511, 1.4282],
         [1.0309, 0.0867, 0.6914,  ..., 2.0216, 0.7514, 1.4282]],

        [[1.0523,    nan, 0.8988,  ..., 2.4333, 0.7175, 1.4358],
         [1.0528, 0.1489, 0.8964,  ..., 2.4315, 0.7167, 1.4356],
         [   nan, 0.1486, 0.8946,  ..., 2.2290, 0.7265, 1.4373],
         ...,
         [1.0293, 0.0674, 0.7749,  ...,    nan, 0.7473, 1.4283],
         [1.0285, 0.0642, 0.7735,  ..., 2.0249, 0.7475, 1.4282],
         [1.0278, 0.0609,    nan,  ..., 2.0232, 0.7476,    nan]],

        [[0.8054,    nan, 0.7196,  ..., 2.0518, 0.6804, 1.4449],
         [   nan, 0.3803, 0.7170,  ..., 2.0529,    nan, 1.4449],
         [0.7976, 0.3777, 0.7145,  ..., 2.0539,    nan, 1.4449],
         ...,
         [   nan,    nan, 0.5117,  ..., 2.3870,    nan, 1.3886],
         [0.4733, 0.1122, 0.5076,  ..., 2.4115, 0.8275, 1.3821],
         [0.4711, 0.1128, 0.5035,  ..., 2.4220, 0.8364, 1.3797]]],
       device='cuda:0')
mask
tensor([[[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]],

        ...,

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]],

        [[1., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.],
         [1., 1., 0.,  ..., 0., 0., 1.],
         ...,
         [0., 1., 1.,  ..., 1., 1., 1.],
         [1., 1., 0.,  ..., 1., 1., 1.],
         [1., 1., 1.,  ..., 1., 0., 1.]]], device='cuda:0')
begin sampling, total number of reverse steps = 200
sampling
tensor([[[0.9678, 0.9660, 0.9653,  ..., 0.9734, 0.9700, 0.9742],
         [0.0289, 0.0309, 0.0348,  ..., 0.0232, 0.0149, 0.0141],
         [0.8132, 0.8032, 0.8028,  ..., 0.8101, 0.8145, 0.8273],
         ...,
         [1.9504, 1.9742, 1.9707,  ..., 2.0550, 2.0432, 2.0547],
         [0.7661, 0.7678, 0.7624,  ..., 0.7470, 0.7350, 0.7511],
         [1.4045, 1.4002, 1.3988,  ..., 1.4255, 1.4107, 1.4355]],

        [[0.9945, 0.9747, 0.9677,  ..., 0.8247, 0.8203, 0.8171],
         [0.4087, 0.4082, 0.4149,  ..., 0.1129, 0.1673, 0.0941],
         [1.0145, 1.0164, 1.0007,  ..., 1.0128, 1.0261, 1.0690],
         ...,
         [1.6436, 1.7451, 1.7579,  ..., 1.5687, 1.6121, 1.6092],
         [0.8776, 0.8785, 0.8640,  ..., 0.8242, 0.8231, 0.8251],
         [1.6431, 1.6462, 1.6372,  ..., 1.4219, 1.3523, 1.4191]],

        [[0.9247, 0.9220, 0.9920,  ..., 0.9658, 0.8419, 0.8332],
         [0.5078, 0.5043, 0.5017,  ..., 0.1060, 0.1367, 0.0987],
         [0.9477, 0.9335, 0.8342,  ..., 0.8674, 0.8568, 0.9061],
         ...,
         [1.7703, 1.8510, 1.0769,  ..., 1.9130, 1.9184, 1.9201],
         [0.9621, 0.7866, 0.7454,  ..., 0.7397, 0.7320, 0.7311],
         [1.4684, 1.4727, 1.4770,  ..., 1.4442, 1.4066, 1.4468]],

        ...,

        [[1.0385, 1.0373, 1.0369,  ..., 1.0341, 1.0309, 1.0294],
         [0.0952, 0.0983, 0.1033,  ..., 0.0951, 0.1594, 0.1064],
         [0.7187, 0.7230, 0.7071,  ..., 0.6624, 0.6871, 0.7192],
         ...,
         [1.7569, 1.9914, 1.9877,  ..., 2.0245, 2.0181, 2.0225],
         [0.7893, 0.7895, 0.7397,  ..., 0.7531, 0.7479, 0.7552],
         [1.0379, 1.3643, 1.3769,  ..., 1.4297, 1.3960, 1.4294]],

        [[1.0576, 1.0532, 1.0189,  ..., 1.0310, 1.0305, 1.0312],
         [0.9459, 0.1680, 0.1669,  ..., 0.0672, 0.1297, 0.0648],
         [0.8974, 0.8989, 0.8673,  ..., 0.7915, 0.7488, 0.9807],
         ...,
         [2.1509, 2.4060, 2.2282,  ..., 1.0714, 1.9877, 2.0284],
         [0.7276, 0.7256, 0.7497,  ..., 0.7488, 0.7466, 0.7437],
         [1.4353, 1.4383, 1.4293,  ..., 1.4333, 1.3045, 1.0330]],

        [[0.8180, 0.9658, 0.8066,  ..., 0.9693, 0.4967, 0.4898],
         [0.9479, 0.4064, 0.4072,  ..., 0.9471, 0.2505, 0.1259],
         [0.7301, 0.7330, 0.7497,  ..., 0.4366, 0.4951, 0.5257],
         ...,
         [1.8600, 2.0446, 2.0506,  ..., 2.3907, 2.3997, 2.4081],
         [0.7036, 0.9764, 0.8901,  ..., 0.9866, 0.8241, 0.8283],
         [1.4451, 1.4495, 1.4443,  ..., 1.3901, 1.3022, 1.3888]]],
       device='cuda:0')
generated 500 utterances of random_digit at iteration 13100 in 536 seconds
saved generated samples at iteration 13100

https://raw.githubusercontent.com/Josh-test-lab/website-assets-repository/refs/heads/main/posts/1140311%20meeting/Figure_1.png
缺失值更換為 “1” 對於實驗 1 第 3 個特徵值進行填補的結果。

結論

我對於本文程式碼或許還有許多不足或需要認識之處，希望下一次能改進 predict.py 腳本，達到可以預測的目的。

運行環境

  • 作業系統：Windows 11 24H2
  • 程式語言：Python 3.12.9

參考資料

更新於 2025-03-09 18:05 UTC+0800   da02093e
Creative Commons License
 Meeting, Python, $SSSD^{S4}$
返回 | 首頁

相關文章