santi-pdp/segan

load failed when running train_segan.sh

andreaschandra opened this issue · 6 comments

andreaschandra commented

RuntimeError: module compiled against API version 0xb but this version of numpy is 0xa
Parsed arguments: {'deconv_type': 'deconv', 'd_label_smooth': 0.25, 'z_depth': 256, 'preemph': 0.95, 'seed': 111, 'init_noise_std': 0.0, 'synthesis_path': 'dwavegan_samples', 'e2e_dataset': 'data/segan.tfrecords', 'save_freq': 50, 'g_type': 'ae', 'l1_remove_epoch': 150, 'epoch': 86, 'bias_D_conv': True, 'save_path': 'segan_allbiased_preemph', 'test_wav': None, 'g_learning_rate': 0.0002, 'unrolled_lstm': False, 'z_dim': 256, 'batch_size': 50, 'bias_downconv': True, 'denoise_epoch': 5, 'canvas_size': 16384, 'noise_decay': 0.7, 'g_nl': 'prelu', 'beta_1': 0.5, 'd_learning_rate': 0.0002, 'denoise_lbound': 0.01, 'bias_deconv': True, 'weights': None, 'model': 'gan', 'save_clean_path': 'test_clean_results', 'init_l1_weight': 100.0}
Using device: /cpu:0
Creating GAN model
*** Applying pre-emphasis of 0.95 ***
*** Building Generator ***
Biasing downconv in G
Downconv (50, 16384, 1) -> (50, 8192, 16)
Adding skip connection downconv 0
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 8192, 16) -> (50, 4096, 32)
Adding skip connection downconv 1
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 4096, 32) -> (50, 2048, 32)
Adding skip connection downconv 2
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 2048, 32) -> (50, 1024, 64)
Adding skip connection downconv 3
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 1024, 64) -> (50, 512, 64)
Adding skip connection downconv 4
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 512, 64) -> (50, 256, 128)
Adding skip connection downconv 5
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 256, 128) -> (50, 128, 128)
Adding skip connection downconv 6
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 128, 128) -> (50, 64, 256)
Adding skip connection downconv 7
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 64, 256) -> (50, 32, 256)
Adding skip connection downconv 8
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 32, 256) -> (50, 16, 512)
Adding skip connection downconv 9
-- Enc: prelu activation --
Biasing downconv in G
Downconv (50, 16, 512) -> (50, 8, 1024)
-- Enc: prelu activation --
g_dec_depths: [512, 256, 256, 128, 128, 64, 64, 32, 32, 16, 1]
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 8, 2048) -> (50, 16, 512)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 16, 512)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 16, 1024) -> (50, 32, 256)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 32, 256)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 32, 512) -> (50, 64, 256)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 64, 256)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 64, 512) -> (50, 128, 128)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 128, 128)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 128, 256) -> (50, 256, 128)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 256, 128)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 256, 256) -> (50, 512, 64)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 512, 64)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 512, 128) -> (50, 1024, 64)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 1024, 64)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 1024, 128) -> (50, 2048, 32)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 2048, 32)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 2048, 64) -> (50, 4096, 32)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 4096, 32)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 4096, 64) -> (50, 8192, 16)
-- Dec: prelu activation --
Fusing skip connection of shape (50, 8192, 16)
-- Transposed deconvolution type --
Biasing deconv in G
Deconv (50, 8192, 32) -> (50, 16384, 1)
-- Dec: tanh activation --
Amount of alpha vectors: 21
Amount of skip connections: 10
Last wave shape: (50, 16384, 1)

num of G returned: 23
*** Discriminator summary ***
D block 0 input shape: (50, 16384, 2) *** biasing D conv *** downconved shape: (50, 8192, 16) *** Applying VBN *** Applying Lrelu ***
D block 1 input shape: (50, 8192, 16) *** biasing D conv *** downconved shape: (50, 4096, 32) *** Applying VBN *** Applying Lrelu ***
D block 2 input shape: (50, 4096, 32) *** biasing D conv *** downconved shape: (50, 2048, 32) *** Applying VBN *** Applying Lrelu ***
D block 3 input shape: (50, 2048, 32) *** biasing D conv *** downconved shape: (50, 1024, 64) *** Applying VBN *** Applying Lrelu ***
D block 4 input shape: (50, 1024, 64) *** biasing D conv *** downconved shape: (50, 512, 64) *** Applying VBN *** Applying Lrelu ***
D block 5 input shape: (50, 512, 64) *** biasing D conv *** downconved shape: (50, 256, 128) *** Applying VBN *** Applying Lrelu ***
D block 6 input shape: (50, 256, 128) *** biasing D conv *** downconved shape: (50, 128, 128) *** Applying VBN *** Applying Lrelu ***
D block 7 input shape: (50, 128, 128) *** biasing D conv *** downconved shape: (50, 64, 256) *** Applying VBN *** Applying Lrelu ***
D block 8 input shape: (50, 64, 256) *** biasing D conv *** downconved shape: (50, 32, 256) *** Applying VBN *** Applying Lrelu ***
D block 9 input shape: (50, 32, 256) *** biasing D conv *** downconved shape: (50, 16, 512) *** Applying VBN *** Applying Lrelu ***
D block 10 input shape: (50, 16, 512) *** biasing D conv *** downconved shape: (50, 8, 1024) *** Applying VBN *** Applying Lrelu ***
discriminator deconved shape: (50, 8, 1024)
discriminator output shape: (50, 1)

Not clipping D weights
Initializing optimizers...
Initializing variables...
Sampling some wavs to store sample references...
sample noisy shape: (50, 16384)
sample wav shape: (50, 16384)
sample z shape: (50, 8, 1024)
total examples in TFRecords data/segan.tfrecords: 48650
Batches per epoch: 973
[*] Reading checkpoints...
[!] Load failed
./train_segan.sh: line 20: 13602 Killed CUDA_VISIBLE_DEVICES="" python main.py --init_noise_std 0. --save_path segan_allbiased_preemph --init_l1_weight 100. --batch_size 50 --g_nl prelu --save_freq 50 --preemph 0.95 --epoch 86 --bias_deconv True --bias_downconv True --bias_D_conv True

bailiangze commented

i have this problem ,can you solve it

andreaschandra commented

i have this problem ,can you solve it

yes, i run it with smaller dataset, and if you just want to test the weight that has been train, you just attach it to the script for clean_wav.sh

bailiangze commented

thank your reply。your answer help me solve this problem。Do you have a prtrained model ,because I do not have GPU.this trainning will spend me a lot of time。can you give me a pretrained model ; my tf is 1.9version.. I find the author‘s model is incompatile with me

andreaschandra commented

I see, but I was also using pretrained model from the author. I computed on Linux 16.10 and using the library version from requirements. you can downgrade the tensorflow version and make virtual environment for it

fengqiyun commented

Have you successfully replicated the experiment？I find that the g_adv loss is always 1 in training.And the loss seems to be different from the picture provided by the author.Can you show me the g_loss where you trained? My e-mail：zhouyaofeng@hotmail.com

fycheng-code commented

Could someone tell me where I can find the pretrained model? Thanks!