I get a vary high loss

[sTarAnna]

Thanks for your great job.I rewrite you program to make it run with tensorflow1.12 and python 3. I only use 2BLSTM not 4.At the begining of train stage ,the train loss is 3000k, after 66k steps ,the train loss is 1000k . I want to know thats correct or not.

dc model ` with tf.variable_scope('BLSTM1') as scope:

        layer1_fw = tf.nn.rnn_cell.LSTMCell(self.n_hidden)
        layer1_bw = tf.nn.rnn_cell.LSTMCell(self.n_hidden)

        #dropout
        layer1_fw_dropout = tf.nn.rnn_cell.DropoutWrapper(layer1_fw,self.p_keep_fw)
        layer1_bw_dropout = tf.nn.rnn_cell.DropoutWrapper(layer1_bw,self.p_keep_fw)

        #layer1 outputs
        layer1_outputs,_ = tf.nn.bidirectional_dynamic_rnn(layer1_fw_dropout,layer1_bw_dropout,x,sequence_length=[FRAMES_PER_SAMPLE] * self.batch_size,dtype=tf.float32)

        #拼接两个输出
        layer1_output = tf.concat(layer1_outputs,2)
        #第一层结束

    #第二层BLSTM
    with tf.variable_scope('BLSTM2') as scope:
        layer2_fw = tf.nn.rnn_cell.LSTMCell(self.n_hidden)
        layer2_bw = tf.nn.rnn_cell.LSTMCell(self.n_hidden)

        # dropout
        layer2_fw_dropout = tf.nn.rnn_cell.DropoutWrapper(layer2_fw, self.p_keep_fw)
        layer2_bw_dropout = tf.nn.rnn_cell.DropoutWrapper(layer2_bw, self.p_keep_fw)

        # layer1 output
        layer2_outputs, _ = tf.nn.bidirectional_dynamic_rnn(layer2_fw_dropout, layer2_bw_dropout, layer1_output,
                                                            sequence_length=[FRAMES_PER_SAMPLE] * self.batch_size,
                                                            dtype=tf.float32)

        # 拼接两个输出
        layer2_output = tf.concat(layer2_outputs, 2)
        # 第二层结束

    #feedfoward layer
    with tf.variable_scope('feedfoward') as scope:
        blstm_output = tf.reshape(layer2_output, [-1, self.n_hidden * 2])
        emb_out = tf.matmul(blstm_output,self.weights['out']) + self.biases['out']
        #tanh激活函数
        emb_out = tf.nn.tanh(emb_out)
        reshaped_emb = tf.reshape(emb_out, [-1, NEFF, EMBBEDDING_D])
        # #L2正则化
        normalized_emb = tf.nn.l2_normalize(reshaped_emb, 2)
    return normalized_emb`

[njusq]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

[sTarAnna]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

I rewrite my model and it now have 4blstm layers .I found that the loss value become lower and convergence faster .After training I would use audio_test to see if it works . Thanks for your reply！

[zhr1201]

Thanks for your reply! Btw @njusq我也是nju的，校友你好啊

On 04/18/2019 14:48, njusq wrote:

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[sTarAnna]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot!

[JayYang-Fdu]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot!

you can change the player .

[njusq]

Thanks for your reply! Btw @njusq我也是nju的，校友你好啊

On 04/18/2019 14:48, njusq wrote:

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

啊啊啊鸡冻！好像师兄也是Acoustics的！老师让我先通过复现这个项目来学习相关的知识，从零开始的小白觉得好艰难TAT，只好到GitHub上看看有没有现成的...然后我现在在用这个框架，改掉loss的计算方式做成DANet的，碰到了许多问题，可以向师兄请教吗？

[njusq]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ?

I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot!

I once encountered the same problem opening the file through Windows Media Player. I guess something related the format of the file leads to this problem. Maybe you can try to open the file through Adobe Audition or under your Linux system.

[sTarAnna]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ? I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot!

I once encountered the same problem opening the file through Windows Media Player. I guess something related the format of the file leads to this problem. Maybe you can try to open the file through Adobe Audition or under your Linux system.

可以，解决了，给力。谢谢。

[sTarAnna]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ? I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot!

you can change the player .

yeah. I open it under linux os and it has no problem .Thanks a lot

[sTarAnna]

I ran the program(4 BiLSTM) with tensorflow1.12 and python3.6, too. The mixture can be separated successfully although the loss value seems to be high. Have you ever checked it with audio_test.py ? I got a problem that the separated speech file (.wav) cant play . Could you please tell me how did solve this problem ? Thank you a lot! you can change the player . yeah. I open it under linux os and it has no problem .Thanks a lot 师兄你好呀，我能加一下你的微信吗？我感觉自己快要毕不了业了。。。我自己之前想过要换神经网络，但是不知道为什么会出现矩阵维度不匹配，百度也找不到。。。这个不知道是怎么解决的呀！感谢师兄的回复！

你想换成什么神经网络0 0

[zhr1201]

@njusq 好啊好啊，师弟是卢晶老师组里的吗？

[zhr1201]

@Yangjie55 haoran7123微信 多多交流

[zhr1201]

@sTarAnna O_O O_o o_o o_O O_O

[sTarAnna]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

[JayYang-Fdu]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

突然变成大兄弟哈哈哈，我是想换成ｃｒｎｎ的网络，可是我写的时候，总会在ｌｏｓｓ的时候，说矩阵元素相乘的时候，维度是不匹配的。我就有点蒙。。。师兄你有实现ｃｒｎｎ的网络嘛？

[JayYang-Fdu]

@njusq 好啊好啊，师弟是卢晶老师组里的吗？

真羡慕师兄你们哦，我这本科毕设都搞得要头秃了哈哈哈。。。还不知道研究生可咋办哈哈。。也是自己太菜了，难受了难受了。。。

[JayYang-Fdu]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

然后就是论文里面的公式和那些参数，讲实话，我又看了一遍，还是没有看懂什么意思。。。很尴尬。。师兄你有联系方式吗？

[sTarAnna]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

突然变成大兄弟哈哈哈，我是想换成ｃｒｎｎ的网络，可是我写的时候，总会在ｌｏｓｓ的时候，说矩阵元素相乘的时候，维度是不匹配的。我就有点蒙。。。师兄你有实现ｃｒｎｎ的网络嘛？

我也是做毕设的0 0 我没实现CRNN ，我按照18年的GCDC实现了GCN做DC的

[JayYang-Fdu]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

突然变成大兄弟哈哈哈，我是想换成ｃｒｎｎ的网络，可是我写的时候，总会在ｌｏｓｓ的时候，说矩阵元素相乘的时候，维度是不匹配的。我就有点蒙。。。师兄你有实现ｃｒｎｎ的网络嘛？

我也是做毕设的0 0 我没实现CRNN ，我按照18年的GCDC实现了GCN做DC的

那你超级厉害呀，我感觉很多东西理解不了。。。唉。。心慌慌。我问你几个问题阔以 不： 1.感觉周浩然师兄在神经网络后面没有加feedforwordlayer，原文中好像有哦（可能是我没看懂），加上师兄用的这个layernormbaisccell和basiccell的神经单元我不知道有什么区别耶。 2. 这个应该是程序里面的loss的计算方式，可是里面的D的-二分之一，那里我看不懂。。。 3.还有原文中的这句话：,and the embedding can be considered a permutation-and cardinality-independent encoding of the network’s estimate of the signal partition.所以这个embedding到底是什么。。我到现在都没理解。 4.还有原文中的V = fθ(x) ∈RN×K，这个我地方我没看懂唉，还请指教，感谢！

[JayYang-Fdu]

@sTarAnna O_O O_o o_o o_O O_O

怎么了大兄弟0 0

突然变成大兄弟哈哈哈，我是想换成ｃｒｎｎ的网络，可是我写的时候，总会在ｌｏｓｓ的时候，说矩阵元素相乘的时候，维度是不匹配的。我就有点蒙。。。师兄你有实现ｃｒｎｎ的网络嘛？

我也是做毕设的0 0 我没实现CRNN ，我按照18年的GCDC实现了GCN做DC的

我能借鉴一下你的model的code嘛？（ps:我有点不要脸哈哈哈哈） 你实现的这个方法我听都没听过。。。膜拜一下大佬！！

[njusq]

我能借鉴一下你的model的code嘛？（ps:我有点不要脸哈哈哈哈） 你实现的这个方法我听都没听过。。。膜拜一下大佬！！

@Yangjie55 也在学习过程中的我和你交流吧...wechat（去掉#号）: s#q#i#a#n#9#6#0#7#2#4

[sTarAnna]

@sTarAnna O_O O_o o_o o_O O_O 怎么了大兄弟0 0 突然变成大兄弟哈哈哈，我是想换成ｃｒｎｎ的网络，可是我写的时候，总会在ｌｏｓｓ的时候，说矩阵元素相乘的时候，维度是不匹配的。我就有点蒙。。。师兄你有实现ｃｒｎｎ的网络嘛？ 我也是做毕设的0 0 我没实现CRNN ，我按照18年的GCDC实现了GCN做DC的

那你超级厉害呀，我感觉很多东西理解不了。。。唉。。心慌慌。我问你几个问题阔以 不： 1.感觉周浩然师兄在神经网络后面没有加feedforwordlayer，原文中好像有哦（可能是我没看懂），加上师兄用的这个layernormbaisccell和basiccell的神经单元我不知道有什么区别耶。 2. 这个应该是程序里面的loss的计算方式，可是里面的D的-二分之一，那里我看不懂。。。 3.还有原文中的这句话：,and the embedding can be considered a permutation-and cardinality-independent encoding of the network’s estimate of the signal partition.所以这个embedding到底是什么。。我到现在都没理解。 4.还有原文中的V = fθ(x) ∈RN×K，这个我地方我没看懂唉，还请指教，感谢！

1.加了 `# one layer of embedding output with tanh activation function out_concate = tf.reshape(state_concate4, [-1, self.n_hidden * 2]) emb_out = tf.matmul(out_concate, self.weights['out']) + self.biases['out'] emb_out = tf.nn.tanh(emb_out) reshaped_emb = tf.reshape(emb_out, [-1, NEFF, EMBBEDDING_D])

normalization before output

    normalized_emb = tf.nn.l2_normalize(reshaped_emb, 2)
    return normalized_emb`

2.你看的不是正式版吧0 0正式版公式不长这样 3.DC通过亲和矩阵，实现了置换无关的方法。引用GCDC内的一句话，网络让相同源的向量平行，否则正交。表现在矩阵里就是相同源的值比较大，不同源的比较小。构建亲和矩阵解决了置换问题（另一种可用的是PIT）。This implies that this objective function encourages the mapped embedding vectors to become parallel if they are dominated by the same source and become orthogonal otherwise. 4.R N* K 就是V是一个N* K矩阵 5.GCDC : https://ieeexplore.ieee.org/document/8461746

[JayYang-Fdu]

....不是正版吗...可是我看的就是周师兄在github上贴的那个呀...感觉你好厉害呀...在本科就被拉开了...苦笑中...介不介意加个微信聊呀...带我走进语音处理中..😁

[sTarAnna]

....不是正版吗...可是我看的就是周师兄在github上贴的那个呀...感觉你好厉害呀...在本科就被拉开了...苦笑中...介不介意加个微信聊呀...带我走进语音处理中..😁

1.这是DC正式发表的论文：https://ieeexplore.ieee.org/document/7471631 这篇论文的结构是2层BLSTM。

2. 这里https://www.isca-speech.org/archive/Interspeech_2016/搜 Single-channel multi-speaker separation using deep clustering ，这篇论文是探讨了4层BLSTM以及更宽的网络的0 0。里面还有对K-means的改进和End2End2的讨论。
3. 信号处理那部分我自己也很懵逼0 0

[JayYang-Fdu]

....不是正版吗...可是我看的就是周师兄在github上贴的那个呀...感觉你好厉害呀...在本科就被拉开了...苦笑中...介不介意加个微信聊呀...带我走进语音处理中..grin

1.这是DC正式发表的论文：https://ieeexplore.ieee.org/document/7471631 这篇论文的结构是2层BLSTM。

2. 这里https://www.isca-speech.org/archive/Interspeech_2016/搜 Single-channel multi-speaker separation using deep clustering ，这篇论文是探讨了4层BLSTM以及更宽的网络的0 0。里面还有对K-means的改进和End2End2的讨论。
3. 信号处理那部分我自己也很懵逼0 0

好的好的　感谢大兄弟啦　那我把这些文章读了先！　再一次感谢！！！