我替换成 deepspeech-0.9.2-models.pbmm会报错

[tailangjun]

感觉默认的 deepspeech是v0.1，我就换成 0.9.2或者 0.9.3，都会报错

Traceback (most recent call last):
File "/home/tailangjun/Documents/AIGenHuman/NerfLips/DFRF/data_util/deepspeech_features/extract_ds_features.py", line 133, in
main()
File "/home/tailangjun/Documents/AIGenHuman/NerfLips/DFRF/data_util/deepspeech_features/extract_ds_features.py", line 125, in main
extract_features(
File "/home/tailangjun/Documents/AIGenHuman/NerfLips/DFRF/data_util/deepspeech_features/extract_ds_features.py", line 82, in extract_features
conv_audios_to_deepspeech(
File "/home/tailangjun/Documents/AIGenHuman/NerfLips/DFRF/data_util/deepspeech_features/deepspeech_features.py", line 41, in conv_audios_to_deepspeech
graph, logits_ph, input_node_ph, input_lengths_ph = prepare_deepspeech_net(
File "/home/tailangjun/Documents/AIGenHuman/NerfLips/DFRF/data_util/deepspeech_features/deepspeech_features.py", line 102, in prepare_deepspeech_net
graph_def.ParseFromString(f.read())
google.protobuf.message.DecodeError: Error parsing message with type 'tensorflow.GraphDef'

请问是不是要修改啥代码呀

[beria-moon]

请问你解决了吗

[tailangjun]

请问你解决了吗

我使用 pbmm_to_pb.py可以将 pbmm转换pb，但是维度上会报错

deepspeech-0.1.0
Tensor("deepspeech/logits:0", shape=(?, ?, 29), dtype=float32)
正常输出 (120, 29)

deepspeech-0.8.2
Tensor("deepspeech/logits:0", shape=(16, 1, 29), dtype=float32)
ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)'

deepspeech-0.9.3
Tensor("deepspeech/logits:0", shape=(16, 1, 256), dtype=float32)
ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)'

[beria-moon]

我用的0.1.0 输出是（16，29，1） 894269281 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "sstzal/DFRF" ***@***.***>; 发送时间: 2024年4月25日(星期四) 中午12:43 ***@***.***>; ***@***.******@***.***>; 主题: Re: [sstzal/DFRF] 我替换成 deepspeech-0.9.2-models.pbmm会报错 (Issue #36) 请问你解决了吗 我使用 pbmm_to_pb.py可以将 pbmm转换pb，但是维度上会报错 deepspeech-0.1.0 Tensor("deepspeech/logits:0", shape=(?, ?, 29), dtype=float32) 正常输出 (120, 29) deepspeech-0.8.2 Tensor("deepspeech/logits:0", shape=(16, 1, 29), dtype=float32) ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)' deepspeech-0.9.3 Tensor("deepspeech/logits:0", shape=(16, 1, 256), dtype=float32) ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)' — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: ***@***.***>

[tailangjun]

我用的0.1.0 输出是（16，29，1） 894269281 @.***  
…
------------------ 原始邮件 ------------------ 发件人: "sstzal/DFRF" @.>; 发送时间: 2024年4月25日(星期四) 中午12:43 @.>; @.@.>; 主题: Re: [sstzal/DFRF] 我替换成 deepspeech-0.9.2-models.pbmm会报错 (Issue #36) 请问你解决了吗 我使用 pbmm_to_pb.py可以将 pbmm转换pb，但是维度上会报错 deepspeech-0.1.0 Tensor("deepspeech/logits:0", shape=(?, ?, 29), dtype=float32) 正常输出 (120, 29) deepspeech-0.8.2 Tensor("deepspeech/logits:0", shape=(16, 1, 29), dtype=float32) ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)' deepspeech-0.9.3 Tensor("deepspeech/logits:0", shape=(16, 1, 256), dtype=float32) ValueError: Cannot feed value of shape (1, 240, 494) for Tensor 'deepspeech/input_node:0', which has shape '(1, 16, 19, 26)' — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you commented.Message ID: @.***>

估计是对 network_output的处理方式不太一样

            network_output = sess.run(
                self.logits_ph,
                feed_dict={
                    self.input_node_ph: input_vector[np.newaxis, ...],
                    self.input_lengths_ph: [input_vector.shape[0]]})'
            
            ds_features = network_output[::2,0,:]

AD-NeRF、DFRF、ER-NeRF在这块也是有点不同的