Converting from tensorflow to pytorch

[slala2121]

Platform (like ubuntu 16.04/win10): ubuntu 16.04

Python version: 3

Source framework with version (like Tensorflow 1.4.1 with GPU): Tensorflow 1.15.0

Destination framework with version (like CNTK 2.3 with GPU): Pytorch 1.6.0

Since the models under Tensorflow 1.11+ are saved as .meta, .index, .data, how can tf to IR step (#22) here be modified for this type of saved model format? Seems like .ckpt is used for TensorflowParser._load_weights but this is not available under this version.

[zzb112233]

Platform (like ubuntu 16.04/win10): ubuntu 16.04

Python version: 3

Source framework with version (like Tensorflow 1.4.1 with GPU): Tensorflow 1.15.0

Destination framework with version (like CNTK 2.3 with GPU): Pytorch 1.6.0

Since the models under Tensorflow 1.11+ are saved as .meta, .index, .data, how can tf to IR step (#22) here be modified for this type of saved model format? Seems like .ckpt is used for TensorflowParser._load_weights but this is not available under this version.

Hello,have you solved this problem?

[slala2121]

No but the issue can be closed since I found a workaround.

…

On Wed, Oct 21, 2020 at 1:22 AM zzb112233 ***@***.***> wrote: Platform (like ubuntu 16.04/win10): ubuntu 16.04 Python version: 3 Source framework with version (like Tensorflow 1.4.1 with GPU): Tensorflow 1.15.0 Destination framework with version (like CNTK 2.3 with GPU): Pytorch 1.6.0 Since the models under Tensorflow 1.11+ are saved as .meta, .index, .data, how can tf to IR step (#22 <#22>) here be modified for this type of saved model format? Seems like .ckpt is used for TensorflowParser._load_weights but this is not available under this version. Hello,have you solved this problem? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#885 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN3GCFRH6GYG3LVXZN6PG2LSL2K57ANCNFSM4P6YX3JA> .

[waleedrazakhan92]

No but the issue can be closed since I found a workaround.
…
can you share the workaround?

[slala2121]

I just loaded the weights using the tf graph. Then I wrote a script to map those weights to a corresponding pytorch graph. You can do this by inspecting the tf graph via tensorboard.

…

On Thu, Oct 22, 2020 at 3:59 AM waleedrazakhan92 ***@***.***> wrote: No but the issue can be closed since I found a workaround. … <#m_-8538388121266979976_> can you share the workaround? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#885 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AN3GCFQB3N7Q6JYTEL2ENE3SMAGBRANCNFSM4P6YX3JA> .