Conda Environment

[halil-erdogan]

I have added spec-file to get same environment and reproduce results easily. @sertansenturk May you check it?

Instruction

To use the spec file to create an identical environment on the same machine or another machine:
conda create --name myenv --file spec-file.txt

After that, you can check this pull request which is about mock training and accept my pull request.

[sertansenturk]

There are a few issues with this request:

1. You should add an "installation instructions" to the README. It should include:
   1.1. installing conda (even if it's following the instructions on their website)
   1.2. creating the environment (basically what you have written above)
   1.3 activating/running/deactivating the conda environment

2. Could you export the environment to yml instead of txt for the sake of readability and convenience? (instructions: https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#sharing-an-environment)

3. We do not have any code using these packages at the moment (in fact we don't have any code in the repo yet).
   At this stage, we should simply create a conda environment with the bare minimum (i.e. the correct Python version with no external dependencies). The environment should be populated later with the appropriate packages, when an additional feature requires them.

4. Also see the inline comments in the added file.

[halil-erdogan]

Thanks for clarification.

I have change my pull request according to your requests. Now, you can find followings in the Readme:

• How to install conda
• How to get same environment
• Activate and deactivate environment

Also, I have added environment.yml instead of spec-file.txt

[sertansenturk]

No need to refer to the original sources. Let's keep the README focused.

[sertansenturk]

Are these two options identical in function? If yes, we can use one of them, that would be enough.

[halil-erdogan]

Yes, both of them worked for me. I have added both of them, if people get error from one of them. But, we can remove one of them.

[sertansenturk]

Please remove one...

[sertansenturk]

When I run this command I get:

Collecting package metadata: done
Solving environment: failed

ResolvePackageNotFound: 
  - ncurses==6.1=h0a44026_1
  - sqlite==3.26.0=ha441bb4_0
  - xz==5.2.4=h1de35cc_4
  - zlib==1.2.11=h1de35cc_3
  - python==3.7.2=haf84260_0
  - openssl==1.1.1a=h1de35cc_0
  - libedit==3.1.20181209=hb402a30_0
  - libcxx==4.0.1=hcfea43d_1
  - tk==8.6.8=ha441bb4_0
  - libcxxabi==4.0.1=hcfea43d_1
  - libffi==3.2.1=h475c297_4
  - readline==7.0=h1de35cc_5

I have Ubuntu 16.04.

[sertansenturk]

When I remove the hashes after second equals sign each line, i.e.:

name: turkishMusicGen
channels:
  - defaults
dependencies:
  - ca-certificates=2018.12.5
  - certifi=2018.11.29
  - libcxx=4.0.1
  - libcxxabi=4.0.1
  - libedit=3.1.20181209
  - libffi=3.2.1
  - ncurses=6.1
  - openssl=1.1.1a
  - pip=18.1
  - python=3.7.2
  - readline=7.0
  - setuptools=40.6.3
  - sqlite=3.26.0
  - tk=8.6.8
  - wheel=0.32.3
  - xz=5.2.4
  - zlib=1.2.11
prefix: /Users/herdogan/anaconda3/envs/turkishMusicGen

I get rid of most errors. The remaining errors are:

Collecting package metadata: done
Solving environment: failed

ResolvePackageNotFound: 
  - libcxxabi=4.0.1
  - libcxx=4.0.1

When I removed these two packages the environment was created successfully. The environment.yml file that worked in my case is:

name: turkishMusicGen
channels:
  - defaults
dependencies:
  - ca-certificates=2019.1.23-0
  - certifi=2018.11.29
  - libedit=3.1.20181209
  - libffi=3.2.1
  - ncurses=6.1
  - openssl=1.1.1a
  - pip=18.1
  - python=3.7.2
  - readline=7.0
  - setuptools=40.6.3
  - sqlite=3.26.0
  - tk=8.6.8
  - wheel=0.32.3
  - xz=5.2.4
  - zlib=1.2.11
prefix: /Users/herdogan/anaconda3/envs/turkishMusicGen

[sertansenturk]

It seems conda is not as OS-agnostic as it looks from outside. See the Github issue here: conda/conda#6073

I am getting inclined to switch to Docker, but

1. it is outside the scope of this ticket
2. You are not familiar with Docker for the time being so I am not in favor of increasing the complexity of our setup. We need to spend our time on the models.

so let's take care of the reviews, and then make sure that the environment works in Magdalena's computer & cluster.

[sertansenturk]

I have given out some reviews. It seems that conda may not be the ideal solution in the long run (see: this comment).

In addition, we should have considered that the code of the state-of-the-art music generation methods was built using different libraries and potentially different versions. It's also not a secret that both the numerical stability and the behavior of DL libraries have changed throughout versions. I'm afraid we may need to have an environment per method for the sake of reproducibility.

For this reason, I'm inclined to close this ticket once the reviews are done and we make sure we get an expected behavior on Magdalena's side.

To reflect the reproducibility concerns, I'm adding a criterion to the evaluation of open source models #19.

[halil-erdogan]

I agree with you. If I have time, I will try to use Docker. If we can use directly Docker, it would be perfect. Otherwise, as you said, we will provide different .yml file for each methods-architectures, however, our problem with conda (non-os-agnostic .yml) will continue.

[sertansenturk]

I agree with you. If I have time, I will try to use Docker. If we can use directly Docker, it would be perfect. Otherwise, as you said, we will provide different .yml file for each methods-architectures, however, our problem with conda (non-os-agnostic .yml) will continue.

Let's keep Docker out of the scope for the time being. It's a vast area, and without guidance, it will distract you from working on the actual problem -> music generation.

[halil-erdogan]

@sertansenturk Hocam, should we delete this branch? I think, we won't go with conda at that point.