Gaby neubias

[turekg]

In principle the yaml work is done. Still nebulous about the output with the "extra" quotes around the arrays. BUT I cannot find anywhere examples of values that are arrays. And I've tried all sorts of things to see if I can get rid of them. Anyways, we can go from here...

[turekg]

I've figured out why the test fails, it's because it reads the wrong input file... Will fix.

[turekg]

Apparently I am not allowed to upload a yaml file, so here is the content of one for testing with java (SEM 2D example)

language: python
framework: tensorflow
inputs:
  name: inputs
  axes: byxc
  data_type: float32
  data_range: '[-inf, inf]'
  shape:
    min: '[1, 4, 4, 1]'
    step: '[1, 4, 4, 0]'
outputs:
  name: activation_26
  axes: byxc
  data_type: float32
  data_range: '[-inf, inf]'
  halo: '[0, 22, 22, 0]'
  shape:
    scale: '[1, 1, 1, 1]'
    offset: '[0, 0, 0, 0]'
training:
  source: de.csbdresden.n2v.train.N2VTraining
  kwargs: '{"means": ["39137.754"], "stds": ["18713.775"], "n_dim": 2, "axes": "YXC",
    "n_channel_in": 1, "n_channel_out": 1, "unet_residual": false, "unet_n_depth":
    2, "unet_kern_size": 3, "unet_n_first": 32, "unet_last_activation": "linear",
    "unet_input_shape": [null, null, 1], "train_loss": "mse", "train_epochs": 2, "train_steps_per_epoch":
    2, "train_learning_rate": 0.0004, "train_batch_size": 128, "train_tensorboard":
    true, "train_checkpoint": "weights_best.h5", "train_reduce_lr": {"factor": 0.5,
    "patience": 10, "verbose": true}, "batch_norm": true, "n2v_perc_pix": 1.6, "n2v_patch_shape":
    [64, 64], "n2v_manipulator": "uniform_withCP", "n2v_neighborhood_radius": 5, "probabilistic":
    false}'
prediction:
  preprocess:
    spec: de.csbdresden.n2v.predict.N2VPrediction::preprocess
    kwargs: '{ mean: [39137.754], stdDev: [18713.775]}'
  postprocess:
    spec: de.csbdresden.n2v.predict.N2VPrediction::postprocess
    kwargs: '{ mean: [39137.754], stdDev: [18713.775]}'

[tibuch]

Looks good to me. Were you able to run the exported file in Fiji?

[turekg]

Looks good to me. Were you able to run the exported file in Fiji?

I am happy to test but I have no idea where to begin. Perhaps @frauzufall can do this quickly. I've added her as a reviewer

[frauzufall]

Sorry I hope to have everything ready to test in a couple of hours. But I have more requests without testing it :)

• Can we name the created ZIP export.modelzoo.zip? This way I can associate a reader in Fiji without having to extract anything.
• Can we add source with the value n2v / denoiseg below framework? It is mentioned in the official spec, but not coherently. I want to use it to know which prediction to use automatically. It might move later somewhere else.
• you can remove spec from preprocessing and postprocessing, not sure if that is the correct entry there and we don't need it
• the source of training needs to be something different since it was trained in python. @tibuch maybe has an idea which identifier to put there.

Are we sure about the language python there? Because this zip file cannot be used e.g. to do prediction in python, right? It was exported to match the Java side. I'm unsure. What do you think @tibuch?

[turekg]

I can make these changes very quickly. I will set the source for training to be the training class in python

Sorry I hope to have everything ready to test in a couple of hours. But I have more requests without testing it :)

* Can we name the created ZIP `export.modelzoo.zip`? This way I can associate a reader in Fiji without having to extract anything.

* Can we add `source` with the value `n2v` / `denoiseg` below `framework`? It is mentioned in the official spec, but not coherently. I want to use it to know which prediction to use automatically. It might move later somewhere else.

* you can remove `spec` from `preprocessing` and `postprocessing`, not sure if that is the correct entry there and we don't need it

* the `source` of `training` needs to be something different since it was trained in python. @tibuch maybe has an idea which identifier to put there.

Are we sure about the language python there? Because this zip file cannot be used e.g. to do prediction in python, right? It was exported to match the Java side. I'm unsure. What do you think @tibuch?

[tibuch]

I would take the N2V.train() method from python as training source.

The language should be python, because it is trained in python. At least that is how I understand it. And it could be used in python as well. We just don't have the functionality implemented.

I think for now this is okay.

[turekg]

Changes committed :0)

[frauzufall]

Ok, I'm testing it now, here is what's happing:

• inputs and outputs need to be lists
• the quotes make the entries be interpreted as strings, they can't be there
• preprocess and postprocess need to be lists as well

Then it works! Here is the version that is loaded without issues:

language: python
framework: tensorflow
inputs:
  - name: inputs
    axes: byxc
    data_type: float32
    data_range: [-inf, inf]
    shape:
      min: [1, 4, 4, 1]
      step: [1, 4, 4, 0]
outputs:
  - name: activation_26
    axes: byxc
    data_type: float32
    data_range: [-inf, inf]
    halo: [0, 22, 22, 0]
    shape:
      scale: [1, 1, 1, 1]
      offset: [0, 0, 0, 0]
training:
  source: de.csbdresden.n2v.train.N2VTraining
  kwargs: {"means": ["39137.754"], "stds": ["18713.775"], "n_dim": 2, "axes": "YXC",
    "n_channel_in": 1, "n_channel_out": 1, "unet_residual": false, "unet_n_depth":
    2, "unet_kern_size": 3, "unet_n_first": 32, "unet_last_activation": "linear",
    "unet_input_shape": [null, null, 1], "train_loss": "mse", "train_epochs": 2, "train_steps_per_epoch":
    2, "train_learning_rate": 0.0004, "train_batch_size": 128, "train_tensorboard":
    true, "train_checkpoint": "weights_best.h5", "train_reduce_lr": {"factor": 0.5,
    "patience": 10, "verbose": true}, "batch_norm": true, "n2v_perc_pix": 1.6, "n2v_patch_shape":
    [64, 64], "n2v_manipulator": "uniform_withCP", "n2v_neighborhood_radius": 5, "probabilistic":
    false}
prediction:
  preprocess:
    - spec: de.csbdresden.n2v.predict.N2VPrediction::preprocess
      kwargs: { mean: [39137.754], stdDev: [18713.775]}
  postprocess:
    - spec: de.csbdresden.n2v.predict.N2VPrediction::postprocess
      kwargs: { mean: [39137.754], stdDev: [18713.775]}

Extra questions:

• the input and output names are just a placeholder or are they actually coming from the model? because we need the real names from the model tensors there.
• are min and step coming from the config or are they placeholders? the values should be different
• it still needs at least name, best case would be to generate a unique name during export, maybe just the timestamp, and a description saying that this was generated e.g. with n2v in python..
• we also need a format_version. but we need to discuss this since language and framework are probably also used by others who write their own format. so what identifies the format we are versioning?

Feels to me like we need another meeting for the details @turekg @tibuch

[frauzufall]

Thanks @turekg for already working on the other issues, we'll get there!! :sun_with_face:

[frauzufall]

Also, I asked the modelzoo people about the zip naming and they support using .bioimage.io.zip instead of their current .model.zip and our .modelzoo.zip, so that's what we should use too.

[turekg]

• @frauzufall perhaps the square brackets need to be escaped, because they have a specific meaning when a dictionary is exported: they produce a list output. Therefore the dictionary entry must be quoted. Perhaps there's some way to escape the brackets... Ah, I'll work in reverse.... Read that version in and see what it's supposed to be....
• The input name is hard coded, the output name is the value from model.keras_model.layers[-1].name
• min and step are computed as agreed: min: [1, 2^unet_n_depth, 2^unet_n_depth, n_channels_in] step: [1, 2^unet_n_depth, 2^unet_n_depth, 0]
• additional entries we can discuss
• as for the version, it really should be the schema version. So until you have a schema you won't have that value

[turekg]

Okay, I have figured it out. In order to make array values appear within brackets (ie. [1,2,3,4]) without quotes I have to set the default output style to be one long string. Otherwise,the array above would show up as -1 -2 -3 -4 which is the correct representation of an array according to the spec. But that should be irrelevant to the reader....It only makes it less legible humans :-) I will implement this fix now

[turekg]

OK, I think this is a good as it's going to get. The zip file contains two yml files in two formats. Deboh confirms she's using a library so "in principle" it should be able to read either format. The only sticking point may be the long dump of config.json. Let me know. Unfortunately the plugin is not working for me it throws an error right away. Archive.zip

[turekg]

OK! It works now! I even added the new fields for the test images. @tibuch you will do the bit that creates the images?

[tibuch]

My bad! Removed an import by accident!

[turekg]

@tibuch after you did whatever you needed to do to my branch the builds fail with a memory error

terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc
Fatal Python error: Aborted

in Travis, but they run OK on my Mac. I will try it on Jenkins.... Maybe also test with newer csbdeep?

[tibuch]

Sorry, I just rebased the branch onto master and accidentally removed an import. I have a running version on my system, but I am still working on the export.

Are you working on it as well? I could push my current version. Else I would push it once I am done with everything.

[turekg]

No go ahead push when you’re all done, but it’s not an import problem, I’ve fixed that. It’s something else. Maybe it’ll go away when you push again. gaby

On 23. Jun 2020, at 15:07, tibuch notifications@github.com wrote:

Sorry, I just rebased the branch onto master and accidentally removed an import. I have a running version on my system, but I am still working on the export.

Are you working on it as well? I could push my current version. Else I would push it once I am done with everything.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/juglab/n2v/pull/80#issuecomment-648133746, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKXJWKQXBOHUD22EKMBWHPDRYCSJFANCNFSM4N773L4Q.

[tibuch]

I changed version: 0.2.0-csbdeep to format_version: 0.2.0 to be compatible with the current java-produced yaml. I think eventually we will change it back to version: 0.2.0-csbdeep.

Additionally the test-images are now part of the output.

The structN2Vmask is exported as well, but Fiji complains if it is set to None.

[tibuch]

I forgot to update the tests after adding the changes to the exporter. Hopefully it works now.

The lzw-compression test fails on jug-srv1.

[turekg]

Make sure you have installed the extra library you need for tifffiles to read lzw compressed files. It is in setup.py

[tibuch]

hmm do I have to do this manually? I just install it with pip install -e . and then it doesn't seem to work.

[frauzufall]

@tibuch I updated the java code to set the version to 0.2.0-csbdeep

[tibuch]

How is the field called? version or format_version?

[frauzufall]

I think format_version. I'm on my phone, can't look

[frauzufall]

It's called format_version. The reference_input of the output node shape is missing:

  shape:
    reference_input: input
    scale: [1.0, 1.0, 1.0, 1.0]
    offset: [0, 0, 0, 0]

[frauzufall]

The name of the input tensor is input. The name of the output tensor is activation_11/Identity.

With these changes (adding reference_input and changes the tensor names), it works for me.

Also, min and shape don't look right... Can you verify these numbers?

[tibuch]

Make sure you have installed the extra library you need for tifffiles to read lzw compressed files. It is in setup.py

Used a python from outside of the env to run the test :facepalm:

[tibuch]

Thanks @frauzufall and @turekg.