Saw an open issue about improving outputs so I took a stab at it. Didn't want to do too much as I saw you may already have some updates in mind regarding the file name / directory structure.

Summary of changes:

• Pixray will now take an optional parameter "--debug": a boolean value that indicates whether or not to output a debug log with the final output.
• The debug log currently includes the settings used to generate an image. (More can be added later).
• A reusable file path calculation function in utility class.
• Add some unit tests.

Summary of changes:

• Added --overlay_until option.
  • Takes an integer argument that is the number of iterations.
  • Default value is None.

Tests can be run by executing python -m unittest tests/test_pixray.py from the main pixray directory.

I had to lower num_cuts when running.

    prompts="warrior. concept art. trending on artstation",
    drawer="super_resolution",
    size=[512, 512],
    num_cuts=8,
    quality="normal",
    learning_rate=0.1,
    init_image="human.jpg",

I keep getting this error...

ImportError: cannot import name 'SimpleTokenizer' from 'tokenizer' (C:\Users\micro\anaconda3\envs\pixel\lib\site-packages\tokenizer_init_.py)

Summary of changes:

• Added a new parse_unit function that parses strings with units ("20 iterations", "50%", etc) to a raw iteration integer.
• Refactored how parameters with pipes are handled slightly.
• Overlay related arguments now use strings with units specified rather than integers.
• Added test cases.

Haven't had time to hook up the other arguments yet but I can continue with that.

Hey! I like Pixray, it can make some cool art. (better then me D:) However, on my Chromebook at school, it cannot load the Input or Output areas without a proxy. I'm using the Lightspeed blocker if that helps.

Hello! It looks like the model for vdiff isn't available at the URL currently in Pixray anymore. Attempting to use the vdiff drawer gives me this:

(base) [email protected]:~/pixray$ python pixray.py --drawer=vdiff --prompts="an excessively fuzzy panda" --outdir panda
Running with 30x1 = 30 cuts
Using seed: 7387649654636579532
Downloading models/yfcc_2.pth from https://v-diffusion.s3.us-west-2.amazonaws.com/yfcc_2.pth, please wait
--2022-05-04 13:39:59--  https://v-diffusion.s3.us-west-2.amazonaws.com/yfcc_2.pth
Resolving v-diffusion.s3.us-west-2.amazonaws.com (v-diffusion.s3.us-west-2.amazonaws.com)... 52.218.228.113
Connecting to v-diffusion.s3.us-west-2.amazonaws.com (v-diffusion.s3.us-west-2.amazonaws.com)|52.218.228.113|:443... connected.
HTTP request sent, awaiting response... 404 Not Found
2022-05-04 13:39:59 ERROR 404: Not Found.

Ignoring non-zero exit:  b''
Traceback (most recent call last):
  File "/home/fox/pixray/pixray.py", line 2135, in <module>
    main()
  File "/home/fox/pixray/pixray.py", line 2129, in main
    do_init(settings)
  File "/home/fox/pixray/pixray.py", line 613, in do_init
    drawer.load_model(args, device)
  File "/home/fox/pixray/vdiff.py", line 85, in load_model
    model.load_state_dict(torch.load(checkpoint, map_location='cpu'))
  File "/home/fox/miniconda3/lib/python3.9/site-packages/torch/serialization.py", line 608, in load
    return _legacy_load(opened_file, map_location, pickle_module, **pickle_load_args)
  File "/home/fox/miniconda3/lib/python3.9/site-packages/torch/serialization.py", line 777, in _legacy_load
    magic_number = pickle_module.load(f, **pickle_load_args)
EOFError: Ran out of input

And spits out a 0 byte vfcc_2.pth file in the models folder. Attempting to follow the link it gives also yields a 404 page saying the bucket no longer exists.

Hey @dribnet 👋🏼

This PR updates pixray's Cog predictor classes to be compatible with Cog's new Python API.

Cog v0.1.0 has a new predictor API that makes use of Python's built-in type annotations to declare input and output types. The new API also has a different way of declaring inputs based on pydantic, a python package for validating Python models. Instead of using the @cog.input decorators, inputs are now declared inline as parameters to the predict() method.

There's a bunch of other useful foundational stuff in this new release of Cog that gets us closer to having a standardize type system that leans on JSON Schema and OpenAPI instead of re-inventing our own thing. For more details, see the release notes here: https://github.com/replicate/cog/releases/tag/v0.1.0

Process

Here's the process I followed to set things up, make changes, and test:

1. created a new model https://replicate.com/zeke/pydantic-pixray
2. forked this repo, recursively cloned it, and cog pushed it to zeke/pydantic-pixray using "old cog" (0.0.x)
3. verified that my unchanged fork of pixray worked by running some predictions
4. upgraded my local Cog version to the latest, 0.1.1
5. updated all the Predictors in cogrun.py to use cog.BasePredictor, cog.Input, cog.Path etc.
6. set the image field in cog.yaml to publish to my own copy of the model and the predict field to cogrun.py:PixrayVdiff
7. published using cog push
8. ran cog predict from a GCP image with a GPU. See https://gist.github.com/zeke/a36c059bebb751fb21b26c1d14ed1996

Progress

I am now able to cog build, cog push, and cog predict the changes herein using the latest version of Cog, but still hitting a few snags:

• The PixrayVdiff predictor (which I think corresponds to https://replicate.com/pixray/text2image) produces output, but it's yielding the same image over and over. See https://gist.github.com/zeke/a36c059bebb751fb21b26c1d14ed1996
• Some of the existing predictors accept a kwargs argument, but the new version of Cog has a strict list of allowed input types. In order to be compatible with Cog's new type checking stuff, these predictors that accept arbitrary keyword arguments will need to be expanded to explicitly list all the arguments and their types.

Next steps

@dribnet hopefully this gives you a head start for updating pixray to work with the new version of Cog. Let me know if this all makes sense, and if you need more help getting these changes shipped.

@dribnet Can you please check and merge this feature?

It is meant to take a new argument: --config-file <path_to_yaml> without using run.pyscript. (Sorry for delays, this is my first PR on a repo I forked ;))

See ya! And, btw, great job on pixray...

Only supports "rectangular pixels".

250it [00:31,  7.82it/s]
vs.
235it [02:22,  1.73it/s]

Also uses a little less memory 6433MiB vs 8747MiB

This is a little Docker gotchya -- at some point the apt repositories will change, then apt-get install will fail because it's using the old cached output of update.

We should make it harder to trip up on this in Cog, or document it, or something...

ERROR: Could not find a version that satisfies the requirement torch==1.9.0+cu102 (from versions: 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.10.1, 1.10.2, 1.11.0, 1.12.0, 1.12.1, 1.13.0)

Python 3.8.5 is the oldest version available on miniconda.

(3.8) [email protected] pixray % pip install -r requirements.txt
Looking in links: https://download.pytorch.org/whl/torch_stable.html
Collecting git+https://github.com/bfirsh/[email protected] (from -r requirements.txt (line 28))
  Cloning https://github.com/bfirsh/taming-transformers.git (to revision 7a6e64ee) to /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-zaqc64ss
  Running command git clone --filter=blob:none --quiet https://github.com/bfirsh/taming-transformers.git /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-zaqc64ss
  WARNING: Did not find branch or tag '7a6e64ee', assuming revision or ref.
  Running command git checkout -q 7a6e64ee
  Resolved https://github.com/bfirsh/taming-transformers.git to commit 7a6e64ee
  Preparing metadata (setup.py) ... done
Collecting git+https://github.com/openai/CLIP (from -r requirements.txt (line 29))
  Cloning https://github.com/openai/CLIP to /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-wdkc5c31
  Running command git clone --filter=blob:none --quiet https://github.com/openai/CLIP /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-wdkc5c31
  Resolved https://github.com/openai/CLIP to commit d50d76daa670286dd6cacf3bcd80b5e4823fc8e1
  Preparing metadata (setup.py) ... done
Collecting git+https://github.com/pvigier/[email protected] (from -r requirements.txt (line 30))
  Cloning https://github.com/pvigier/perlin-numpy (to revision 6f077f8) to /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-iyuy_gin
  Running command git clone --filter=blob:none --quiet https://github.com/pvigier/perlin-numpy /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-iyuy_gin
  WARNING: Did not find branch or tag '6f077f8', assuming revision or ref.
  Running command git checkout -q 6f077f8
  Resolved https://github.com/pvigier/perlin-numpy to commit 6f077f8
  Preparing metadata (setup.py) ... done
Collecting git+https://github.com/fbcotter/pytorch_wavelets (from -r requirements.txt (line 46))
  Cloning https://github.com/fbcotter/pytorch_wavelets to /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-a0yl7grc
  Running command git clone --filter=blob:none --quiet https://github.com/fbcotter/pytorch_wavelets /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-a0yl7grc
  Resolved https://github.com/fbcotter/pytorch_wavelets to commit 9a0c507f04f43c5397e384bb6be8340169b2fd9a
  Preparing metadata (setup.py) ... done
Collecting git+https://github.com/pixray/[email protected] (from -r requirements.txt (line 49))
  Cloning https://github.com/pixray/aphantasia (to revision 7e6b3bb) to /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-jbsijre8
  Running command git clone --filter=blob:none --quiet https://github.com/pixray/aphantasia /private/var/folders/01/sn57hs8566145w17svn1c9780000gn/T/pip-req-build-jbsijre8
  WARNING: Did not find branch or tag '7e6b3bb', assuming revision or ref.
  Running command git checkout -q 7e6b3bb
  Resolved https://github.com/pixray/aphantasia to commit 7e6b3bb
  Preparing metadata (setup.py) ... done
ERROR: Could not find a version that satisfies the requirement torch==1.9.0+cu102 (from versions: 1.8.1, 1.9.0, 1.9.1, 1.10.0, 1.10.1, 1.10.2, 1.11.0, 1.12.0, 1.12.1, 1.13.0)
ERROR: No matching distribution found for torch==1.9.0+cu102
(3.8) [email protected] pixray % python --version
Python 3.8.5

I am trying to install using the requirements.txt. However, I am getting this error:

Collecting package metadata (current_repodata.json): ...working... done Solving environment: ...working... failed with initial frozen solve. Retrying with flexible solve. Collecting package metadata (repodata.json): ...working... done Solving environment: ...working... failed with initial frozen solve. Retrying with flexible solve.

PackagesNotFoundError: The following packages are not available from current channels:

• torch==1.9.0+cu102

Current channels:

• https://repo.anaconda.com/pkgs/main/win-64
• https://repo.anaconda.com/pkgs/main/noarch
• https://repo.anaconda.com/pkgs/r/win-64
• https://repo.anaconda.com/pkgs/r/noarch
• https://repo.anaconda.com/pkgs/msys2/win-64
• https://repo.anaconda.com/pkgs/msys2/noarch

To search for alternate channels that may provide the conda package you're looking for, navigate to

https://anaconda.org

My computer is windows 10, and created a Python 3.8 virtual env using Conda. Here is my Cuda settings:

import torch
import tensorflow as tf
import tensorflow.keras as ks

print(tf)
print(ks)
print(torch.cuda.is_available())
print(torch.version.cuda)
print(torch.backends.cudnn.version())
print('//////////')

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
X_train = torch.FloatTensor([0., 1., 2.])
X_train = X_train.to(device)


print(X_train.is_cuda)
print(torch.cuda.current_device())
print(torch.cuda.device_count())
print(torch.cuda.get_device_name(0))

Output

<module 'tensorflow' from 'C:\Venv\conda_python3_8_tensorflow_gen_art\lib\site-packages\tensorflow\init.py'> <module 'tensorflow.keras' from 'C:\Venv\conda_python3_8_tensorflow_gen_art\lib\site-packages\tensorflow\keras\init.py'> True 11.3 8302 ////////// True 0 1 NVIDIA GeForce GTX 1070 Ti

Process finished with exit code 0

The file download utility has been changed from a linux utility to a python library. It downloads at a good speed without too much RAM or CPU usage. It requires a new import, unfortunately, but it is a common one.

This is tested to be working on my local Windows 10 machine and a Google Colab instance.

Fixes #47 Partial for #76