thinkingmachines/christmais

Text to abstract art generation for the holidays!

christmais's Issues

Write shell script

Write shell script for:

• Parsing (christmais.scripts.parse)
• Style Transfer (christmais.scripts.style_transfer)

Use them for usage instructions as required in #55

Add Dockerfile

We might be using GCP to speed up processing (not yet sure). Easier to set everything up via Dockerfile

Implement Trainer

Hi @ibtingzon , I'll just write the workflow here so you can have a basis too (for easier syncs):

How to get the objective function

1. Initialize n number of Artists, store them in a list called artists (lol)
2. All of them should generate an image (of type PIL.Image) using draw(). We can use a list comprehension here
3. One PIL.Image has a shape of (224, 224, 3). If you initialized n number of Artists, we should stack them into a tensor of shape (n, 224, 224, 3). (Optional since predict() can use PIL.Image directly)
4. Initialize a Predictor class and call predict() on this tensor. Not sure if the current implementation does batch compute, if not, fix Predictor. If successful, it should return a list of shape (n,).
5. The elements of the list are the "fitness" of each Artist. Optimize using the techniques that we have.

Optimization Step

1. In the list artists, we can get the gene by calling get_gene(). This should give you a matrix for each artist (as far as I remember its shape (10, 26). Reshape them if necessary (np.ravel(), etc. etc.).
2. Once we get the "new genes", call draw_from_gene() for each Artist again in the list artists.
3. Repeat step 2 of "How to get the objective function"

Sample

# Generate 10 artists
artists = [Artist(seed, (224,224)) for i in range(10)]
# Draw an image for each artist
imgs = [artist.draw() for artist in artists]
# Compute fitness
target = 'iron'
fitness = [p.predict(img, target=target)[0] for img in imgs]

Update trainer-naive-workflow for a more descriptive walkthrough

Fix make dev

Add apt-get installation for make dev and make build.

Add tests for drawer

Requires testing output artifacts

Add option to adjust transparency for each layer

Write API documentation

Write a pure magenta.js port?

Add option to just load the trained embedding

Add option to control the number of lines drawn

Add Usage to README

Expose the locations of circles and lines in an image

Add tests for DrawingSystem

Add usage instructions

For now, just write usage instructions on how to run the shell script.

Remove JS dependency in drawer

We're just sourcing these dependencies from a CDN, so there's really no need to implement that. You can remove those and just have a single drawer, similar to parser and styler instead.

Write notes on what worked and what did not work

Add option to override circles per layer

Fix christmais_time