Syntax Test¶

This is a simple notebook for experimenting with Forgather's syntax using the low-level API

See: Low-level API

In [1]:

from forgather.config import ConfigEnvironment
from forgather.latent import Latent
from forgather.codegen import generate_code
from forgather.nb.notebooks import display_codeblock
from pprint import pp

# Construct a configuration environment
env = ConfigEnvironment()

from forgather.config import ConfigEnvironment from forgather.latent import Latent from forgather.codegen import generate_code from forgather.nb.notebooks import display_codeblock from pprint import pp # Construct a configuration environment env = ConfigEnvironment()

Define a configuration¶

In [2]:

# Define a configuration
document = """
random_tensor: !call:torch:randn [ 2, 2 ]
"""

# Convert the configuration to a graph
config = env.load_from_string(document).config
pp(config)

# Define a configuration document = """ random_tensor: !call:torch:randn [ 2, 2 ] """ # Convert the configuration to a graph config = env.load_from_string(document).config pp(config)

{'random_tensor': SingletonNode('torch:randn', *(2, 2), identity=139888396855632, **{})}

Materialize the "random_tensor" target¶

In [3]:

Latent.materialize(config, mtargets="random_tensor")

Latent.materialize(config, mtargets="random_tensor")

Out[3]:

tensor([[ 0.7128, -0.5204],
        [ 1.4749,  0.8098]])

Convert a configuraiton into executable Python code¶

Note that generate_code() takes an optional Jinja2 template as an arguemnt, which can be used to customize the output.

In [4]:

code = generate_code(config)
display_codeblock("python", code)

code = generate_code(config) display_codeblock("python", code)

from torch import randn

def construct(
):
    
    
    return {
        'random_tensor': randn(
            2,
            2,
        ),
    }

Run generated code¶

In [5]:

exec(code)
construct()

exec(code) construct()

Out[5]:

{'random_tensor': tensor([[-0.7962,  0.5893],
         [ 0.8036, -0.2863]])}

In [ ]: