Decorators for reducing pytorch boilerplate
Project description
What is this?
Functions and decorators I found myself rewriting for every pytorch project
How do I use this?
pip install trivial-torch-tools
from trivial_torch_tools import Sequential, init
import torch.nn as nn
class Model(nn.Module):
@init.to_device()
# ^ does self.to() and defaults to GPU if available (uses default_device variable)
@init.save_and_load_methods(model_attributes=["layers"], basic_attributes=["input_shape"])
# ^ creates self.save(path=self.path) and self.load(path=self.path)
def __init__(self):
self.input_shape = (81,81,3)
layers = Sequential(input_shape=self.input_shape)
# ^ dynamically compute the output shape/size of layers (the nn.Linear below)
layers.add_module('conv1' , nn.Conv2d(input_shape[0], 32, kernel_size=8, stride=4, padding=0))
layers.add_module('relu1' , nn.ReLU())
layers.add_module('flatten' , nn.Flatten(start_dim=1, end_dim=-1))
layers.add_module('linear1' , nn.Linear(in_features=layers.output_size, out_features=10))
layers.add_module('sigmoid1', nn.Sigmoid())
self.layers = layers
# layers.output_size
# layers.output_shape
# layers.layer_shapes
# available tools
from trivial_torch_tools import *
core.default_device # defaults to cuda if available
core.to_tensor(nested_lists_of_arrays_tuples_and_more) # aggresively converts objects to tensors
# decorators for def __init__()
@model.init.to_device(device=default_device)
@model.init.save_and_load_methods(basic_attributes=[], model_attributes=[], path_attribute="path")
@model.init.forward_sequential_method
# decorators for def forward(): # or whatever
@model.convert_each_arg.to_tensor() # Use to_tensor(which_args=[0]) to only convert first arg
@model.convert_each_arg.to_device() # Use to_device(which_args=[0]) to only convert first arg
@model.convert_each_arg.to_batched_tensor(number_of_dimensions=4) # 4 works for color images
@model.convert_each_arg.torch_tensor_from_opencv_format()
image.tensor_from_path(path)
image.pil_image_from_tensor(tensor)
image.torch_tensor_from_opencv_format(tensor_or_array)
image.opencv_tensor_from_torch_format(tensor)
image.opencv_array_from_pil_image(image_obj)
OneHotifier.tensor_from_argmax(tensor) # [0.1,99,0,0,] => [0,1,0,0,]
OneHotifier.index_from_one_hot(tensor) # [0,1,0,0,] => 2
OneHotifier.index_tensor_from_onehot_batch(tensor) # [[0,1,0,0,]] => [2]
import torch
converter = OneHotifier(possible_values=[ "thing0", ('thing', 1), {"thing":2} ])
converter.to_one_hot({"thing":2}) # >>> tensor([0,0,1])
converter.from_one_hot(torch.tensor([0,0,1])) # >>> {"thing":2}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Close
Hashes for trivial_torch_tools-0.6.0.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | 3fcb69df534cca32e50ea2c7404ac5226b743bd75750c78cc8e2056d12ce1658 |
|
MD5 | 17db46a1737e55775e7696c38cace114 |
|
BLAKE2b-256 | 2a9011ce89c667bdc4cffc9351a942aac49082ac42912fd88d1bce753f50ba22 |
Close
Hashes for trivial_torch_tools-0.6.0-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 577b2334fcd170bbc32a9703fb6cf051aa6d690f32096d52ce0c2a397f1b760c |
|
MD5 | b46df71d04153647204f9f26b0cdec51 |
|
BLAKE2b-256 | 92ffe0cfcc12e0330df2a316f97a911dca2b99c98314e62fea4c500a7fffca50 |