moskomule / homura

homura is a library for fast prototyping DL research

About!

I'm Ryuichiro Hataya, PhD, a postdoctoral researcher at RIKEN AIP (High-Dimensional Causal Analysis Team).

News

• Our papers "Transformers as Stochastic Optimizers" (Hataya & Imaizumi) and "Automatic Domain Adaptation by Transformers in In-Context Learning" (Hataya & Matsui & Imaizumi) have been accepted at the ICML 2024 Workshop on In-Context Learning.
• I will visit CENTAI @ Torino🇮🇹.
• I will visit Wien🇦🇹 to attend ICML.
• I will visit Aix-Marseille University @ Marseille🇫🇷.
• I will visit A*STAR CFAR @ Singapore🇸🇬.
• Our paper "Self-attention Networks Localize When QK-eigenspectrum Concentrates" is accepted at ICML.
• I will attend FIMI @ Bristol🇬🇧.
• I will attend DLWS @ Tokyo🇯🇵.
• I will attend MLSS @ Okinawa🇯🇵 as an organizer.
• Our paper “Sketch-based Semantic Retrieval of Medical Images” (Kobayashi et al.) has been accepted at Medical Image Analysis.
• I will attend QTML @ Geneve🇨🇭.
• I will visit Nicolaus Copernicus University Poland @ Torun🇵🇱 from Sep 25th to 30th and then attend ICCV @ Paris🇫🇷.
• I will visit the Okinawa Institute of Science and Technology🇯🇵 from Aug 29th to Sep 1st.
• Our paper "An Empirical Investigation of Pre-trained Model Selection for Out-of-Distribution Generalization and Calibration" is accepted at ICCV 2023 Workshop on Uncertainty Quantification for Computer Vision.
• Our paper "Will Large-scale Generative Models Corrupt Future Datasets?" is accepted at ICCV 2023.
• Our paper "Towards AI-driven radiology education: A self-supervised segmentation-based framework for high-precision medical image editing" is accepted at MICCAI 2023.
• I will attend AISTATS @ València🇪🇸 at the end of April and visit IIT @ Genova🇮🇹 at the beginning of May.
• I will visit the Vietnam Institute for Advanced Study in Mathematics @ Ha Noi🇻🇳 on April 2023.
• I will visit EPFL CIS @ Lausanne🇨🇭 and Fraunhofer IIS @ Nürnberg🇩🇪 from 8th to 15th March 2023.
• Our paper "Nyström Method for Accurate and Scalable Implicit Differentiation" is accepted at AISTATS 2023.
• I received a doctoral degree as a representative student of the Graduate School of Information Science and Technology, UTokyo.
• I visted at IIT @ Genova on July 2022.
• I successfully defended my PhD thesis!
• Our paper "DJMix: Unsupervised Task-agnostic Image Augmentation for Improving Robustness of Convolutional Neural Networks" is accepted at IJCNN 2022.
• I am promoted to Excellent Research Assistant @ UTokyo.
• Our paper "Meta Approach to Data Augmentation Optimization" has been accepted to WACV 2022.

Stats

homura's Issues

dataset DataSet(root, train, transform, download) is expected, but <class 'torchvision.datasets.cifar.CIFAR10'> has arguments of set() instead.

Traceback (most recent call last):
File "cifar.py", line 3, in
from homura import callbacks, lr_scheduler, optim, reporters
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/init.py", line 8, in
Registry.import_modules('homura.vision')
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/register.py", line 97, in import_modules
module = importlib.import_module(package_name)
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/importlib/init.py", line 127, in import_module
return _bootstrap._gcd_import(name[level:], package, level)
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/vision/init.py", line 2, in
from .data import DATASET_REGISTRY
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/vision/data/init.py", line 2, in
from .datasets import VisionSet, DATASET_REGISTRY
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/vision/data/datasets.py", line 230, in
{'cifar10': VisionSet(datasets.CIFAR10, "~/.torch/data/cifar10", 10,
File "", line 9, in init
File "/home/x/anaconda3/envs/senetv38/lib/python3.8/site-packages/homura/vision/data/datasets.py", line 86, in post_init
raise RuntimeError(f"dataset DataSet(root, train, transform, download) is expected, "
RuntimeError: dataset DataSet(root, train, transform, download) is expected, but <class 'torchvision.datasets.cifar.CIFAR10'> has arguments of set() instead.

ImportError: cannot import name 'CosineAnnealingWarmRestarts'

Hello, as I import homura, I got this issue
'''

import homura
Traceback (most recent call last):
File "", line 1, in
File "/home/wangwen/anaconda3/envs/py36/lib/python3.6/site-packages/homura/init.py", line 3, in
from . import callbacks, debug, liblog, lr_scheduler, optim, reporters, trainers, metrics, modules, utils, vision
File "/home/wangwen/anaconda3/envs/py36/lib/python3.6/site-packages/homura/lr_scheduler.py", line 5, in
from torch.optim.lr_scheduler import (StepLR as _StepLR, MultiStepLR as _MultiStepLR, LambdaLR as _LambdaLR,
ImportError: cannot import name 'CosineAnnealingWarmRestarts'
'''

About the pip install homura under conda env

Hi I install through pip install -U git+https://github.com/moskomule/homura, and update setuptools in conda environment, but faced the issue as below:

(py36) [ychan@server]$ pip install -U git+https://github.com/moskomule/homura
Collecting git+https://github.com/moskomule/homura
Cloning https://github.com/moskomule/homura to /tmp/pip-req-build-lyxuuvoc
Running command git clone -q https://github.com/moskomule/homura /tmp/pip-req-build-lyxuuvoc
ERROR: Command errored out with exit status 1:
command: /home/csgrad/ychan/anaconda3/envs/py36/bin/python -c 'import sys, setuptools, tokenize; sys.argv[0] = '"'"'/tmp/pip-req-build-lyxuuvoc/setup.py'"'"'; file='"'"'/tmp/pip-req-build-lyxuuvoc/setup.py'"'"';f=getattr(tokenize, '"'"'open'"'"', open)(file);code=f.read().replace('"'"'\r\n'"'"', '"'"'\n'"'"');f.close();exec(compile(code, file, '"'"'exec'"'"'))' egg_info --egg-base /tmp/pip-pip-egg-info-sm_mm7_d
cwd: /tmp/pip-req-build-lyxuuvoc/
Complete output (7 lines):
Traceback (most recent call last):
File "", line 1, in
File "/tmp/pip-req-build-lyxuuvoc/setup.py", line 4, in
readme = f.read()
File "/home/csgrad/ychan/anaconda3/envs/py36/lib/python3.6/encodings/ascii.py", line 26, in decode
return codecs.ascii_decode(input, self.errors)[0]
UnicodeDecodeError: 'ascii' codec can't decode byte 0xf0 in position 388: ordinal not in range(128)
----------------------------------------
WARNING: Discarding git+https://github.com/moskomule/homura. Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.
ERROR: Command errored out with exit status 1: python setup.py egg_info Check the logs for full command output.

Do anyone know how to fix it? Thanks

`trainer.resume` does not work as expected

• fails when DDP? https://github.com/moskomule/senet.pytorch/issues/41
• does not load scheduler etc.

Profiler support

Support PyTorch's profiler. See

https://www.mattari-benkyo-note.com/2021/04/03/pytorch-profiler/

WIN10, pip install git+https://github.com/moskomule/homura UnicodeDecodeError: 'gbk' codec can't decode byte

pip install git+https://github.com/moskomule/homura
Looking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple
Collecting git+https://github.com/moskomule/homura
Cloning https://github.com/moskomule/homura to c:\users\tenglong\appdata\local\temp\pip-req-build-cv6aijof
Complete output from command python setup.py egg_info:
Traceback (most recent call last):
File "", line 1, in
File "C:\Users\TengLong\AppData\Local\Temp\pip-req-build-cv6aijof\setup.py", line 4, in
readme = f.read()
UnicodeDecodeError: 'gbk' codec can't decode byte 0xb0 in position 271: illegal multibyte sequence

How can I record the accuracy of the model?

I would like to ask you how to record the accuracy of model training and test when the model is trained。

thanks！

The latest update broke tqdm

At every iteration, a new line is inserted and breaks everything.

the function of imagenet_loaders don't exist in homura.vision.data

when i use imagenet.py in senet.pytorch, but i find the function of imagenet_loaders don't exist in homura.vision.data.

How to assign GPU id?

I want to train cifar10 with two gpus. How to assign the training process with specific two gpus?

callbacks/reporter APIs

complicated

0.2

• Use ABCMeta for abstract classes
• Use decorator for a simpler interface.
• Add simple tests as much as possible
• Documentation

How to use mixup for training

    tqdm_rep = reporters.TQDMReporter(range(args.epochs), callbacks=[callbacks.AccuracyCallback(1)])
       trainer = Trainer(model, optimizer, F.cross_entropy, scheduler=scheduler, callbacks=[tqdm_rep])
       for _ in tqdm_rep:
            trainer.train(train_loader)
            trainer.test(test_loader)

How to use mixup augmentation methods for training.

ValueError: not enough values to unpack (expected 2, got 1)

hello,
when i call trainer.train（customdataloader），i get error following:
File "cifar.py", line 40, in | 0/11143 [00:00<?, ?it/s]
main()
File "cifar.py", line 27, in main
trainer.train(trainloader)
File "/home/lab205/anaconda3/envs/senet/lib/python3.7/site-packages/homura/trainers.py", line 236, in train
self._loop(data_loader, mode=mode)
File "/home/lab205/anaconda3/envs/senet/lib/python3.7/site-packages/homura/trainers.py", line 215, in _loop
self._iteration(data, mode)
File "/home/lab205/anaconda3/envs/senet/lib/python3.7/site-packages/homura/trainers.py", line 182, in _iteration
results = self.iteration(data)
File "/home/lab205/anaconda3/envs/senet/lib/python3.7/site-packages/homura/trainers.py", line 336, in iteration
input, target = data
ValueError: not enough values to unpack (expected 2, got 1)

but i am sure that
(for data in customdataloader) the lenth is 2

How to uninstall this module?

I want to uninstall this module, how to do it?

More simple APIs

Callbacks and reporters are not simple enough.

Trainer.iteration should be something like

class SupervisedTrainer(TrainerBase):
...
def iteration(self, data):
    input, target = data
    output = self.model(input)
    loss = self.loss_f(output, target)
    if self.is_train:
        self.optimizer.zero_grad()
        loss.backward()
        self.optimizer.ste()
    self.report('loss', loss)
    self.report('accuracy', homura.accuracy(output, target))

Then, a simple training code is like

def main(cfg):
    model = MODEL_REGISTRY...
    dataset = DATASET_REGISTRY...
    train_loader, test_loader = dataset(...)
    optimizer = ...
    trainer = Trainer(model, optimizer, loss_f)
    trainer.add_callbacks(...)  # reporters other than TQDMReporter
    for ep in trainer.epoch_iters:
        trainer.train(...)

DDP yileds looower accuracy

DDP yileds looower accuracy. Maybe MetricCallback.reduce is wrong.

How to resolve the exception: AssertionError: root does not exist

Traceback (most recent call last):
File "/home/fxz/gitRepos/senet.pytorch/cifar.py", line 39, in
main()
File "/home/fxz/gitRepos/senet.pytorch/cifar.py", line 11, in main
train_loader, test_loader = DATASET_REGISTRY("cifar10")(args.batch_size, num_workers=args.num_workers)
File "/home/fxz/.local/lib/python3.8/site-packages/homura/vision/data/datasets.py", line 237, in get_dataloader
train_set, test_set, val_set = self.get_dataset(train_size, test_size, val_size,
File "/home/fxz/.local/lib/python3.8/site-packages/homura/vision/data/datasets.py", line 145, in get_dataset
assert (download or self.root.exists()), "root does not exist"
AssertionError: root does not exist

Easy Trainer override

Add add_end_epoch(self, name, value) etc. Then overrideing iteration, loop will be easier.

def iteration(self, data: Iterable[torch.Tensor], is_train: bool) -> Iterable[torch.Tensor]:
        input, target = self.to_device(data)
        self.add_start_epoch("input", input)
        output = self.model(input)
        loss = self.loss_f(output, target)
        if is_train:
            self.optimizer.zero_grad()
            loss.backward()
            self.optimizer.step()
        return loss, output

TypeError: `scheduler.func` is expected to be subclass of `_LRScheduler` but got <class 'type'>

when i added cifar-10 datasets to the ~/.torch/data. then i run the cifar.py ,it remind me this error:

D:\Installation\ANACONDA\envs\senet\python.exe E:\study\AI_acc\senet.pytorch-master\cifar.py
Files already downloaded and verified
Files already downloaded and verified
0%| | 0/200 [00:00<?, ?it/s]Traceback (most recent call last):
File "E:\study\AI_acc\senet.pytorch-master\cifar.py", line 38, in
main()
File "E:\study\AI_acc\senet.pytorch-master\cifar.py", line 22, in main
with Trainer(model, optimizer, F.cross_entropy, scheduler=scheduler, callbacks=_callbacks) as trainer:
File "D:\Installation\ANACONDA\envs\senet\lib\site-packages\homura\trainers.py", line 519, in init
super(SupervisedTrainer, self).init(model, optimizer, loss_f, callbacks=callbacks, scheduler=scheduler,
File "D:\Installation\ANACONDA\envs\senet\lib\site-packages\homura\trainers.py", line 122, in init
self.set_scheduler()
File "D:\Installation\ANACONDA\envs\senet\lib\site-packages\homura\trainers.py", line 454, in set_scheduler
raise TypeError(f"scheduler.func is expected to be subclass of _LRScheduler"
TypeError: scheduler.func is expected to be subclass of _LRScheduler but got <class 'type'>
0%| | 0/200 [00:00<?, ?it/s]

Process finished with exit code 1

ImportError: cannot import name 'ConcatTransform' from 'homura.vision.transforms'

I cannot find ConcatTransform ,TransformBase from the lib，

Use pdoc?

https://pdoc3.github.io/pdoc/

GPU memory leak?

While DDP training, GPU memory consumption gradually and slowly grows.

Parameter scheduler

PyTorch only offers lr_scheduler.

How to install the old version of homura

I wonder to know how to install the old version of homura, thx.

why callbacks can't import from homura?

Feature requests (ToDo)

• (Categorical Conditionl Batchnorm see ( https://github.com/pfnet-research/sngan_projection/blob/master/source/links/conditional_batch_normalization.py ))
• Enable to plot matplotlib objects on Tensorboard and Visdom

Update attention/transformers

Give a ERROR

When I use the package, it give a error about ImportError: cannot import name 'callbacks' from 'homura'.Why?

ImportError: cannot import name 'callbacks' from partially initialized module 'homura' , I already have homura installed

raise KeyError( KeyError: 'Name resnet20 is already used, try another name!'

run cifar.py always remind me this error:
raise KeyError( KeyError: 'Name resnet20 is already used, try another name!'
it seems like error in register.py

`
def register_from_dict(self,
name_to_func: Dict[str, Callable]):
for k, v in name_to_func.items():
self.register(v, name=k)

def register(self,
             func: Callable = None,
             *,
             name: Optional[str] = None):
    if func is None:
        return functools.partial(self.register, name=name)

    _type = self.type
    if _type is not None and not isinstance(_type, types.FunctionType):
        if not (isinstance(func, _type) or issubclass(func, _type)):
            raise TypeError(
                f'`func` is expected to be subclass of {_type}.')

    if name is None:
        name = func.__name__

    if self._registry.get(name) is not None:
        raise KeyError(
            f'Name {name} is already used, try another name!')

`