search

nitzanlab / biolord

https://biolord.readthedocs.io
BSD 3-Clause "New" or "Revised" License
67 stars 4 forks source link

several errors in trial runs #6

Closed shobhitagrawal1 closed 1 year ago

shobhitagrawal1 commented 1 year ago

Report

hi, really interesting method, congrats. Trying to run the demo running into several problems. the parameters commented out below throw errors at subsequent steps -- for exampl decoder_width is an unrecognised parameter for model etc. After commenting the parameters which were causing a problem : Support for training_epoch_end has been removed in v2.0.0. biolordTrainingPlan implements this method. You can use the on_train_epoch_end hook instead. To access outputs, save them in-memory as instance attributes. You can find migration examples in https://github.com/Lightning-AI/lightning/pull/16520

Would be grateful for any assistance :)

module_params = {

"decoder_width": 1024,

#"decoder_depth": 4,
"attribute_nn_width": 512,
"attribute_nn_depth": 2,
"n_latent_attribute_categorical": 4,
"loss_ae": "gauss",
"reconstruction_penalty": 1e2,
"unknown_attribute_penalty": 1e1,
"unknown_attribute_noise_param": 1e-1,
"attribute_dropout_rate": 0.1,
"use_batch_norm": False,
"use_layer_norm": False,
"seed": 42,

}

model = biolord.Biolord( adata=adata, n_latent=32, model_name="spatio_temporal_infected", module_params=module_params, train_classifiers=False, split_key="split_random", )

trainer_params = { "n_epochs_warmup": 0,

"latent_lr": 1e-4,

"latent_wd": 1e-4,

"decoder_lr": 1e-4,

"decoder_wd": 1e-4,

"attribute_nn_lr": 1e-2,
"attribute_nn_wd": 4e-8,
"step_size_lr": 45,
"cosine_scheduler": True,
"scheduler_final_lr": 1e-5,}

model.train( max_epochs=500, batch_size=512, plan_kwargs=trainer_params, early_stopping=True, early_stopping_patience=20, check_val_every_n_epoch=10, num_workers=1, enable_checkpointing=False, )

Version information

anndata 0.9.1 biolord 0.0.1 matplotlib 3.7.2 numpy 1.24.4 pandas 2.0.3 scanpy 1.9.3 scipy 1.11.1 scvi 1.0.2 seaborn 0.12.2 session_info 1.0.0

PIL 9.4.0 absl NA aiohttp 3.8.4 aiosignal 1.3.1 anyio NA async_timeout 4.0.2 attr 23.1.0 backoff 2.2.1 brotli NA bs4 4.12.2 certifi 2023.05.07 cffi 1.15.1 charset_normalizer 2.0.4 chex 0.1.7 click 8.1.4 contextlib2 NA croniter NA cycler 0.10.0 cython_runtime NA dateutil 2.8.2 deepdiff 6.3.1 docrep 0.3.2 etils 1.3.0 fastapi 0.100.0 flax 0.7.0 frozenlist 1.3.3 fsspec 2023.6.0 gmpy2 2.1.2 h5py 3.9.0 idna 3.4 importlib_metadata NA importlib_resources NA jax 0.4.13 jaxlib 0.4.13 joblib 1.3.1 kiwisolver 1.4.4 lightning 2.0.5 lightning_cloud NA lightning_fabric 2.0.5 lightning_utilities 0.9.0 llvmlite 0.40.1 ml_collections NA ml_dtypes 0.2.0 mpl_toolkits NA mpmath 1.2.1 msgpack 1.0.5 mudata 0.2.3 multidict 6.0.4 multipart 0.0.6 multipledispatch 0.6.0 natsort 8.4.0 numba 0.57.1 numpyro 0.12.1 nvfuser NA opt_einsum v3.3.0 optax 0.1.5 ordered_set 4.1.0 packaging 23.1 patsy 0.5.3 pkg_resources NA psutil 5.9.5 pydantic 1.10.11 pygments 2.15.1 pyparsing 3.0.9 pyro 1.8.5 pytorch_lightning 2.0.5 pytz 2023.3 requests 2.29.0 rich NA setuptools 67.8.0 six 1.16.0 sklearn 1.3.0 sniffio 1.3.0 socks 1.7.1 soupsieve 2.4.1 sparse 0.14.0 starlette 0.27.0 statsmodels 0.14.0 sympy 1.11.1 threadpoolctl 3.1.0 toolz 0.12.0 torch 2.0.1 torchaudio 2.0.2 torchmetrics 1.0.0 torchvision 0.15.2 tqdm 4.65.0 tree 0.1.8 typing_extensions NA urllib3 1.26.16 uvicorn 0.22.0 websocket 1.6.1 websockets 11.0.3 xarray 2023.6.0 yaml 6.0 yarl 1.9.2 zipp NA zoneinfo NA

Python 3.9.17 (main, Jul 5 2023, 20:41:20) [GCC 11.2.0] Linux-4.18.0-305.12.1.el8_4.x86_64-x86_64-with-glibc2.31

zoepiran commented 1 year ago

Hi,

I am sorry for the inconvenience - if you git clone the repo the framework can be run without errors. we will soon make a new release which will also include updated tutorials!

Please excuse the inconvenience, Zoe

shobhitagrawal1 commented 1 year ago

thanks for replying will try your recommendation. Looking forward to your new tutorials!

shobhitagrawal1 commented 1 year ago

Hi Zoe, everything ran fine, except that the parameter "loss ae" was not found, which I commented out for now. It seems it is not mentioned in the _module.py, init.. - i did not delve deeper (whether by default it uses MSE or Gaussian)... thanks !

zoepiran commented 1 year ago

i am sorry; to match current scvi we changed "loss_ae"-> "gene_likelihood" (if tour input is normalized pass "normal", and if raw counts use "nb" )