Error when working with Norman pretrained model

[igrabski]

Thanks for making this tool available!

I am trying to run the tutorial located here, but have been encountering some issues. When I run the following code:

from gears import PertData, GEARS
from gears.utils import dataverse_download
from zipfile import ZipFile

dataverse_download('https://dataverse.harvard.edu/api/access/datafile/6979957', 'norman_umi_go.tar.gz')

## Extract and set up dataloader directory
import tarfile
with tarfile.open('norman_umi_go.tar.gz', 'r:gz') as tar:
    tar.extractall()

dataverse_download('https://dataverse.harvard.edu/api/access/datafile/6979956', 'model.zip')

## Extract and set up model directory
with ZipFile(('model.zip'), 'r') as zip:
    zip.extractall(path = './')

data_path = './'
data_name = 'norman_umi_go'
model_name = 'gears_misc_umi_no_test'

pert_data = PertData(data_path)
pert_data.load(data_path = data_path + data_name)
pert_data.prepare_split(split = 'no_test', seed = 1)
pert_data.get_dataloader(batch_size = 32, test_batch_size = 128)

gears_model = GEARS(pert_data, device = 'cpu',
                        weight_bias_track = False,
                        proj_name = 'gears',
                        exp_name = model_name)

I get the error below:

AttributeError: 'Series' object has no attribute 'nonzero'

I was able to resolve the error by running:

pert_data.adata.X = pert_data.adata.X.toarray()

and the gears_model command works. However, when I then run:

gears_model.load_pretrained('./model_ckpt')

I get this error:

TypeError: GEARS.model_initialize() got an unexpected keyword argument 'cell_fitness_pred'

Please let me know how I can resolve this -- thanks so much!

[yuzzini]

Same error here.

[longliveJESSIE]

I met with the same first problem too. It seems the problem is caused by the data type of sparse matrix. The second problem did not happen to me with the 0.1.1 version of gears. I manually switched to df09d7a version. Maybe you can also try that version.

[cchu70]

I got a different error at gears_model.load_pretrained('./model_ckpt') in the tutorial_inference_Norman.ipynb notebook. It seems that the pert_data only has 5045 genes, while the pretrained model is expecting 5054 (last two digits swapped). Did others run into this, too?

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[11], line 5
      1 gears_model = GEARS(pert_data, device = 'cpu', 
      2                         weight_bias_track = False, 
      3                         proj_name = 'gears', 
      4                         exp_name = model_name)
----> 5 gears_model.load_pretrained(f'{output_dir}/model_ckpt')

File ~/projects/perturb_protein/src/GEARS/gears/gears.py:272, in GEARS.load_pretrained(self, path)
    269         new_state_dict[name] = v
    270     state_dict = new_state_dict
--> 272 self.model.load_state_dict(state_dict)
    273 self.model = self.model.to(self.device)
    274 self.best_model = self.model

File /data/ClaudiaC/envs/GEARS_env/lib/python3.10/site-packages/torch/nn/modules/module.py:2593, in Module.load_state_dict(self, state_dict, strict, assign)
   2585         error_msgs.insert(
   2586             0,
   2587             "Missing key(s) in state_dict: {}. ".format(
   2588                 ", ".join(f'"{k}"' for k in missing_keys)
   2589             ),
   2590         )
   2592 if len(error_msgs) > 0:
-> 2593     raise RuntimeError(
   2594         "Error(s) in loading state_dict for {}:\n\t{}".format(
   2595             self.__class__.__name__, "\n\t".join(error_msgs)
   2596         )
   2597     )
   2598 return _IncompatibleKeys(missing_keys, unexpected_keys)

RuntimeError: Error(s) in loading state_dict for GEARS_Model:
	size mismatch for indv_w1: copying a param with shape torch.Size([5054, 64, 1]) from checkpoint, the shape in current model is torch.Size([5045, 64, 1]).
	size mismatch for indv_b1: copying a param with shape torch.Size([5054, 1]) from checkpoint, the shape in current model is torch.Size([5045, 1]).
	size mismatch for indv_w2: copying a param with shape torch.Size([1, 5054, 65]) from checkpoint, the shape in current model is torch.Size([1, 5045, 65]).
	size mismatch for indv_b2: copying a param with shape torch.Size([1, 5054]) from checkpoint, the shape in current model is torch.Size([1, 5045]).
	size mismatch for gene_emb.weight: copying a param with shape torch.Size([5054, 64]) from checkpoint, the shape in current model is torch.Size([5045, 64]).
	size mismatch for emb_pos.weight: copying a param with shape torch.Size([5054, 64]) from checkpoint, the shape in current model is torch.Size([5045, 64]).
	size mismatch for cross_gene_state.network.0.weight: copying a param with shape torch.Size([64, 5054]) from checkpoint, the shape in current model is torch.Size([64, 5045]).