OutOfMemoryError when running run_tensor_cell2cell_pipeline on GPU

[joseph-siefert]

The memory at each iteration seems to accumulate. Is it possible to release it or chuck it differently to circumvent this issue?

Running Elbow Analysis
 24%|██▍       | 6/25 [23:50<1:15:30, 238.44s/it]
---------------------------------------------------------------------------
OutOfMemoryError                          Traceback (most recent call last)
Cell In[10], line 1
----> 1 tensor2 = c2c.analysis.run_tensor_cell2cell_pipeline(tensor,
      2                                                      meta_tensor,
      3                                                      copy_tensor=True
      4                                                      rank=None, 
      5                                                      tf_optimization='robust',
      6                                                      random_state=0, 
      7                                                      device='cuda', 
      8                                                      elbow_metric='error', 
      9                                                      smooth_elbow=False, 
     10                                                      upper_rank=25, 
     11                                                      tf_init='random', 
     12                                                      tf_svd='numpy_svd',
     13                                                      cmaps=None, 
     14                                                      sample_col='Element', 
     15                                                      group_col='Category', 
     16                                                      fig_fontsize=14, 
     17                                                      output_folder=output_folder, 
     18                                                      output_fig=True, 
     19                                                      fig_format='pdf', 
     20                                                     )

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/cell2cell/analysis/tensor_pipelines.py:172, in run_tensor_cell2cell_pipeline(interaction_tensor, tensor_metadata, copy_tensor, rank, tf_optimization, random_state, backend, device, elbow_metric, smooth_elbow, upper_rank, tf_init, tf_svd, cmaps, sample_col, group_col, fig_fontsize, output_folder, output_fig, fig_format, **kwargs)
    170 if rank is None:
    171     print('Running Elbow Analysis')
--> 172     fig1, error = interaction_tensor.elbow_rank_selection(upper_rank=upper_rank,
    173                                                           runs=elbow_runs,
    174                                                           init=tf_init,
    175                                                           svd=tf_svd,
    176                                                           automatic_elbow=True,
    177                                                           metric=elbow_metric,
    178                                                           output_fig=output_fig,
    179                                                           smooth=smooth_elbow,
    180                                                           random_state=random_state,
    181                                                           fontsize=fig_fontsize,
    182                                                           filename=elbow_filename,
    183                                                           tol=tol, n_iter_max=n_iter_max,
    184                                                           **kwargs
    185                                                           )
    187     rank = interaction_tensor.rank
    189 # Factorization

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/cell2cell/tensor/tensor.py:518, in BaseTensor.elbow_rank_selection(self, upper_rank, runs, tf_type, init, svd, metric, random_state, n_iter_max, tol, automatic_elbow, manual_elbow, smooth, mask, ci, figsize, fontsize, filename, output_fig, verbose, **kwargs)
    516         fig = None
    517 elif runs > 1:
--> 518     all_loss = _multiple_runs_elbow_analysis(tensor=self.tensor,
    519                                              upper_rank=upper_rank,
    520                                              runs=runs,
    521                                              tf_type=tf_type,
    522                                              init=init,
    523                                              svd=svd,
    524                                              metric=metric,
    525                                              random_state=random_state,
    526                                              mask=mask,
    527                                              n_iter_max=n_iter_max,
    528                                              tol=tol,
    529                                              verbose=verbose,
    530                                              **kwargs
    531                                              )
    533     # Same outputs as runs = 1
    534     loss = np.nanmean(all_loss, axis=0).tolist()

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/cell2cell/tensor/factorization.py:342, in _multiple_runs_elbow_analysis(tensor, upper_rank, runs, tf_type, init, svd, metric, random_state, mask, n_iter_max, tol, verbose, **kwargs)
    340 else:
    341     rs = None
--> 342 tl_object, errors = _compute_tensor_factorization(tensor=tensor,
    343                                                   rank=r,
    344                                                   tf_type=tf_type,
    345                                                   init=init,
    346                                                   svd=svd,
    347                                                   random_state=rs,
    348                                                   mask=mask,
    349                                                   n_iter_max=n_iter_max,
    350                                                   tol=tol,
    351                                                   verbose=verbose,
    352                                                   **kwargs)
    354 if metric == 'error':
    355     # This helps to obtain proper error when the mask is not None.
    356     if mask is None:

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/cell2cell/tensor/factorization.py:92, in _compute_tensor_factorization(tensor, rank, tf_type, init, svd, random_state, mask, n_iter_max, tol, verbose, **kwargs)
     89         return_errors = kwargs['return_errors']
     91 if tf_type == 'non_negative_cp':
---> 92     cp_tf = non_negative_parafac(tensor=tensor,
     93                                  rank=rank,
     94                                  init='random' if mask is not None else init,
     95                                  svd=svd,
     96                                  random_state=random_state,
     97                                  mask=mask,
     98                                  n_iter_max=n_iter_max,
     99                                  tol=tol,
    100                                  verbose=verbose,
    101                                  **kwargs)
    103     if return_errors:
    104         cp_tf, errors = cp_tf

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/decomposition/_nn_cp.py:126, in non_negative_parafac(tensor, rank, n_iter_max, init, svd, tol, random_state, verbose, normalize_factors, return_errors, mask, cvg_criterion, fixed_modes)
    124 accum = tl.reshape(weights, (-1, 1)) * accum * tl.reshape(weights, (1, -1))
    125 if mask is not None:
--> 126     tensor = tensor * mask + tl.cp_to_tensor(
    127         (weights, factors), mask=1 - mask
    128     )
    130 mttkrp = unfolding_dot_khatri_rao(tensor, (weights, factors), mode)
    132 numerator = tl.clip(mttkrp, a_min=epsilon, a_max=None)

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/cp_tensor.py:483, in cp_to_tensor(cp_tensor, mask)
    478     full_tensor = T.dot(
    479         factors[0] * weights, T.transpose(khatri_rao(factors, skip_matrix=0))
    480     )
    481 else:
    482     full_tensor = T.sum(
--> 483         khatri_rao([factors[0] * weights] + factors[1:], mask=mask), axis=1
    484     )
    486 return fold(full_tensor, 0, shape)

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/backend/__init__.py:206, in BackendManager.dispatch_backend_method.<locals>.wrapped_backend_method(*args, **kwargs)
    202 def wrapped_backend_method(*args, **kwargs):
    203     """A dynamically dispatched method
    204 
    205     Returns the queried method from the currently set backend"""
--> 206     return getattr(
    207         cls._THREAD_LOCAL_DATA.__dict__.get("backend", cls._backend), name
    208     )(*args, **kwargs)

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/tenalg/core_tenalg/_khatri_rao.py:101, in khatri_rao(matrices, weights, skip_matrix, reverse, mask)
     98     matrices = matrices[::-1]
     99     # Note: we do NOT use .reverse() which would reverse matrices even outside this function
--> 101 return T.kr(matrices, weights=weights, mask=mask)

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/backend/__init__.py:206, in BackendManager.dispatch_backend_method.<locals>.wrapped_backend_method(*args, **kwargs)
    202 def wrapped_backend_method(*args, **kwargs):
    203     """A dynamically dispatched method
    204 
    205     Returns the queried method from the currently set backend"""
--> 206     return getattr(
    207         cls._THREAD_LOCAL_DATA.__dict__.get("backend", cls._backend), name
    208     )(*args, **kwargs)

File /mambaforge/envs/scanpy/lib/python3.9/site-packages/tensorly/backend/core.py:1097, in Backend.kr(self, matrices, weights, mask)
   1093     res = self.reshape(a * b, (-1, n_col))
   1095 m = self.reshape(mask, (-1, 1)) if mask is not None else 1
-> 1097 return res * m

OutOfMemoryError: CUDA out of memory. Tried to allocate 4.20 GiB (GPU 0; 15.90 GiB total capacity; 7.36 GiB already allocated; 3.66 GiB free; 11.41 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

[earmingol]

What size is your tensor? this depends a lot on the size of the tensor because when unfolding it to performs underlying calculations can demand a lot of memory

[joseph-siefert]

I've experimented with different sizes, but still have the same issue. For example, a tensor of (42, 2628, 27, 27), which is roughly half the size I need, still runs out of memory. There is plenty of CPU memory available, so if it can be stored on CPU between iterations that would help. Seems to progressively accumulate each iteration. Otherwise, if it can be chunked differently for the GPU processes.

[earmingol]

Can you try running the decomposition with a given number of factors (e.g. 10 factors) and see if you still get this error?
If so, the issue is the memory that you have in your GPU, which from what I see is 16GB.

In my opinion, the issue is the size of your tensor and the limited memory in your GPU. I would recommend prioritizing ligand-receptor pairs (e.g. by signaling pathway, expression level, among other options) to reduce the number of interactions from 2,628 to ~200-500 pairs.