Dev

config/ablation_study.yml

@@ -20,8 +20,8 @@ trainer:
     - class_path: lightning.pytorch.loggers.WandbLogger
       init_args:
         project: ${project}
-        save_dir: /lustre/fswork/projects/rech/xeg/uat95fg/ #/pasteur/zeus/projets/p02/ml4ig_hot/Users/jkalfon/ #/data/log/
-        offline: False
+        save_dir: /pasteur/zeus/projets/p02/ml4ig_hot/Users/jkalfon/ #/lustre/fswork/projects/rech/xeg/uat95fg/ #/pasteur/zeus/projets/p02/ml4ig_hot/Users/jkalfon/ #/data/log/
+        offline: True
   callbacks:
     #- class_path: lightning.pytorch.callbacks.StochasticWeightAveraging
     #  init_args:
@@ -53,6 +53,7 @@ model:
   prenorm: True
   fused_mlp: False
   fused_bias_fc: False
+  depth_atinput: False
   drop_path_rate: 0
   freeze_embeddings: True
   normalization: log
@@ -63,7 +64,7 @@ data:
     - NCBITaxon:9606
     - NCBITaxon:10090
   gene_position_tolerance: 10_000
-  gene_embeddings: /lustre/fswork/projects/rech/xeg/uat95fg/gene_embeddings.parquet
+  gene_embeddings: ./data/main/gene_embeddings.parquet # /lustre/fswork/projects/rech/xeg/uat95fg/gene_embeddings.parquet
   collection_name: scPRINT-V2 full #scPRINT-V2 (good quality)
   how: random expr
   max_len: 2200

config/base_v2.yml

@@ -8,7 +8,7 @@ log_graph: True
 trainer:
   precision: 16-mixed
   strategy: ddp_find_unused_parameters_true
-  gradient_clip_val: 40
+  gradient_clip_val: 10
   log_every_n_steps: 100
   limit_train_batches: 20000
   gradient_clip_algorithm: norm
@@ -33,7 +33,7 @@ trainer:
         save_last: True
 scprint_training:
   run_full_forward: True
-  noise: [0.6]
+  noise: [0.9]
   do_ecs: True
   do_cce: True
   mask_ratio: ["TF"]
@@ -42,9 +42,10 @@ scprint_training:
 scprint_early_stopping:
   patience: 4
 model:
-  dropout: 0.1
-  num_heads_kv: 2
+  dropout: 0
+  num_heads_kv: 4
   transformer: flash
+  finetune_gene_emb: True
   mvc_decoder: inner product
   residual_in_fp32: True
   checkpointing: True
@@ -68,6 +69,7 @@ model:
   #  self_reported_ethnicity_ontology_term_id: 10
   #  sex_ontology_term_id: 2
   #  organism_ontology_term_id: 8
+  class_compression: "none"
 data:
   organisms:
     - NCBITaxon:9606
@@ -76,10 +78,10 @@ data:
   gene_embeddings: ./data/main/gene_embeddings.parquet
   collection_name: scPRINT-V2 full #scPRINT-V2 (good quality) # scPRINT-V2 (some)
   how: random expr
-  max_len: 2200
+  max_len: 3200
   pin_memory: True
   prefetch_factor: 3
-  metacell_mode: 0.2
+  metacell_mode: 0
   weight_scaler: 200
   do_gene_pos: ./data/main/biomart_pos.parquet
   add_zero_genes: 0

docs/notebooks/cancer_usecase.ipynb

@@ -35,6 +35,9 @@
     "5. [Denoising and differential expression](#denoising-and-differential-expression)\n",
     "6. [Gene network inference](#gene-network-inference)\n",
     "\n",
+    "Please note that although your adata needs to be processed in a similar way to what is shown in `1.` You can perform the analaysis in `2-3-4`, `5` or `6` in any order as they are completely distinct.\n",
+    "\n",
+    "\n",
     "> In the notebook [cancer_usecase_part2.ipynb](./cancer_usecase_part2.ipynb) you will see how to analyse cell type specific gene regulatory networks.\n",
     "\n"
    ]