save tile embeddings

README.md

@@ -41,7 +41,7 @@ pip install slide2vec
 
    A good starting point is the default configuration file `slide2vec/configs/default.yaml` where parameters are documented.<br>
    We've also added default configuration files for each of the foundation models currently supported:
-   - tile-level: `uni`, `uni2`, `virchow`, `virchow2`, `prov-gigapath`, `h-optimus-0`, `h-optimus-1`
+   - tile-level: `uni`, `uni2`, `virchow`, `virchow2`, `prov-gigapath`, `h-optimus-0`, `h-optimus-1`, `h0-mini`
    - slide-level: `prov-gigapath`, `titan`, `prism`
 
 

slide2vec/configs/default.yaml

@@ -44,6 +44,7 @@ model:
   batch_size: 1
   tile_size: ${tiling.params.tile_size}
   patch_size: 256 # if level is "region", size used to unroll the region into patches
+  save_tile_embeddings: false # whether to save tile embeddings alongside the pooled slide embedding when level is "slide"
 
 speed:
   fp16: false # use mixed precision during model inference

slide2vec/configs/prism.yaml

@@ -1,7 +1,6 @@
-# csv: "/data/temporary/clement/mutation-prediction/csvs/tcga-coad/slide2vec-msi.csv"
-csv: "/data/temporary/clement/mutation-prediction/csvs/tcga-ucec/slide2vec-msi.csv"
+csv: # path to csv containing slide paths
 
-output_dir: "output"
+output_dir: "output" # output directory
 visualize: true
 
 tiling:
@@ -16,16 +15,9 @@ tiling:
 model:
   level: "slide"
   name: "prism"
-  batch_size: 128
+  batch_size: 32
+  save_tile_embeddings: true # whether to save tile embeddings alongside the pooled slide embedding
 
 speed:
   fp16: true
-  num_workers_embedding: 16
-
-wandb:
-  enable: true
-  project: "mut-pred"
-  username: "clemsg"
-  exp_name: "features"
-  # tags: ["features", "tcga-coad", "${model.name}", "${model.level}", "${tiling.params.tile_size}"]
-  tags: ["features", "tcga-ucec", "${model.name}", "${model.level}", "${tiling.params.tile_size}"]
+  num_workers_embedding: 16

slide2vec/configs/prov-gigapath-slide.yaml

@@ -1,6 +1,7 @@
 csv: # path to csv containing slide paths
 
 output_dir: "output" # output directory
+visualize: true
 
 tiling:
   params:
@@ -10,7 +11,8 @@ tiling:
 model:
   level: "slide" # level at which to extract the features ("tile", "region" or "slide")
   name: "prov-gigapath"
-  batch_size: 1
+  batch_size: 32
+  save_tile_embeddings: true # whether to save tile embeddings alongside the pooled slide embedding
 
 speed:
   fp16: true # use mixed precision during model inference

slide2vec/configs/titan.yaml

@@ -1,8 +1,7 @@
-csv: "/data/temporary/clement/leopard/csvs/cologne-slide2vec.csv" # path to csv containing slide paths
+csv: # path to csv containing slide paths
 
 output_dir: "output" # output directory
-resume:
-resume_dirname:
+visualize: true
 
 tiling:
   params:
@@ -12,14 +11,8 @@ tiling:
 model:
   level: "slide" # level at which to extract the features ("tile", "region" or "slide")
   name: "titan"
-  batch_size: 1
+  batch_size: 32
+  save_tile_embeddings: true # whether to save tile embeddings alongside the pooled slide embedding
 
 speed:
-  fp16: true # use mixed precision during model inference
-
-wandb:
-  enable: true
-  project: "leopard"
-  username: "clemsg"
-  exp_name: "features"
-  tags: ["features", "cologne", "${model.name}", "${model.level}", "${tiling.params.tile_size}"]
+  fp16: true # use mixed precision during model inference

slide2vec/embed.py

@@ -219,6 +219,8 @@ def main(args):
 
                 name = wsi_fp.stem.replace(" ", "_")
                 feature_path = features_dir / f"{name}.pt"
+                if cfg.model.save_tile_embeddings:
+                    feature_path = features_dir / f"{name}-tiles.pt"
                 tmp_feature_path = tmp_dir / f"{name}-rank_{distributed.get_global_rank()}.h5"
 
                 # get feature dimension and dtype using a dry run