Support for Apple silicon using MPS #6

batch-dl+direct.sh

@@ -72,6 +72,8 @@ export MODEL_ARGS
 export JOB_QUEUE=job_queue.txt
 [[ -f ${JOB_QUEUE} ]] && rm ${JOB_QUEUE}
 
+TAIL=tail
+[[ "`uname -s`" == "Darwin" ]] && TAIL=gtail
 
 run_dl() {
 	SUBJ=$1
@@ -127,5 +129,5 @@ true > ${JOB_QUEUE}
 # create first N_PARALLEL_CPU dummy entries. Otherwise jobs will only start once N_PARALLEL_CPU jobs are queued
 for i in `seq 1 ${N_PARALLEL_CPU}` ; do echo dummy >> ${JOB_QUEUE} ; done
 
-tail -n+0 -f ${JOB_QUEUE} --pid ${PID_DL} | parallel -j ${N_PARALLEL_CPU} run_direct {}
+${TAIL} -n+0 -f ${JOB_QUEUE} --pid ${PID_DL} | parallel -j ${N_PARALLEL_CPU} run_direct {}
 

dl+direct.sh

@@ -91,7 +91,7 @@ echo
 # convert into freesurfer space (resample to 1mm voxel, orient to LIA)
 python ${SCRIPT_DIR}/conform.py "${T1}" "${DST}/T1w_norm.nii.gz"
 
-HAS_GPU=`python -c 'import torch; print(torch.cuda.is_available())'`
+HAS_GPU=`python -c 'import torch; print(torch.cuda.is_available() or (hasattr(torch.backends, "mps") and torch.backends.mps.is_available()))'`
 if [ ${HAS_GPU} != 'True' ] ; then
 	echo "WARNING: No GPU/CUDA device found. Running on CPU might take some time..."
 fi
@@ -110,6 +110,7 @@ if [ ${DO_SKULLSTRIP} -gt 0 ] ; then
 	IN_VOLUME=${DST}/T1w_norm_noskull.nii.gz
 	BET_INPUT_VOLUME=${DST}/T1w_norm.nii.gz
 	MASK_VOLUME=${DST}/T1w_norm_noskull_mask.nii.gz
+	export PYTORCH_ENABLE_MPS_FALLBACK=1
 
 	python ${SCRIPT_DIR}/bet.py ${BET_OPTS} "${BET_INPUT_VOLUME}" "${IN_VOLUME}" || die "hd-bet failed"
 else

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "DL-DiReCT"
-version = "1.0.1"
+version = "1.0.2"
 description = "DL+DiReCT - Direct Cortical Thickness Estimation using Deep Learning-based Anatomy Segmentation and Cortex Parcellation"
 readme = "README.md"
 authors = [ {name = "Michael Rebsamen"} ]
@@ -20,7 +20,7 @@ classifiers = [
 
 dependencies = [
 	"antspyx>=0.3.5",
-	"HD_BET @ https://github.com/MIC-DKFZ/HD-BET/archive/refs/heads/master.zip",
+	"HD_BET @ https://github.com/mrunibe/HD-BET/archive/refs/heads/master.zip",
 	"nibabel>=3.2.1",
 	"numpy<2.0.0",
 	"pandas>=0.25.3",

src/DeepSCAN_Anatomy_Newnet_apply.py

@@ -399,7 +399,7 @@ def load_checkpoint(checkpoint_file, device):
         sys.exit(1)
 
     print('loading checkpoint {}'.format(checkpoint_file)) if VERBOSE else False
-    return torch.load(checkpoint_file, map_location=device)
+    return torch.load(checkpoint_file, weights_only=True, map_location=device)
 
 
 def validate_input(t1, t1_data):
@@ -435,7 +435,7 @@ def validate_input(t1, t1_data):
     if not os.path.exists(output_dir):
         os.makedirs(output_dir)
 
-    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    device = torch.device('cuda' if torch.cuda.is_available() else ('mps' if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available() else 'cpu'))
     checkpoint = load_checkpoint(model_file, device)
     target_label_names = checkpoint['label_names']
     # number of last labels to ignore for hard segmentation (argmax), e.g. left-hemi, right-hemi, brain

src/bet.py

@@ -46,7 +46,9 @@
 
     print('Brain extraction using HD-BET [https://doi.org/10.1002/hbm.24750] ...')
 
-    if not torch.cuda.is_available():
+    if hasattr(torch.backends, 'mps') and torch.backends.mps.is_available():
+	    hdbet_device = 'mps'
+    elif not torch.cuda.is_available():
         print('No GPU found. Running hd-bet in fast mode, check results! Make sure you have enough memory.')
         hdbet_mode = 'fast'
         hdbet_device = 'cpu'