Fixed COCO-Stuff splits

Fixed Dataset for FSL step (filter not active yet)
Fixed Word-Embeddings for SPNet in COCO

coco.sh

@@ -1,28 +1,43 @@
 #!/bin/bash
 
 export CUDA_VISIBLE_DEVICES=$1
-port=$2
-alias exp="python -m torch.distributed.launch --master_port ${port} --nproc_per_node=1 run.py --opt_level O0"
+alias exp="python -m torch.distributed.launch --master_port $2 --nproc_per_node=1 run.py --num_workers 8"
 shopt -s expand_aliases
 
-task=7mc
-ds=coco
+gen_par="--dataset coco --epochs 20 --batch_size 24 --crop_size 512 --val_interval 1"
+#gen_par="--task ${task} --dataset voc --epochs 30 --batch_size 24 --crop_size 512 --val_interval 1"
+#gen_par="--task voc --dataset coco-stuff --lr 0.01 --epochs 20 --batch_size 24 --crop_size 512 --val_interval 1"
 
-path=checkpoints/step/spn-${ds}
+lr=0.01
 
-gen_par="--task ${task} --dataset ${ds} --batch_size 10 --crop_size 512"
+exp --method SPN --name SPN ${gen_par} --step 0 --lr ${lr} --task 20-0
+exp --method SPN --name SPN ${gen_par} --step 0 --lr ${lr} --task 20-1
+exp --method SPN --name SPN ${gen_par} --step 0 --lr ${lr} --task 20-2
 
-lr=0.0001 # for 1-shot, 15-1
-iter=200
 
-for is in 0 1 2 3 4; do
-inc_par="--ishot ${is} --input_mix novel --val_interval 1000 --ckpt_interval 5 --no_pooling"
-for ns in 1; do
-  exp --method AMP --name AMP --iter 0 --lr ${lr} ${gen_par} ${inc_par} --step 1 --nshot ${ns} --step_ckpt ${path}/FT_0.pth
-  exp --method FT --name FT --iter ${iter} --lr ${lr} ${gen_par} ${inc_par} --step 1 --nshot ${ns} --step_ckpt ${path}/FT_0.pth
-    for step in 2 3 4 5 6 7; do
-        exp --method AMP --name AMP --iter 0 --lr ${lr} ${gen_par} ${inc_par} --step ${step} --nshot ${ns}
-        exp --method FT --name FT --iter ${iter} --lr ${lr} ${gen_par} ${inc_par} --step ${step} --nshot ${ns}
-    done
-  done
-done
+#export CUDA_VISIBLE_DEVICES=$1
+#port=$2
+#alias exp="python -m torch.distributed.launch --master_port ${port} --nproc_per_node=1 run.py --opt_level O0"
+#shopt -s expand_aliases
+#
+#task=7mc
+#ds=coco
+#
+#path=checkpoints/step/spn-${ds}
+#
+#gen_par="--task ${task} --dataset ${ds} --batch_size 10 --crop_size 512"
+#
+#lr=0.0001 # for 1-shot, 15-1
+#iter=200
+#
+#for is in 0 1 2 3 4; do
+#inc_par="--ishot ${is} --input_mix novel --val_interval 1000 --ckpt_interval 5 --no_pooling"
+#for ns in 1; do
+#  exp --method AMP --name AMP --iter 0 --lr ${lr} ${gen_par} ${inc_par} --step 1 --nshot ${ns} --step_ckpt ${path}/FT_0.pth
+#  exp --method FT --name FT --iter ${iter} --lr ${lr} ${gen_par} ${inc_par} --step 1 --nshot ${ns} --step_ckpt ${path}/FT_0.pth
+#    for step in 2 3 4 5 6 7; do
+#        exp --method AMP --name AMP --iter 0 --lr ${lr} ${gen_par} ${inc_par} --step ${step} --nshot ${ns}
+#        exp --method FT --name FT --iter ${iter} --lr ${lr} ${gen_par} ${inc_par} --step ${step} --nshot ${ns}
+#    done
+#  done
+#done

data/coco-stuff/make_annotation.py

@@ -0,0 +1,30 @@
+from PIL import Image
+import numpy as np
+import os
+import os.path as osp
+from tqdm import tqdm
+
+orig_path = '/dataset/vandal/COCO'  # insert here the source path where original COCO annotations are
+dest_path = '.'  # the destination folder, which should be this one'
+
+for split in ["train2017", "val2017"]:
+    annotations = f"{orig_path}/annotations/{split}"
+    nov_ann = f"{dest_path}/annotations/{split}"
+
+    # clear folder if exists
+    if osp.exists(nov_ann):
+        print("Removing existing")
+        os.rmdir(nov_ann)
+    os.makedirs(nov_ann)
+
+    # remap labels in the novel interval (+1 for Stuff, +1 and stuff on 0 for objects)
+    mapping = np.zeros((256,), dtype=np.int8)
+    for i, cl in enumerate(range(91)):
+        mapping[cl] = i + 1
+    mapping[255] = 255
+    target_transform = lambda x: Image.fromarray(mapping[x])
+
+    for f in tqdm(os.listdir(annotations)):
+        lbl = Image.open(osp.join(annotations, f))
+        lbl = target_transform(lbl)
+        lbl.save(osp.join(nov_ann, f))