2024 update

.gitignore

@@ -1,5 +1,6 @@
 data
 work_dirs
+.conda
 
 # Created by https://www.toptal.com/developers/gitignore/api/python
 # Edit at https://www.toptal.com/developers/gitignore?templates=python

configs/_base_/datasets/fungi_bs16_swin_384_class-balanced.py

@@ -62,7 +62,7 @@ train_dataloader = dict(
         oversample_thr=1e-2,
         dataset=dict(
             type=dataset_type,
-            data_root='data/fungi2023/',
+            data_root='data/fungi2024/',
             ann_file='FungiCLEF2023_train_metadata_PRODUCTION.csv',
             data_prefix='DF20/',
             pipeline=train_pipeline)),
@@ -74,7 +74,7 @@ val_dataloader = dict(
     num_workers=8,
     dataset=dict(
         type=dataset_type,
-        data_root='data/fungi2023/',
+        data_root='data/fungi2024/',
         ann_file='FungiCLEF2023_val_metadata_PRODUCTION.csv',
         data_prefix='DF21/',
         pipeline=test_pipeline),
@@ -87,7 +87,7 @@ test_dataloader = dict(
     num_workers=8,
     dataset=dict(
         type=dataset_type,
-        data_root='data/fungi2023/',
+        data_root='data/fungi2024/',
         ann_file='FungiCLEF2023_val_metadata_PRODUCTION.csv',
         data_prefix='DF21/',
         pipeline=test_pipeline),

configs/_base_/default_runtime.py

@@ -31,7 +31,7 @@ env_cfg = dict(
     mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
 
     # set distributed parameters
-    dist_cfg=dict(backend='nccl'),
+    dist_cfg=dict(backend='nccl', timeout=1800),
 )
 
 # set visualizer

configs/fungi2023/swinv2_base_w24_b16x4-fp16_fungi+val_res_384_cb_epochs_6.py

@@ -0,0 +1,45 @@
+_base_ = [
+    '../_base_/models/swin_transformer_v2/base_384_aug.py', '../_base_/datasets/fungi_bs16_swin_384_class-balanced.py',
+    '../_base_/schedules/fungi_bs64_adamw_swin.py', '../_base_/default_runtime.py'
+]
+
+# model settings
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-base-w12_3rdparty_in21k-192px_20220803-f7dc9763.pth'  # noqa
+model = dict(
+    backbone=dict(
+        window_size=[24, 24, 24, 12],
+        pretrained_window_sizes=[12, 12, 12, 6],
+        init_cfg=dict(
+            type='Pretrained',
+            checkpoint=checkpoint,
+            prefix='backbone',
+        )),
+    head=dict(num_classes=1604),
+    train_cfg=dict(_delete_=True),
+)
+
+train_dataloader = dict(
+    batch_size=16,
+    num_workers=14)
+
+val_dataloader = dict(
+    batch_size=16,
+    num_workers=12)
+
+train_cfg = dict(max_epochs=6)
+
+optim_wrapper = dict(type='AmpOptimWrapper', optimizer=dict(lr=2.5e-4 * 64 / 512))
+
+# learning policy
+param_scheduler = [
+    # warm up learning rate scheduler
+    dict(
+        type='LinearLR',
+        start_factor=0.01,
+        by_epoch=False,
+        end=4200),
+    # main learning rate scheduler
+    dict(type='CosineAnnealingLR', eta_min=0, by_epoch=False, begin=4200)
+]
+
+custom_imports = dict(imports=['mmpretrain_custom'], allow_failed_imports=False)

configs/fungi2023/swinv2_base_w24_b16x8-fp16_fungi+val_res_384_cb_epochs_6.py

@@ -0,0 +1,45 @@
+_base_ = [
+    '../_base_/models/swin_transformer_v2/base_384_aug.py', '../_base_/datasets/fungi_bs16_swin_384_class-balanced.py',
+    '../_base_/schedules/fungi_bs64_adamw_swin.py', '../_base_/default_runtime.py'
+]
+
+# model settings
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-base-w12_3rdparty_in21k-192px_20220803-f7dc9763.pth'  # noqa
+model = dict(
+    backbone=dict(
+        window_size=[24, 24, 24, 12],
+        pretrained_window_sizes=[12, 12, 12, 6],
+        init_cfg=dict(
+            type='Pretrained',
+            checkpoint=checkpoint,
+            prefix='backbone',
+        )),
+    head=dict(num_classes=1604),
+    train_cfg=dict(_delete_=True),
+)
+
+train_dataloader = dict(
+    batch_size=16,
+    num_workers=14)
+
+val_dataloader = dict(
+    batch_size=16,
+    num_workers=12)
+
+train_cfg = dict(max_epochs=6)
+
+optim_wrapper = dict(type='AmpOptimWrapper')
+
+# learning policy
+param_scheduler = [
+    # warm up learning rate scheduler
+    dict(
+        type='LinearLR',
+        start_factor=0.01,
+        by_epoch=False,
+        end=4200),
+    # main learning rate scheduler
+    dict(type='CosineAnnealingLR', eta_min=0, by_epoch=False, begin=4200)
+]
+
+custom_imports = dict(imports=['mmpretrain_custom'], allow_failed_imports=False)

configs/fungi2023/swinv2_base_w24_b8x8-fp16_fungi+val_res_384_cb_epochs_6.py

@@ -0,0 +1,45 @@
+_base_ = [
+    '../_base_/models/swin_transformer_v2/base_384_aug.py', '../_base_/datasets/fungi_bs16_swin_384_class-balanced.py',
+    '../_base_/schedules/fungi_bs64_adamw_swin.py', '../_base_/default_runtime.py'
+]
+
+# model settings
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-base-w12_3rdparty_in21k-192px_20220803-f7dc9763.pth'  # noqa
+model = dict(
+    backbone=dict(
+        window_size=[24, 24, 24, 12],
+        pretrained_window_sizes=[12, 12, 12, 6],
+        init_cfg=dict(
+            type='Pretrained',
+            checkpoint=checkpoint,
+            prefix='backbone',
+        )),
+    head=dict(num_classes=1604),
+    train_cfg=dict(_delete_=True),
+)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=10)
+
+val_dataloader = dict(
+    batch_size=16,
+    num_workers=10)
+
+train_cfg = dict(max_epochs=6)
+
+optim_wrapper = dict(type='AmpOptimWrapper', optimizer=dict(lr=2.5e-4 * 64 / 512))
+
+# learning policy
+param_scheduler = [
+    # warm up learning rate scheduler
+    dict(
+        type='LinearLR',
+        start_factor=0.01,
+        by_epoch=False,
+        end=4200),
+    # main learning rate scheduler
+    dict(type='CosineAnnealingLR', eta_min=0, by_epoch=False, begin=4200)
+]
+
+custom_imports = dict(imports=['mmpretrain_custom'], allow_failed_imports=False)

configs/fungi2023/swinv2_large_w24_b8x8-fp16_fungi+val_res_384_cb_epochs_6.py

@@ -0,0 +1,45 @@
+_base_ = [
+    '../_base_/models/swin_transformer_v2/large_384_aug.py', '../_base_/datasets/fungi_bs16_swin_384_class-balanced.py',
+    '../_base_/schedules/fungi_bs64_adamw_swin.py', '../_base_/default_runtime.py'
+]
+
+# model settings
+checkpoint = 'https://download.openmmlab.com/mmclassification/v0/swin-v2/pretrain/swinv2-large-w12_3rdparty_in21k-192px_20220803-d9073fee.pth'  # noqa
+model = dict(
+    backbone=dict(
+        window_size=[24, 24, 24, 12],
+        pretrained_window_sizes=[12, 12, 12, 6],
+        init_cfg=dict(
+            type='Pretrained',
+            checkpoint=checkpoint,
+            prefix='backbone',
+        )),
+    head=dict(num_classes=1604),
+    train_cfg=dict(_delete_=True),
+)
+
+train_dataloader = dict(
+    batch_size=8,
+    num_workers=10)
+
+val_dataloader = dict(
+    batch_size=32,
+    num_workers=10)
+
+train_cfg = dict(max_epochs=6)
+
+optim_wrapper = dict(type='AmpOptimWrapper', optimizer=dict(lr=2.5e-4 * 64 / 512))
+
+# learning policy
+param_scheduler = [
+    # warm up learning rate scheduler
+    dict(
+        type='LinearLR',
+        start_factor=0.01,
+        by_epoch=False,
+        end=4200),
+    # main learning rate scheduler
+    dict(type='CosineAnnealingLR', eta_min=0, by_epoch=False, begin=4200)
+]
+
+custom_imports = dict(imports=['mmpretrain_custom'], allow_failed_imports=False)

launch.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+#nvidia-smi
+#echo CUDA_VISIBLE_DEVICES=$CUDA_VISIBLE_DEVICES
+unset CUDA_VISIBLE_DEVICES
+export KMP_AFFINITY=none
+python -u tools/train.py configs/fungi2023/swinv2_base_w24_b8x8-fp16_fungi+val_res_384_cb_epochs_6.py --launcher="slurm" # --auto-scale-lr

train.sh

@@ -0,0 +1,6 @@
+#!/bin/bash
+
+conda activate .conda/
+
+set -x
+bash tools/dist_train.sh configs/fungi2023/swinv2_base_w24_b16x4-fp16_fungi+val_res_384_cb_epochs_6.py 4

train_distributed.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+
+conda activate .conda/
+
+set -x
+export SRUN_CPUS_PER_TASK=${SLURM_CPUS_PER_TASK}
+PYTHONPATH="$(dirname $0)":$PYTHONPATH srun --kill-on-bad-exit=1 launch.sh