Using 4 channels is better

ngxbac · Jul 8, 2019 · c6f5c1b · c6f5c1b
1 parent cd0a09c
commit c6f5c1b
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 8 deletions.
diff --git a/bin/train.sh b/bin/train.sh
@@ -1,10 +1,10 @@
 #!/usr/bin/env bash
 
-export CUDA_VISIBLE_DEVICES=2,3
+export CUDA_VISIBLE_DEVICES=2
 RUN_CONFIG=config.yml
 
 
-LOGDIR=/raid/bac/kaggle/logs/recursion_cell/test/c123_s1_smooth_nadam_rndsite/se_resnext50_32x4d/
+LOGDIR=/raid/bac/kaggle/logs/recursion_cell/test/c1234_s1_smooth_nadam_rndsite/se_resnext50_32x4d/
 catalyst-dl run \
     --config=./configs/${RUN_CONFIG} \
     --logdir=$LOGDIR \

diff --git a/configs/config.yml b/configs/config.yml
@@ -1,7 +1,7 @@
 model_params:
   model: cell_senet
   model_name: se_resnext50_32x4d
-  n_channels: 3
+  n_channels: 4
   num_classes: 1108
 
 args:
@@ -35,7 +35,7 @@ stages:
     momentum_range: [0.85, 0.95]
 
   data_params:
-    batch_size: 64
+    batch_size: 32
     num_workers: 8
     drop_last: False
 
@@ -44,7 +44,7 @@ stages:
     valid_csv: "./csv/valid_0.csv"
     root: "/raid/data/kaggle/recursion-cellular-image-classification/"
     sites: [1]
-    channels: [1, 2, 3]
+    channels: [1, 2, 3, 4]
 
   stage1:
 
@@ -57,7 +57,7 @@ stages:
         callback: LabelSmoothCriterionCallback
       optimizer:
         callback: OptimizerCallback
-        accumulation_steps: 2
+        accumulation_steps: 4
       accuracy:
         callback: AccuracyCallback
         accuracy_args: [1]

diff --git a/src/make_submission.py b/src/make_submission.py
@@ -36,13 +36,13 @@ def predict_all():
     test_csv = '/raid/data/kaggle/recursion-cellular-image-classification/test.csv'
     # test_csv = './csv/valid_0.csv'
 
-    experiment = 'c123_s1_smooth_nadam_rndsite'
+    experiment = 'c1234_s1_smooth_nadam_rndsite'
     model_name = 'se_resnext50_32x4d'
 
     log_dir = f"/raid/bac/kaggle/logs/recursion_cell/test/{experiment}/{model_name}/"
     root = "/raid/data/kaggle/recursion-cellular-image-classification/"
     sites = [1]
-    channels = [1,2,3]
+    channels = [1,2,3,4]
 
     model = cell_senet(
         model_name="se_resnext50_32x4d",