Tune hyper params

ikergarcia1996 · Mar 28, 2022 · 3d22f58 · 3d22f58
1 parent 4515317
commit 3d22f58
Show file tree

Hide file tree

Showing 4 changed files with 11 additions and 13 deletions.
diff --git a/training_scripts/GPU/TEDD_1140_base.sh b/training_scripts/GPU/TEDD_1140_base.sh
@@ -16,7 +16,6 @@ python3 train.py --train_new \
   --dropout_encoder_features 0.3 \
   --control_mode keyboard \
   --val_check_interval 0.5 \
-  --hide_map_prob 0.4 \
   --devices 1
 
 
diff --git a/training_scripts/GPU/TEDD_1140_large.sh b/training_scripts/GPU/TEDD_1140_large.sh
@@ -16,7 +16,6 @@ python3 train.py --train_new \
   --dropout_encoder_features 0.3 \
   --control_mode keyboard \
   --val_check_interval 0.5 \
-  --hide_map_prob 0.4 \
   --devices 1
 
 
diff --git a/training_scripts/TPU/TEDD_1140_base.sh b/training_scripts/TPU/TEDD_1140_base.sh
@@ -5,23 +5,23 @@ python3 train.py --train_new \
   --val_dir  ../gtaai_datasets/dev \
   --output_dir models/tedd_1104_base \
   --encoder_type transformer \
-  --dataloader_num_workers 32 \
-  --batch_size 8 \
+  --dataloader_num_workers 16 \
+  --batch_size 16 \
   --accumulation_steps 1 \
   --max_epochs 40 \
-  --cnn_model_name efficientnet_b4 \
+  --cnn_model_name convnext_base \
   --num_layers_encoder 4 \
   --embedded_size 512 \
-  --learning_rate 1e-5 \
+  --learning_rate 5e-5 \
   --mask_prob 0.2 \
   --dropout_cnn_out 0.3 \
   --dropout_encoder 0.1 \
   --dropout_encoder_features 0.3 \
   --control_mode keyboard \
   --val_check_interval 0.5 \
-  --hide_map_prob 0.4 \
+  --precision 16 \
   --devices 8 \
   --accelerator tpu \
-  --report_to tensorboard
+  --report_to wandb
 
 
diff --git a/training_scripts/TPU/TEDD_1140_large.sh b/training_scripts/TPU/TEDD_1140_large.sh
@@ -5,21 +5,21 @@ python3 train.py --train_new \
   --val_dir  ../gtaai_datasets/dev \
   --output_dir models/tedd_1104_large \
   --encoder_type transformer \
-  --dataloader_num_workers 32 \
+  --dataloader_num_workers 16 \
   --batch_size 8 \
   --accumulation_steps 1 \
   --max_epochs 40 \
-  --cnn_model_name efficientnet_b7 \
-  --num_layers_encoder 4 \
+  --cnn_model_name convnext_large \
+  --num_layers_encoder 6 \
   --embedded_size 512 \
-  --learning_rate 1e-5 \
+  --learning_rate 5e-5 \
   --mask_prob 0.2 \
   --dropout_cnn_out 0.3 \
   --dropout_encoder 0.1 \
   --dropout_encoder_features 0.3 \
   --control_mode keyboard \
   --val_check_interval 0.5 \
-  --hide_map_prob 0.4 \
+  --precision 16 \
   --devices 8 \
   --accelerator tpu \
   --report_to wandb