Add prompt_method and neg_prob to the phrasecut dataset

naamiinepal · Dec 30, 2023 · 962f68b · 962f68b
1 parent aedcf8e
commit 962f68b
Show file tree

Hide file tree

Showing 4 changed files with 1,779 additions and 29 deletions.
diff --git a/configs/data/phrasecut.yaml b/configs/data/phrasecut.yaml
@@ -6,6 +6,8 @@ train_ds:
   tokenizer_pretrained_path: ${tokenizer_pretrained_path}
   transforms: ${train_transforms}
   return_tensors: pt
+  prompt_method: shuffle+
+  neg_prob: 0.2
 
 val_ds:
   _target_: src.data.core_datasets.phrasecutdataset.PhraseCutDataset
@@ -14,6 +16,8 @@ val_ds:
   tokenizer_pretrained_path: ${tokenizer_pretrained_path}
   transforms: ${val_transforms}
   return_tensors: pt
+  prompt_method: fixed
+  neg_prob: 0
 
 test_ds:
   _target_: src.data.core_datasets.phrasecutdataset.PhraseCutDataset
@@ -22,6 +26,8 @@ test_ds:
   tokenizer_pretrained_path: ${tokenizer_pretrained_path}
   transforms: ${test_transforms}
   return_tensors: pt
+  prompt_method: fixed
+  neg_prob: 0
 
 batch_size: 16
 num_workers: 4

diff --git a/configs/experiment/phrasecut.yaml b/configs/experiment/phrasecut.yaml
@@ -35,8 +35,8 @@ model:
     factor: 0.1
     patience: 5
   net:
-    freeze_image_encoder: false
-    freeze_text_encoder: false
+    freeze_image_encoder: true
+    freeze_text_encoder: true
     add_pos_enc: false
     decoder_layer_kwargs:
       nhead: 8
@@ -47,7 +47,7 @@ model:
   compile: false # Torch compile works only when CUDA Compat >= 7.0
 
 data:
-  batch_size: 32
+  batch_size: 64
 
 logger:
   wandb:
@@ -61,7 +61,7 @@ logger:
 # Additional parameters    #
 ############################
 
-model_pretrained_path: openai/clip-vit-base-patch32
+model_pretrained_path: openai/clip-vit-base-patch16
 tokenizer_pretrained_path: ${model_pretrained_path}
 image_pretrained_path: ${model_pretrained_path}
 text_pretrained_path: ${model_pretrained_path}
@@ -73,19 +73,29 @@ img_mean: [0.48145466, 0.4578275, 0.40821073]
 img_std: [0.26862954, 0.26130258, 0.27577711] 
 
 # Image pre-processing configs
-train_transforms: 
+train_transforms:
   _target_: albumentations.Compose
   transforms:
-    - _target_: albumentations.Resize
-      height: ${img_size}
+    - _target_: albumentations.SmallestMaxSize
+      max_size: ${img_size}
+    - _target_: albumentations.Rotate
+      limit: 10
+      border_mode: 1 # cv2.BORDER_REPLICATE
+      p: 0.2
+    - _target_: albumentations.RandomCrop
       width: ${img_size}
+      height: ${img_size}
+    - _target_: albumentations.RandomBrightnessContrast
+      contrast_limit: 0.1
+      brightness_limit: 0.1
+      brightness_by_max: false
     - _target_: albumentations.Normalize
       mean: ${img_mean}
       std: ${img_std}
     - _target_: albumentations.pytorch.ToTensorV2
       transpose_mask: true
 
-eval_transforms:
+eval_transforms: 
   _target_: albumentations.Compose
   transforms:
     - _target_: albumentations.Resize