Update README.md

Author: FeilongTangmonash

README.md

@@ -241,9 +241,10 @@ For multi-node distributed training, configure your training script according to
 
 ### Attentive Probe Evaluation
 
-#### Codec Evaluation
 
-To evaluate the encoder with codec-style patch selection, first navigate to the evaluation directory:
+#### Sampling Evaluation
+
+To evaluate the encoder with uniform frame sampling, first navigate to the evaluation directory:
 
 ```bash
 cd eval_encoder
@@ -252,33 +253,27 @@ cd eval_encoder
 Then run the following command:
 
 ```bash
-torchrun --nproc_per_node=8 --master_port=29512 attentive_prob_codec.py \
+torchrun --nproc_per_node=8 --master_port=29507 attentive_probe.py \
   --eval_freq 1 \
   --default_lr_list 0.0001 \
-  --batch_size 4 \
+  --batch_size 32 \
   --default_weight_decay 0 \
   --dali_py_num_workers 8 \
-  --model_family llava_vit_codec \
+  --model_family llava_vit_sampling \
   --dataset diving48 \
-  --num_frames 64 \
+  --num_frames 8 \
   --model_weight lmms-lab/onevision-encoder-large \
   --model_name hf_llava_vit_large_ln \
   --embedding_size 1024 \
-  --default_epoch 30 \
-  --data_root /path/to/your/data_attentive_probe/ \
-  --cache_dir /path/to/your/cache_residuals/ \
-  --K_keep 2048 \
-  --mv_compensate median
+  --frames_token_num 256
 ```
 
-**Codec-Specific Parameters:**
-- `cache_dir`: Directory for cached codec patches. This is where the codec-selected patches will be stored/loaded.
-- `K_keep`: Number of patches to keep. For example, 256 patches per frame × 8 frames = 2048 total patches. Adjust based on your frame count and desired compression ratio.
-- `mv_compensate`: Motion vector compensation method (e.g., `median`).
+**Sampling-Specific Parameters:**
+- `frames_token_num`: Number of tokens per frame (e.g., 256 tokens for standard sampling).
 
-#### Sampling Evaluation
+#### Codec Evaluation
 
-To evaluate the encoder with uniform frame sampling, first navigate to the evaluation directory:
+To evaluate the encoder with codec-style patch selection, first navigate to the evaluation directory:
 
 ```bash
 cd eval_encoder
@@ -287,23 +282,29 @@ cd eval_encoder
 Then run the following command:
 
 ```bash
-torchrun --nproc_per_node=8 --master_port=29507 attentive_probe.py \
+torchrun --nproc_per_node=8 --master_port=29512 attentive_prob_codec.py \
   --eval_freq 1 \
   --default_lr_list 0.0001 \
-  --batch_size 32 \
+  --batch_size 4 \
   --default_weight_decay 0 \
   --dali_py_num_workers 8 \
-  --model_family llava_vit_sampling \
+  --model_family llava_vit_codec \
   --dataset diving48 \
-  --num_frames 8 \
+  --num_frames 64 \
   --model_weight lmms-lab/onevision-encoder-large \
   --model_name hf_llava_vit_large_ln \
   --embedding_size 1024 \
-  --frames_token_num 256
+  --default_epoch 30 \
+  --data_root /path/to/your/data_attentive_probe/ \
+  --cache_dir /path/to/your/cache_residuals/ \
+  --K_keep 2048 \
+  --mv_compensate median
 ```
 
-**Sampling-Specific Parameters:**
-- `frames_token_num`: Number of tokens per frame (e.g., 256 tokens for standard sampling).
+**Codec-Specific Parameters:**
+- `cache_dir`: Directory for cached codec patches. This is where the codec-selected patches will be stored/loaded.
+- `K_keep`: Number of patches to keep. For example, 256 patches per frame × 8 frames = 2048 total patches. Adjust based on your frame count and desired compression ratio.
+- `mv_compensate`: Motion vector compensation method (e.g., `median`).
 
 #### Shared Parameters