where to download training data and checkpoints

.gitignore

@@ -38,6 +38,7 @@ runs/
 outputs/
 runs_hyperparam/
 checkpoints/
+data/
 *.pth
 *.ckpt
 *.pt

README.md

@@ -24,13 +24,19 @@ chmod +x /usr/local/bin/cog
 
 ## Cog
 
+download the weights
+
+```bash
+gdown https://drive.google.com/uc?id=1Gn3UdoKffKJwz84GnGx-WMFTwZuvDsuf -O ./checkpoints/
+```
+
 build the image
 
 ```bash
 cog build --separate-weights
 ```
 
-push the image
+push a new image
 
 ```bash
 cog push
@@ -82,6 +88,14 @@ clip/
 └── requirements.txt      # Python dependencies
 ```
 
+## Training Data
+
+To run training on your own, you can find the training data [here](https://drive.google.com/drive/folders/11M6nSuSuvoU2wpcV_-6KFqCzEMGP75q6?usp=drive_link) and put it in the a directory at the root of the project called `./data`.
+
+## Checkpoints
+
+To run predictions with cog or locally on an existing checkpoint, you can find a checkpoint and configuration files [here](https://drive.google.com/drive/folders/1Gn3UdoKffKJwz84GnGx-WMFTwZuvDsuf?usp=sharing) and put them in the a directory at the root of the project called `./checkpoints`.
+
 ## Model Architecture
 
 - Base: CLIP ViT-Large/14

script/hyperparameter_tuning.py

@@ -227,13 +227,13 @@ def run_hyperparameter_search(data_paths, n_trials=100):
 if __name__ == "__main__":
     # List of dataset paths to optimize
     data_paths = [
-        '../finetune/blog/bryant/random',
-        '../finetune/blog/bryant/adjusted',
-        '../finetune/blog/youtube/random',
-        '../finetune/blog/youtube/adjusted',
-        '../finetune/blog/combined/random',
-        '../finetune/blog/combined/adjusted',
-        '../finetune/blog/bryant_train_youtube_val/default'
+        './data/blog/datasets/bryant/random',
+        './data/blog/datasets/bryant/adjusted',
+        './data/blog/datasets/youtube/random',
+        './data/blog/datasets/youtube/adjusted',
+        './data/blog/datasets/combined/random',
+        './data/blog/datasets/combined/adjusted',
+        './data/blog/datasets/bryant_train_youtube_val/default'
     ]
     
     # Run hyperparameter search

script/train.py

@@ -273,56 +273,30 @@ def main():
     config = {
         "class_labels": class_labels,
         "num_classes": len(class_labels),
-        "clip_model": "openai/clip-vit-large-patch14",
-        "batch_size": 32,
-        "unfreeze_layers": 4,
-        "learning_rate": 5.305885796107412e-06,
-        "weight_decay": 4.543630233732527e-07,
-        "gradient_clip_max_norm": 0.6446650879658523,
-        "augmentation_strength": 0.5827616006715585,
-        "crop_scale_min": 0.7872781274088598,
-        "max_frames": 15,
-        "sigma": 0.286510943464138,
-        "data_path": "../finetune/blog/bryant/random",
-        "num_epochs": 50,
-        "patience": 10,
+        "data_path": './data/blog/datasets/bryant/random',
+        "batch_size": 8,
+        "learning_rate": 2e-6,
+        "weight_decay": 0.007,
+        "num_epochs": 2,
+        "patience": 10,  # for early stopping
+        "max_frames": 10,
+        "sigma": 0.3,
         "image_size": 224,
+        "flip_probability": 0.5,
+        "rotation_degrees": 15,
+        "brightness_jitter": 0.2,
+        "contrast_jitter": 0.2,
+        "saturation_jitter": 0.2,
+        "hue_jitter": 0.1,
+        "crop_scale_min": 0.8,
         "crop_scale_max": 1.0,
-        "normalization_mean": [
-            0.485,
-            0.456,
-            0.406
-        ],
-        "normalization_std": [
-            0.229,
-            0.224,
-            0.225
-        ],
+        "normalization_mean": [0.485, 0.456, 0.406],
+        "normalization_std": [0.229, 0.224, 0.225],
+        "unfreeze_layers": 3,
+        # "clip_model": "openai/clip-vit-large-patch14",
+        "clip_model": "openai/clip-vit-base-patch32",
+        "gradient_clip_max_norm": 1.0,
         "overfitting_threshold": 10,
-        # "data_path": '../finetune/blog/bryant/random',
-        # "batch_size": 8,
-        # "learning_rate": 2e-6,
-        # "weight_decay": 0.007,
-        # "num_epochs": 2,
-        # "patience": 10,  # for early stopping
-        # "max_frames": 10,
-        # "sigma": 0.3,
-        # "image_size": 224,
-        # "flip_probability": 0.5,
-        # "rotation_degrees": 15,
-        # "brightness_jitter": 0.2,
-        # "contrast_jitter": 0.2,
-        # "saturation_jitter": 0.2,
-        # "hue_jitter": 0.1,
-        # "crop_scale_min": 0.8,
-        # "crop_scale_max": 1.0,
-        # "normalization_mean": [0.485, 0.456, 0.406],
-        # "normalization_std": [0.229, 0.224, 0.225],
-        # "unfreeze_layers": 3,
-        # # "clip_model": "openai/clip-vit-large-patch14",
-        # "clip_model": "openai/clip-vit-base-patch32",
-        # "gradient_clip_max_norm": 1.0,
-        # "overfitting_threshold": 10,
         "run_dir": run_dir,
     }
     train_and_evaluate(config)

script/visualization/visualize.py

@@ -204,9 +204,7 @@ def run_visualization(run_dir, data_path=None, test_csv=None):
 
 if __name__ == "__main__":
     # Find the most recent run directory
-    # run_dir = get_latest_run_dir()
-    run_dir = "/home/bawolf/workspace/break/clip/runs_hyperparam/hyperparam_20241106_124214/search_combined_adjusted/trial_combined_adjusted_20241106-195023/"
-    # run_dir = "/home/bawolf/workspace/break/clip/runs/run_20241024-150232_otherpeopleval_large_model"
-    # run_dir = "/home/bawolf/workspace/break/clip/runs/run_20241022-122939_3moves_balanced"
-    data_path = "/home/bawolf/workspace/break/finetune/blog/combined/all"
-    run_visualization(run_dir, data_path=data_path)
+    run_dir = get_latest_run_dir()
+    
+    # add a data_path argument to visualize a specific dataset
+    run_visualization(run_dir)

script/visualization/viz_cross_compare.py

@@ -5,10 +5,10 @@ from visualize import run_visualization
 def get_opposite_dataset_path(run_folder):
     # Map run folders to their corresponding opposite dataset training files
     dataset_mapping = {
-        'search_bryant_adjusted': '../finetune/blog/youtube/adjusted',
-        'search_bryant_random': '../finetune/blog/youtube/random',
-        'search_youtube_adjusted': '../finetune/blog/bryant/adjusted',
-        'search_youtube_random': '../finetune/blog/bryant/random'
+        'search_bryant_adjusted': './data/blog/datasets/youtube/adjusted',
+        'search_bryant_random': './data/blog/datasets/youtube/random',
+        'search_youtube_adjusted': './data/blog/datasets/bryant/adjusted',
+        'search_youtube_random': './data/blog/datasets/bryant/random'
     }
     
     for folder_prefix, dataset_path in dataset_mapping.items():