[Docs] Custom Dataset Training with PoseC3D

[mohammed-alaa40123]

The doc issue

The tutorial needs to be updated because it says
ann_file_train = 'data/posec3d/custom_dataset_train.pkl' # Your annotation for training
ann_file_val = 'data/posec3d/custom_dataset_val.pkl' # Your annotation for validation

and there is only one ann_file in the config file then in the train/test/val data loaders do the splitting.

Another issue I encountered while using the ntu_pose_extraction I get only the annotations stored in the pkl file based on the old documentation in the website https://mmaction2.readthedocs.io/en/0.x/supported_datasets.html#the-format-of-posec3d-annotations
But the new format of the pose3d annotations suggests this https://github.com/open-mmlab/mmaction2/blob/main/tools/data/skeleton/README.md#the-format-of-annotations

I need some explanation if possible as I get this error
File "/content/drive/MyDrive/mmaction/mmaction2/mmaction/datasets/pose_dataset.py", line 66, in load_data_list
split, annos = data_list['split'], data_list['annotations']
TypeError: list indices must be integers or slices, not str

because my annotations.pkl file doesn't match the new pose_dataset format

dict_keys(['keypoint', 'keypoint_score', 'frame_dir', 'img_shape', 'original_shape', 'total_frames', 'label'])
This is the keys I find in the annotations pickle file There is no split nor annotations

[mohammed-alaa40123]

import os
import random
import pickle

Path to the folder containing the pickle files (relative to the current working directory)

pkl_folder = ''

List to store paths of all pickle files

pkl_files = []

Iterate over each file in the folder

for file_name in os.listdir(pkl_folder):
if file_name.endswith('.pkl'):
pkl_files.append(os.path.join(pkl_folder, file_name))

Shuffle the list of pickle files

random.shuffle(pkl_files)

Calculate the split index for training and validation

split_index = int(0.8 * len(pkl_files))

Split the list of pickle files into training and validation sets

train_files = pkl_files[:split_index]
val_files = pkl_files[split_index:]

Function to merge pickle files

def merge_pickles(pickle_files, output_file):
split = {'xsub_train': [], 'xsub_val': [], 'xview_train': [], 'xview_val': []}
annotations = []

for pkl_file in pickle_files:
    with open(pkl_file, 'rb') as f:
        data = pickle.load(f)
        # Update split dictionary and combine annotations

        annotations.append(data)
        # Assign videos to splits based on their indices
        if pkl_file in train_files:
            split['xsub_train'].append(data['frame_dir'])
        else:
            split['xsub_val'].append(data['frame_dir'])

# Save merged data into a single pickle file
merged_data = {'split': split, 'annotations': annotations}
with open(output_file, 'wb') as f:
    pickle.dump(merged_data, f)

Merge pickle files for training and validation

merge_pickles(train_files + val_files, 'merged_data.pkl')

print("Merge completed!")

[mohammed-alaa40123]

This should solve the problem in the annotations.pkl file :)