Training Yolo-v2 on Colab

This repository explains how to run the notebook to train Yolo inside a Colab notebook. It save the weights hourly makeing it easy to resume from a previous checkpoint.

How tu use it

1. Setup data

The data should be a .zip file containing the following strucutre

├── data.zip
    ├── cfg     <-  configurations files for yolo (more details below)
    ├── img     <-  Image files .png or .jpg, etc, and the .txt     with markers

The cfg directory should contain:

• obj.data: the path for the train.txt file;
• obj.names: the labels for the in the dataset;
• test.txt: the path wherein the training images are stored;
• train.txt: the path wherein the test images are stored;
• yolo-obj.cfg: last, but not least, the main file to use during training call of YoLo. Create your cfg file by following the instructions described in AlexeyAB. See more in cfg examples.

The img directory is pretty straightfoward.

1.1 Generating train.txt and test.txt

To generate these files use the magical script created by leandrobmarinho, named of write_img_names.py. Take a look in his repo, as well as, star him :)

Using the notebook Cloud_Yolo_Train

• GLOBAL CONFIGURATIONS

DATAFILE_ID = '1UUScyA913_iWAZsS8wWgG0pAeMflAbZ-'
OUTPUT_FOLDER_ID = '1h6tpmWqPd8nVd-4bc8xNPKehf-JBvm54'
FINETUNING = False
RESUME_TRAINING = True
CHECKPOINT_FILE_ID = '1MSSs76SKJeg7v0oBkf61kDuSGyUF4zxC'
WEIGHTS_FILE_NAME = 'yolo-v2_final.weights'

• DATAFILE_ID: The ID you outta get of you zip folder into drive.
• OUTPUT_FOLDER_ID: The ID you outta get of you zip folder you desired to save the output files.
• FINETUNING: that's what it is.
• RESUME_TRAINING: set to false when you want to train from scratch.

WARNING: If you set RESUME_TRAINING to True you must upload the .weights from which you want to resume your training.

• CHECKPOINT_FILE_ID: the id of the uploaded weights.
• WEIGHTS_FILE_NAME: the file name of these weights.set to false when you want to train from scratch.

There is nothing to worry about the rest. Run the following cell as you scroll down and the traning should start and upload .weights to the OUTPUT_FOLDER_ID.

Keep in mind to accept drive permissions and authenticate it.