Generate
Generation is recommended to be done in a Google colaboratory notebook so that you can make use of it's compute resources. You should open one of the notebooks in colab and execute the set up code boxes before executing the generate code block. This will create a training set (size 192,000), validation set (size 19,200) and a test set (size 100)using the captcha symbol set 'ABCDeFghijkMnPQRSTUVWXxYZz0123456789#/[]:><%{}-+'
Alternatively you can run our generate.sh script inside a virtual environment once you clone this repository, adjusting the paths to your own specifications.
Training
Training is also recommended being done in Google colaboratory, executing the training code block will train using the previously generated training and validation image sets in batches of 32 for 6 epochs.
Alternatively you can run our train.sh script inside a virtual environment once you clone this repository, adjusting the paths to your own specifications.
Classification
We convert the model to a tflite model by executing the convert code block in colaboratory or using the convert.py file in the tflite folder and then that tflite is transferred to the raspberry pi using git.
And the classify.sh script is executed using the specified model to classify the images inside the venv.
We first execute the get Filenames script by passing our shortname and filename to the endpoint via a get request. We then parse the contents of the response from html to csv using the htmlToCSV python script. Finally we utilise a bulk file download script (FileScrapper) to be run on the raspberry pi using aiohttp- an async I/O API. Which allowed us to speed up and automate the task of getting each of the individual images. This can be further extended to handle larger I/O intensive tasks like this.
We trialled multiple pre-processing approaches, which can be found in our preprocessing file. We found that less seemed to be more in this case and ended up with the preprocessing suggested in the original code we downloaded from blackboard.
We generated images in google colaboratory notebooks.
We altered the provided generate.py to generate images of random length using the numpy rand function in the range 0 to 6 to randomly select the length of all images in the set. We did not use padding at this stage but instead adding spaces to the ends of any image with a length less than 6 while training.
Since some special characters can't be used for filenames we decided to implement a mapping functionality for the image names. We stored the label for each image in a set_labels.txt file and mapped the image name to it via indexes. This label file is called during training.
For the final model we used 192,000 training images, this gave us 6,000 batches per epoch. We generated and used 19,200 validation images as well.
We started by working with models we had from the previous assignment. We each trained a model over what we deemed to be a whole symbol set by combining what our classmates suggested on piazza with what we observed from our own sets. We then began to reduce that symbol set through eliminations and submissions.
"0123456789eghijknpoqsuvwxyzABCDFJKMPQRSTUVWXYZ#[]+:></%{}-|_"
'ABCDeFghijkMnPQRSTUVWXxYZz0123456789#/[]:><%{}-+
| Metric | Time |
|---|---|
| Generate | |
| Train 192000 | 13m 58s |
| Val 19200 | 1m 6s |
| Training | |
| Per Epoch | 10-18m |
| Classification | |
| For 2000 images | 10m |