CNNs

The CNNs repository is a collection of python codes used for training deep convolutional neural networks. The repository's primary focus is on implementing and experimenting with convolutional neural networks (CNNs) using Python.

How to train?

Before running the main.py script in the CNNs repository, you'll need to specify some hyperparameters within the script. main.py is the primary training script for the CNN models, responsible for training the model on a specific dataset, validating the model with specified validation data, and saving the trained model to a specified location. Additionally, the script allows for hyperparameter tuning, enabling you to adjust the model's parameters and fine-tune it to get better results.

Prepare your own data-set

To start on your own dataset, first you need to prepare it. Just split your data into the train and validation set and build a table from them. The table consists of two columns:

• image_name: the path to the image.
• tags: the ground-truth label corresponding to the image.

Here is an example for the train.csv and val.csv files:

image_name	tags
./dataset/train_image1.png	label1
./dataset/train_image2.png	label2
.....	.....

image_name	tags
./dataset/val_image1.png	label1
./dataset/val_image2.png	label2
.....	.....

Set hyper-parameters

Based on your data and the desired output, you need to set hyper-parameters. Here is the definition of some parameters:

• save_dir: all checkpoints and logs are saved in this directory.
• loss_function: there are two main loss functions that you can use during the training process.
• focalLoss_alpha and focalLoss_gamma: if you set the loss function to Focal, then you need to set these params.
• fineTune_enable: if you want to fine-tune an existing network, you need to set it True.
• fineTune_layers: when you set the fineTune_enable True you need to define the layers which you want to train them. In the other words, unfreeze these layers.
• fineTune_batchNorm: if the model has BatchNorm layers, you can unfreeze them using this param.

Define the Network

The model could be an existing network or your own model. Here is an example of how to load the pretrained ImageNet model:

model = models.resnet50(pretrained=True)
number_feats = model.fc.in_features
model.fc = nn.Linear(in_features=number_feats, out_features=number_of_classes)

You may load your own model using below code:

my_model = torch.load('./snapshots/model.pth')
model = my_model['state_dict'].module ## when your model has a state_dict 

Start training

Finally, start to train your model using this command:

python main.py

To monitor the training process use the log file in the save_dir:

tensorboard --logdir=/path/to/event-file

Future work

• Add inference code
• Add new models
• Add normalization
• Add ONNX convertor