CS6910 Assignment 2

Assignment 2 submission for the course CS6910 Fundamentals of Deep Learning.

Team members: N Sowmya Manojna (BE17B007), Shubham Kashyapi (MM16B027)

---

Part-A Training a Smaller Network from Scratch

1) Q1 to Q3 (Model architecture and wandb sweeps)

Part-A-Q1-to-Q3.ipynb loads the training and validation datasets. The different hyperparameter configurations for wandb are specified in the variable sweep_config.

sweep_config = {'name': 'random-test-sweep', 'method': 'random'}
sweep_config['metric'] = {'name': 'val_acc', 'goal': 'maximize'}
parameters_dict = {
                   'first_layer_filters': {'values': [32, 64]},
                   'filter_org': {'values': [0.5, 1, 2]}, # Halving, same, doubling in subsequent layers
                   'data_aug': {'values': [False, True]},
                   'batch_norm': {'values': [False, True]}, 
                   'dropout': {'values': [0.0, 0.2, 0.3]},
                   'kernel_size': {'values': [3,5,7]},
                   'dense_size': {'values': [32, 64, 128]},
                   'activation': {'values': ['relu']},
                   'num_epochs': {'values': [50]}, 
                   'optimizer': {'values': ['adam']},
                   'conv_layers': {'values': [5]}
                  }
sweep_config['parameters'] = parameters_dict

The function CNN_train defines the model architecture, trains the model and logs the metrics to wandb. Link for wandb project https://wandb.ai/cs6910-team/DL-Assignment2-PartA-5April-2

2) Q4 (Training without wandb, evaluating on test set, visualizing test images and filters)

PartA_Q4.ipynb can be used to train a model using the optimal hyperparameters obtained in Q2 and Q3. The code for loading the data, initializing the model architecture and training is similar to the previous notebook. So this notebook can be used to quickly test the code without setting up the wandb sweeps. The optimal hyperparameters are specified as follows. These can be modified for the purpose of testing the code.

optimal = {
              'first_layer_filters': 64,
              'filter_org': 1, # Same number of filters in all convolution layers
              'data_aug': True,
              'batch_norm': True, 
              'dropout': 0.2,
              'kernel_size': 3,
              'dense_size': 128,
              'activation': 'relu',
              'num_epochs': 50, 
              'optimizer': 'adam',
              'conv_layers': 5
          }

The trained model can also be accessed directly at https://drive.google.com/drive/folders/1343Tk13X9iyIxdF4SityFtGOV9soLS5c?usp=sharing

model = tf.keras.models.load_model(pathlib.Path('/content/drive/MyDrive/DL_Assignment2_PartA_Model'))

The code for evaluating this model on the test set (Q4 (a)), visualizing test images (Q4 (b)) and visualizing kernels (Q4 (c)) is also included in the same notebook.

2) Q5 (Guided backpropogation)

PartA_Q5.ipynb The gradient for relu activation is redefined. The saved model is loaded and guided backpropogation is performed. The gradient images are visualized.

Part-B Using Pre-trained Models for Image Classification

1. Dataset

The dataset can be downloaded using the drive_dataset_check.ipynb code present in the main directory.

The code downloads the iNaturalist dataset, unzips it and renames the images sequentially (for easy identification of missing images in case of network discrepancy).

The images are then loaded, split into training & validation sets. The image size is restricted to (256, 256) and all images that do not conform to the specified size are automatically resized.

2. Models

The trained model is made modular by implementing a CNN class.

Instances of the class can be created by specifying the base model, a flag determining whether the weights of the last few layers should be trained and an offset (i.e.) the number of layers from the end that have to be trained. The available options of the parameters are as follows:

• base_model_name: A string that specifies the base model that should be loaded
  • "InceptionV3"
  • "InceptionResNetV2"
  • "ResNet50"
  • "Xception"
• tune: A flag that determines whether the last few layers can be trained (default: False)
  • True
  • False
• offset: An positive integer that determines the number of layers from the end that should be trained. (Active only when the tune parameter is set to True; default: 20)

3. Train without wandb

base_model_name = "InceptionV3"
tune = False
offset = 20

model = CNN(base_model_name, tune, offset)

base_learning_rate = 0.0001
model.compile(optimizer=tf.keras.optimizers.Adam(lr=base_learning_rate),
                loss=tf.keras.losses.CategoricalCrossentropy(from_logits=True),
                metrics=['accuracy'])

4. Train with wandb

sweep_config = {'name': 'random-test-sweep', 'method': 'grid'}
sweep_config['metric'] = {'name': 'val_acc', 'goal': 'maximize'}
parameters_dict = {
                   'base_model_name': {'values': ["InceptionV3", "InceptionResNetV2", "ResNet50", "Xception"]},
                   'tune': {'values': [False, True]},
                  }
sweep_config['parameters'] = parameters_dict

sweep_id = wandb.sweep(sweep_config, project = 'DL-Assignment2-PartB-9April')
wandb.agent(sweep_id, function=pretrain_CNN_sweep)

5. Results

The results of the parameter sweep can be accessed here: Wandb report Part-B

---

Part-C YOLO-v3 for Forest Fire Detection

1. Model - Weights and config.json

The trained weights are taken from OlafenwaMoses/FireNET.

The files required for execution are as follows:

• detection_model-ex-33--loss-4.97.h5
• detection_config.json
• pretrianed-yolov3.h5

Note that the size of pretrianed-yolov3.h5 and detection_model-ex-33--loss-4.97.h5 is greater than 100 Mb. So, the files have to downloaded separately.

2. Process Video

In order to detect fire from the video video1.mp4, run the following:

python3 detect_from_video

3. Dataset

The fire dataset videos are taken from MIVIA – Laboratorio di Macchine Intelligenti per il Riconoscimento di Video, Immagini e Audio and Bilkent University.

4. Results

The final processed video can be accessed here: final.mp4

The youtube link for the same: Final Video

The actual and processed video files can be accessed here: CS6910-YOLOv3