• In this thesis, the development of a discrepancy based domain adaptation models named MBM (Modified Baseline Network - an extension of Deep CORAL) & CDAN (Custom Domain Adaptive Network) are discussed. The models are trained on popular benchmarked visual recognition domain datasets like Office-31, GTSRB and Synthetic Signs for image classification tasks, and their performances are evaluated compared to other available domain adaptation methods.
• The "Magnitude based weight pruning" with Constant Sparsity approach is used to perform target feature extractor optimization.

1. models.py module defines the source & target models. Xception Network & Top layers
2. config.py module defines various parameters like set paths, domain adaptation scenarios, backbone model selection, etc.
3. loss.py defines the domain alignment loss functions. Deep CORAL loss, KL Divergence, etc.
4. preprocessing.py module defines data preprocessing pipeline with various experimental scenarios including Data augmentation methods.
5. train_test.py is a helper module which defines training and evaluation methods, including Evaluatiion metrics like Confusion Matrix, etc.
6. utlis.py defines various plotting, helper methods and various logging paths like tensorboard, csv, model checkpoint, etc.
7. main.py is the runnable script which defines various command line arguments of the experiment. In progress mode = "eval", script is running for mode="train_test"
8. requirements.txt defines the libraries dependency of the experiments.
9. Use shell script run.sh to run multiple experiments.
10. evaluation folder shows the loss/accuracy plots, also can be viewed in Tensorboards.
11. model_data folder stores the intermediate and final weights of the model.
12. logs folder saves the logs for a particular run and create experiments.log file.
13. data folder contains the datasets.
14. Monitor experiments.log for log paths and script progress.
15. Check the tensorboard logs by: tensorboard --lodir "path to tb logs"
16. Check training_logs.csv for model training logs.
17. Log paths: logs/CombinationID_BackboneModel_DomainLossUsed_LambdaWeight_Original/DateTimeStampValue) -> MBM logs/CombinationID_BackboneModel_DomainLossUsed_LambdaWeight/DateTimeStampValue) -> CDAN
18. For Pruning: logs/CombinationID_BackboneModel_DomainLossUsed_LambdaWeight_PrunedValue/DateTimeStampValue) -> CDAN logs/CombinationID_BackboneModel_DomainLossUsed_LambdaWeight_Original_PrunedValue/DateTimeStampValue) -> MBM

1. Create conda environment (tf): Install all the required dependencies using both pip and conda as mentioned in the requirements.txt file.
2. Activate conda environment by: conda activate tf.
3. You may launch the program by executing the main.py script directly from an IDE or via terminal.
4. Also, one can run the shell command sh run.sh in order to run the series of python experiments.

python main.py --combination="Amazon_to_Webcam" --architecture="Xception" --batch_size=16 resize=299
--learning_rate=0.0001 --mode="train_test" --lambda_loss=0.5 --epochs=50
--input_shape=(299,299,3) --output_classes=31 --loss_function="CORAL" --augment --prune --prune_val=0.30 --technique --save_weights --save_model --use_multiGPU