Lets get started with week-4
My mentor and I decided the tasks for this week given as :
- 1. Extracting all the possible audio features
- 2. Extracting all the possible image features (including pretrained models)
- 3. Hyperparameter tunning the models to get better results
Starting with the image clustering
Last week we saw that bio metric clustering results were really nice as compared to the general image clustering , so this week we decided to go with some more feature extraction and hyper tunning all the clustering models in depth. We use a) transfer learning to extract the features and b)converted images to gray scale to extract the raw features from the image, i gray scale the images to see how the images will be clustered without any color component.
Transfer learning:
This week I tried these pretrained models.
1. VGG16
2. VGG19
3. Resnet50
4. Xception
5.Inception-Resnetv2
6.mobilenet
7.densenet121
8.densenet169
9.densetnet201
I partioned the algorithms into two categories 1)Hierarchical based 2) Density based
For every algorithm i ran with
a)t-SNE+PCA reduced
b)PCA reduced features
c)t-sne reduced features
d)U-map reduced features
| S No. | Code Type | Code link |
|---|---|---|
| 1 | Image feature Extraction with pretrained models | Image feature extraction code |
| 2 | Image feature extraction without pretrained models | Image feature extrcation |
| 3 | Density based Image clustering (with grid search) | clustering-code |
| 4 | Hierarchical based Image clustering (with grid search) | clustering-code |
This week we will see a lot of image clustering results . I will advised to go through all the files, it will be interesting to see the results.
Starting with the Hierarchical clustering results
Density based results
| Pre-trained model | Umap reduced | pca reduced | t-SNE reduced | PCA + t-SNE reduced | |
|---|---|---|---|---|---|
| VGG16 | Density-image-results-umap-vgg16.pdf(One drive link) | Density-image-results-pca-vgg16.pdf(One drive link) | Density-image-results-tsne-vgg16.pdf(One drive link) | Density-image-results-pca-tsne-vgg16.pdf(One drive link) | |
| VGG19 | Density-image-results-pca-vgg19.pdf(One drive link) | ||||
| Resnet-50 | Density-image-results-tsne-resnet50.pdf(One drive link) | ||||
| InceptionNet | |||||
| Mobilenet | Density-image-results-umap-mobilenet.pdf(One drive link) | ||||
| InceptionNet-ResnetV2 | |||||
| Densenet121 | Density-image-results-umap-densenet121.pdf(One drive link) | Density-image-results-pca-tsne-densenet121 (One drive link) | |||
| Densenet169 | |||||
| Densenet201 | Density-image-results-tsne-densenet201.pdf(One drive link) | ||||
| Xception | Density-image-results-tsne-xception (One drive link) | Density-image-results-pca-tsne-xception (One drive link) | Gray Scale |
Getting started with the Audio clustering
In the Audio clustering I extracted all the possible audio fetaures (listed in the table below)and then I ran all the clustering models in four cases:
a)Actual features
b)PCA dimensionality reduction
c)t-Sne dimensionality reduction
d)Umap dimensionality reduction
The feature extraction code can be found here https://github.com/Himani2000/GSOC_2020/blob/master/Feature_extraction/Audio/FeatureExtraction_pipeline.ipynb