from yoloexplorer import Explorer
exp = Explorer("VOC.yaml")exp.build_embeddings()
coco_exp = Explorer("coco128.yaml")coco_exp.build_embeddings()#Init coco8 similarly
exp.dash([coco_exp, coco8])#Automatic analysis coming soon with dash(..., analysis=True)

You can now explore multiple pretrained models listed"resnet18", "resnet50", "efficientnet_b0", "efficientnet_v2_s", "googlenet", "mobilenet_v3_small" for extracting better features out of images to improve searching across multiple datasets.

from yoloexplorer import Explorer
exp = Explorer("coco128.yaml", model="resnet50")exp.build_embeddings()
coco_exp = Explorer("coco128.yaml", model="mobilenet_v3_small")coco_exp.build_embeddings()
#Use force=True as a parameter in build_embedding if embeddings already exists
exp.dash([coco_exp, coco8])#Automatic analysis coming soon with dash(..., analysis=True)

You can get the schema of you dataset once the table is built

schema = coco_exp.table.schema

You can use this schema to run queries

SQL query
Let's try this query and print 4 result - Select instances that contain one or more 'person' and 'cat'

df=coco_exp.sql("SELECT * from 'table' WHERE labels like '%person%' and labels LIKE '%cat%'")coco_exp.plot_imgs(ids=df["id"][0:4].to_list())

Result

The above is equivlant to plotting directly with a query:

voc_exp.plot_imgs(query=query,n=4)

Querying by similarity
Now lets say your model confuses between cetain classes( cat & dog for example) so you want to look find images similar to the ones above to investigate.

The id of the first image in this case was 117

imgs,ids=coco_exp.get_similar_imgs(117,n=6)# accepts ids/idx, Path, or img blobvoc_exp.plot_imgs(ids)

The above is equivlant to directly callingplot_similar_imgs

voc_exp.plot_similar_imgs(117,n=6)

NOTE: You can also pass any image file for similarity search, even the ones that are not in the dataset

Similarity Search with SQL Filter (Coming Soon)
Soon you'll be able to have a finer control over the queries by pre-filtering your table

coco_exp.get_similar_imgs(..., query="WHERE labels LIKE '%motorbike%'")coco_exp.plot_similar_imgs(query="WHERE labels LIKE '%motorbike%'")

Additional Details

• Theplot_imgs method can be used to visualize a subset of images from the dataset. Theids argument can be a list of image IDs, or a SQL query that returns a list of image IDs. Then argument specifies the number of images to plot.
• Theplot_similar_imgs method can be used to visualize the topn similar images to a given image. Theimg/idx argument can be the index of the image in the dataset, the path to the image file, or the encoded/binary representation of the image.
• Theplot_similarity_index method can be used to visualize the similarity index of the dataset. The similarity index is a measure of how similar each image is to all the other images in the dataset. Thetop_k argument specifies the percentage of images to keep for the similarity index. Thesim_thres argument specifies the similarity threshold. Thereduce argument specifies whether to reduce the dimensionality of embeddings before calculating the index. Thesorted argument specifies whether to sort the similarity index.

Removing data
You can simply remove images by passing a list ofids from the table.

coco_exp.remove_imgs([100,120,300..n]) # Removes images at the given ids.

Adding data
For adding data from another dataset, you need an explorer object of that dataset with embeddings built. You can then pass that object along with the ids of the imgs that you'd like to add from that dataset.

coco_exp.add_imgs(exp, idxs) #

Note: You can use SQL querying and/or similarity searches to get the desired ids from the datasets.

Persisting the Table: Create new dataset and start training
After making the desired changes, you can persist the table to create the new dataset.

coco_exp.persist()

This creates a new dataset and outputs the training command that you can simply paste in your terminal to train a new model!

Resetting the Table
You can reset the table to its original or last persisted state (whichever is latest)

coco_exp.reset()

voc_exp.plot_similarity_index()

You can also get the the similarity index as a numpy array to perform advanced querys.

sim=voc_exp.get_similarity_index()

Now you can combine the similarity index with other querying options discussed above to create even more powerful queries. Here's an example:

"Let's say you've created a list of candidates you wish to remove from the dataset. Now, you want to filter out the images that have similarity index less than 250, i.e, remove the images that are 90%(sim_thres) or more similar to more than 250 images in the dataset."

ids= [...]# filtered ids listfilter=np.where(sim>250)final_ids=np.intersect1d(ids,filter)# intersect both arraysexp.remove_imgs(final_ids)