Skip to content

Latest commit

 

History

History
123 lines (66 loc) · 5.42 KB

README.md

File metadata and controls

123 lines (66 loc) · 5.42 KB
Raw

statistical-tools

Tools for creating dataset stats

Package requirements

  • numpy
  • matplotlib
  • matplotlib-venn
  • pandas
  • tqdm (for showing progress)
  • nltk, after installing nltk run nltk.download('wordnet') in Python shell

Visual Genome dataset analysis

Human synsets in Visual Genome attribute data

python3 vg_analyze_attributes.py --attributes_json /proj/mediaind/picsom/databases/visualgenome/download/1.2/VG/1.2/attributes.json

Human synsets in Visual Genome relationships data

python3 vg_analyze_relationships.py --rel_counts --relationships_json /proj/mediaind/picsom/databases/visualgenome/download/1.2/VG/1.2/relationships.json

Extracting information on words belonging to different categories, such as time, location, etc

First you will need to download and unzip Stanford Part-Of-Speech (POS) tagger. To perform this run the following command inside the project folder:

bash download_pos_tagger.sh

NOTE: The code assumes that the tag_sentences.py and POS tagger Java files are in the SAME folder. If not, please modify tag_sentences.py to load tagger Jar from your preferred location!

To run the tagger you will also need to make sure that your Java runtime environment is at least version 1.8:

java -version

If the version is lower than 1.8, you will need to make sure that you update your environment to use Java >= 1.8. On CSC Taito environment this can be done with:

module load java/oracle/1.8

Example: Analyze Visual Genome region descriptions, COCO captions and MSR-VTT video captions in PicSOM format

Stage 1: Preprocess that data

1.1) Perform POS tagging, store result in JSON file:

Visual Genome:

python3 tag_sentences.py --dataset vg-regions --num_workers 4 --num_batches 32 --input_file /proj/mediaind/picsom/databases/visualgenome/download/1.2/VG/1.2/region_descriptions.json --output_file output/pos_tags_vg_regions.json

The above command should work with 16 GB job memory, and take around 40 minutes with 4 CPU cores on Taito for Visual Genome region descriptions (processing a total of 5,408,689 captions).

COCO:

python3 tag_sentences.py --dataset coco --num_workers 4 --num_batches 32 --input_file /proj/mediaind/picsom/databases/COCO/download/annotations/captions_train2014.json --output_file output/pos_tags_coco.json

PicSOM / MSR-VTT (gt-raw.txt):

python3 tag_sentences.py --dataset picsom --num_workers 4 --num_batches 32 --input_file /proj/mediaind/picsom/databases/MSR-VTT/textdumps/gt-raw.txt --output_file output/pos_tags_picsom_msrvtt_raw.json

1.2) Infer synsets based on POS tags, store result in JSON file:

Visual Genome:

python3 infer_synsets.py output/pos_tags_vg_regions.json --output_path output

Total synset inference time for Visual Genome regions dataset is around 15 minutes on Taito CPU instance (only one core is used for this task).

COCO:

python3 infer_synsets.py output/pos_tags_coco.json --output_path output

PicSOM:

python3 infer_synsets.py output/pos_tags_picsom_msrvtt_raw.json --output_path output

Stage 2: Match categories and extract stats

2.1) Match synsets to categories

Visual Genome:

python3 match_synsets_to_categories.py --dataset vg-regions --data_file /proj/mediaind/picsom/databases/visualgenome/download/1.2/VG/1.2/region_descriptions.json --synset_file output/pos_tags_vg_regions_syns.json --category location --output_path output

COCO:

python3 match_synsets_to_categories.py --dataset coco --data_file /proj/mediaind/picsom/databases/COCO/download/annotations/captions_train2014.json --synset_file output/pos_tags_coco_syns.json --category location --output_path output

PicSOM:

python3 match_synsets_to_categories.py --dataset picsom --data_file /proj/mediaind/picsom/databases/MSR-VTT/textdumps/gt-raw.txt --synset_file output/pos_tags_picsom_msrvtt_raw_syns.json --category location --output_path output

The above command tries to match synsets that match a user specified category. Synsets that are part of a given category are defined in a dictionary inside categories.py.

The above script outputs 2 files: JSON containing both the original captions and the inferred categories, as well as same data inside a serialized Pandas object that is used for calculating the actual statistics in the next and final step of the pipeline.

2.2) Calculate statistics

Visual Genome:

python3 calculate_category_statistics.py output/pos_tags_vg_regions_syns_location.pandas.pkl --category location --output_name 'Visual Genome Regions' --output_prefix 'vg_regions' --group_by image_id

COCO:

python3 calculate_category_statistics.py output/pos_tags_coco_syns_location.pandas.pkl --category location --output_name 'COCO Train 2014' --output_prefix 'coco' --group_by image_id

PicSOM:

python3 calculate_category_statistics.py output/pos_tags_picsom_msrvtt_raw_syns_location.pandas.pkl --category location --output_name 'PicSOM MSR-VTT RAW' --output_prefix 'msrvtt_gt_raw' --group_by label

NOTE: In case of PicSOM we are grouping by label instead of image_id

Extending generic analysis

  • Different data set handlers can be added to:

    • tag_sentences.py to produce per sentence POS tagging for an arbitrary list of sentences
    • match_synsets_to_categories.py.

  • One can add more category types to categories.py. Right now only location and temporal types are supported.

  • Currently POS tags are not used for inferring relationships and attributes. Adding support for this would require some work.