Skip to content

Latest commit

 

History

History
178 lines (109 loc) · 5.79 KB

google-ai.md

File metadata and controls

178 lines (109 loc) · 5.79 KB
Raw

Google Cloud

Project Settings

  • Project birdmig
  • Region us-central1

DATASET: birdmig1_animals_1000

10.2.2020 Single-label classification 1000 random images from the full c. 5900 image set.

Labels:

  • ignore: "ignore" in tagList or "distortion2" in tagList or "faded" in tagList or "high-pass" in tagList
  • animal (376 images): "migrant" in tagList or "migrant-low" in tagList or "migrant-low" in tagList or "wander" in tagList or "local_individual" in tagList or "local_choir" in tagList or "owl" in tagList or "mystery" in tagList or "mammal" in tagList or "dog" in tagList or "other_animal" in tagList or "bat" in tagList:
  • no-animal (624 images): the rest

MODEL: birdmig1_animals_1000_20200210

Edge High-precision max 10 node hours, c. 2,5 hours used

DATASET:

11.2.2020

MODEL: birdmig1_animals_full_20200212

Edge High-precision max 20 node hours, c. 2,5 hours used

Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 4076: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 5802: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 4875: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 4937: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 3005: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 4719: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 5825: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full.csv line 2921: The label is already a ground truth.

AUC c. 0.96 Precison & recall: 92.57 % Animals: c. 91.8 %

False negatives:

  • Melnig's
  • Scorus's
  • Faint migrants
  • Sounds in the beginning & end of segment
  • Some local fieldfares
  • Birds with other noises (single noises, rain, mechanical noises)
  • Thrushes, can be mixed with human sounds?

False positives

  • Loud planes
  • Human sounds: coughing, footsteps
  • Unforeseen mechanical noises & snaps
  • Uniform background noise, which causes "light" spectrogram
  • Strong wind

Exported model birdmig1_animals_full_20200212.pb

DATASET:

23.2.2020

MODEL: birdmig1_animals_full_20200223

Edge High-precision max 20 node hours, 8 estimated, c. NN hours used

Operation ID: projects/825817147834/locations/us-central1/operations/ICN7374201286633193472 Error Messages: Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 4023: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 4052: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 4800: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 4862: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 5302: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 4655: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 5749: The label is already a ground truth. Error: gs://spectro-us/spectro-1/_data/target_annotations_animals_full_20200223.csv line 5741: The label is already a ground truth.

Possible datasets & models

Check what kind of sounds are problematic and add those

  • grasshoppers

Heavy rain and/or wind separately -> give up on those?

Multi-classification, with heavy rain & heavy wind as separate classes -> System can give up and not report them.

Test with Juha Saari's recording

Train with Xeno-Canto, test with own recordings, or vice versa. Balance bird/non bird 50/50%.

Multi-label classification

  • bird
  • bird chorus
  • bat
  • rain & heavy rain
  • wind & heavy wind

Notes

Cloud hosted / Host your model on Google Cloud for online predictions Edge / Download your model for offline/mobile use

  • optimize for higher accuracy / 360ms latency for Google Pixel 2 ... Best trade-off / 150 ms latency ... Faster predictions / 56 ms latency

Specify your own TRAIN, TEST, VALIDATION split. The tool randomly assigns images, but near-duplicates may end up in TRAIN and VALIDATION which could lead to overfitting and then poor performance on the TEST set.

rule of thumb - the label with the lowest number of examples should have at least 10% of the examples as the label with the highest number of examples.

Using a somewhat novel dataset to fine-tune model structure means your model will generalize better

CSV formats

Single-label classification:

[set,]image_path[,label] TRAIN,gs://My_Bucket/sample1.jpg,cat TEST,gs://My_Bucket/sample2.jpg,dog

Multi:

[set,]image_path[,label1][,label2][,labelN]

Commands

Searching files with regex/wildcard in mongodb:

{ file_id: { $in: [ /^noordwijk*/ ] } }

Create bucket:

gsutil mb -l {REGION} gs://spectro-us/

List directories in bucket:

gsutil ls gs://spectro-us/

Copy single file:

gsutil cp FILENAME gs://spectro-us/spectro-1/_data/

Copies files from current dir to bucket, but does not create folder structure:

gsutil cp -r $(find . -name "*.png") gs://spectro-us/

Rsync all subdirs, excluding mp3-files:

First go to the directory to sync (_exports)

gsutil -m rsync -r -x ".*.mp3$" . gs://spectro-us/lepidoptera/

Move bucket ...

gsutil mb -l us-central1 gs://spectro-us/ gsutil cp -r gs://spectro-1/ gs://spectro-us/