Adapted from https://github.com/ClaudeCoulombe/deep-learning-with-python-notebooks/blob/master/5.3-using-a-pretrained-convnet.ipynb

The data was generated by crawling google and reddit. It was manually curated after reading about canabis diseases by a nonexpert. Please consult a professional if your marijuana is sick.

One source identifies 30 types of cannabis ailments:

1. boron deficiency
2. broad mites
3. bud rot
4. calcium deficiency
5. copper deficiency
6. fungus gnats
7. heat light stress
8. iron deficiency
9. yellow leaf spot leaf septoria
10. light burn
11. magnesium deficiency
12. male plants bananas hermies
13. manganese deficiency
14. molybdenum deficiency
15. nitrogen deficiency
16. nitrogen toxicity
17. nutrient burn
18. overwatering
19. bugs pests symptoms marijuana grow
20. ph fluctuations
21. phosphorus deficiency
22. potassium deficiency
23. root problems
24. root rot
25. spider mites
26. sulfur deficiency
27. underwatering
28. white powdery mildew
29. wind burn
30. zinc deficiency

resources:

• https://www.pinterest.com/search/pins/?q=nitrogen%20deficiency%20canabis&rs=typed&term_meta[]=nitrogen%7Ctyped&term_meta[]=deficiency%7Ctyped&term_meta[]=canabis%7Ctyped
• http://www.marijuana-seeds.net/marijuana-plant-problems
• https://www.growweedeasy.com/cannabis-symptoms-pictures/
• https://towardsdatascience.com/build-a-taylor-swift-detector-with-the-tensorflow-object-detection-api-ml-engine-and-swift-82707f5b4a56

the "other" dataset was generated from:

• https://www.kaggle.com/alxmamaev/flowers-recognition/home
• https://www.kaggle.com/sayangoswami/reddit-memes-dataset/home
• https://www.kaggle.com/jessicali9530/caltech256/home we wanted to build an "other" dataset with images that may fool the model, we collected images likely to be uploaded by random users from their phones, random objects, plants that aren't cannabis and images with overlayed text. Many of the sick samples were from sites that tend to add overlay text describing the cannabis's condition.

healthy 1120 sick 946 not_weed 964

train/validation/test 60/20/20

• data augmentation make validation accuracy worse on first attempt. Must mean the transformations are not representative of valid data for our problem.
  • find transformations that produce valid data for our problem
• resnet50 may overfit on datasets < 10K images
  • try VGG16
• visualize misclassified samples to gain intuition about where the model is struggling
• curate more data
• consider dropping photos of groups of plants

• see projects gcloud projects list
• see current project gcloud config list project
• change project gcloud config set project <project name>
• set bucket name BUCKET_NAME="keras-class-191806"
• set your preferred region REGION=us-east1
• copy model over gsutil cp src gs://$BUCKET_NAME/saved_model.pg NOTE: ml-engine expects your model's protocol buffer file to be name saved_model.{pb,pbtxt}
• register ml-engine entry gcloud ml-engine models create <model name>
• deploy version of model gcloud ml-engine versions create v4 --model=plantDisease01 --origin=gs://keras-class-191806/plantDisease01/vgg16_data_augmented-tf --runtime-version=1.4 the origin arg must be a local dir with the full tf model (protocol buffer and variables)
• see models gcloud ml-engine models list
• craft request payload python -c 'import base64, sys, json; img = base64.b64encode(open(sys.argv[1], "rb").read()).decode(); print(json.dumps({"foo-input": {"b64": img}}))' test_healthy.jpg &> request.json
• craft request python -c 'req = []; [req.append(0.2) for i in range(224*224)]; print(req)' &> request-float32.json
• inference gcloud ml-engine predict --model=plantDisease01 --json-instances=request.json