marco

Automating the Evaluation of Crystallization Experiments

This is a pretrained model described in the paper:

Classification of crystallization outcomes using deep convolutional neural networks.

This model takes images of crystallization experiments as an input:

It classifies it as belonging to one of four categories: crystals, precipitate, clear, or 'others'.

The model is a variant of Inception-v3 trained on data from the MARCO repository.

Model

The model can be downloaded from:

https://storage.googleapis.com/marco-168219-model/savedmodel.zip

Example

1. Install TensorFlow and the Google Cloud SDK.

2. Download and unzip the model:

unzip savedmodel.zip

3. A sample image can be downloaded from:

https://storage.googleapis.com/marco-168219-model/002s_C6_ImagerDefaults_9.jpg

Convert your image into a JSON request using:

python jpeg2json.py 002s_C6_ImagerDefaults_9.jpg > request.json

4. To issue a prediction, run:

gcloud ml-engine local predict --model-dir=savedmodel --json-instances=request.json

The request should return normalized scores for each class:

CLASSES                                            SCORES
[u'Crystals', u'Other', u'Precipitate', u'Clear']  [0.926338255405426, 0.026199858635663986, 0.026074528694152832, 0.021387407556176186]

CloudML Endpoint

The model can also be accessed on Google CloudML by issuing:

gcloud ml-engine predict --model marco_168219_model --json-instances request.json

Ask the author for access privileges to the CloudML instance.

Note

002s_C6_ImagerDefaults_9.jpg is a sample from the MARCO repository, contributed to the dataset under the CC BY 4.0 license.

Author

Vincent Vanhoucke (github: vincentvanhoucke)