Quimify Classifier

Overview

Quimify Classifier is the model that sorts a chemistry query into its category before it reaches the resolvers. Given an input string (a formula or a name), it decides which of these 4 classes it belongs to. It is used in Quimify.

• inorganicFormula, e.g. H2SO4, NaCl
• organicFormula, e.g. CH3-CH2-OH
• inorganicName, e.g. cloruro de sodio
• organicName, e.g. 2-bromo-2-cloropropano

It is not a single multiclass model but a tree of 3 binary classifiers chained with a 0.5 threshold:

input ──► formula-name ──┬─(formula)─► formula-inorganic-organic ──► inorganic / organic
                         └─(name)────► name-inorganic-organic ─────► inorganic / organic

How it works

Each of the 3 classifiers is a character-level Keras Sequential network. The preprocessing step (Model.data_standardization) is the key piece: it lowercases the input and inserts a space after every character via regex, so TextVectorization tokenizes per character instead of per word:

"CH3-CH=CH"  ──►  "c h 3 - c h = c h"

Full pipeline:

TextVectorization(max_tokens=2048, output_sequence_length=16, standardize=data_standardization)
  → Embedding(max_features+1, embedding_dim=1024)
  → Dropout(0.2)
  → Dense(256, relu)
  → GlobalAveragePooling1D()
  → Dropout(0.2)
  → Dense(1)                       # logit

• Compilation: adam optimizer, BinaryCrossentropy(from_logits=True) loss, BinaryAccuracy(threshold=0.0) metric.
• data_standardization is registered with @tf.keras.utils.register_keras_serializable() so it ships inside the SavedModel.
• On export, the model is re-wrapped as [TextVectorization, model, Activation("sigmoid")], so the artifact takes raw strings and returns a probability in [0, 1].

Usage

Clone the repository and install the dependencies (Python, tensorflow==2.12.0, Flask==2.3.1):

pip install -r requirements.txt

1. Prepare the data

data/split.sh reads data/source/ and builds data/split/ with a 90/10 train/test hold-out, one file per line (the format text_dataset_from_directory expects):

cd data
./split.sh

It refuses to run if data/split/ already holds data, so delete it to regenerate.

2. Train a model

src/builder.ipynb is a parameterizable notebook. Pick which model to train through the indices at the top:

model_prefix    = ["", "formula-", "name-"][2]     # -> "name-"
first_category  = ["formula", "inorganic"][1]       # -> "inorganic"
second_category = ["name", "organic"][1]            # -> "organic"
# model_name = "name-inorganic-organic"

It loads data/split/<model_name> with text_dataset_from_directory (validation_split=0.2, seed=32), adapts TextVectorization on the train set, fits for 5 epochs, plots accuracy and loss (matplotlib), evaluates on test, and exports to models/<model_name>-model in SavedModel (tf) format. Run it three times, changing the indices, to regenerate all 3 models.

3. Serve inference

src/api.py loads the 3 models from the given directory and starts the service. It clears the Keras session after every request to avoid a memory leak:

cd src
python api.py ../models

4. Package a build

target.sh copies api.py, model.py, requirements.txt and models/ into target/<date>/:

./target.sh

Structure

data/
  source/            # raw corpus (one entry per line)
    formulas.txt, names.txt
    {inorganic,organic}-{formulas,names}.txt
    random/          # generated subsets (16k, 260k organic)
  split.sh           # builds the train/test tree from source/
models/              # 3 exported models in SavedModel (.pb) format
  formula-name-model/
  formula-inorganic-organic-model/
  name-inorganic-organic-model/
src/
  model.py           # Model class: architecture, standardization, load/predict
  builder.ipynb      # training notebook (data → fit → export)
  api.py             # inference service (Flask)
target.sh            # packages a dated build for deployment
requirements.txt     # Flask==2.3.1, tensorflow==2.12.0