quebic-source / tf2crf

CRF layer for tensorflow 2 keras

tf2crf

• a simple CRF layer for tensorflow 2 keras
• support keras masking

Install

$ pip install tf2crf

Features

• easy to use CRF layer with tensorflow
• now support the ModelWithCRFLossDSCLoss with DSC loss, which increases f1 score with unbalanced data (refer the paper Dice Loss for Data-imbalanced NLP Tasks)

Attention

• A dense layer is needed before the CRF layer to convert inputs to shape (batch_size, timesteps, num_classes). The 'num_class' is how many tags or catogories the model predicts.
• I have changed the previous way that putting loss function and accuracy function in the CRF layer. Instead I choose to use ModelWappers (refered to jaspersjsun), which is more clean and flexible.

Tips

tensorflow >= 2.1.0 Recommmend use the latest tensorflow-addons which is compatiable with your tf version.

Example

import tensorflow as tf
from tensorflow.keras.layers import Input, Embedding, Bidirectional, GRU, Dense
from tensorflow.keras.models import Model
from tf2crf import CRF, ModelWithCRFLoss

inputs = Input(shape=(None,), dtype='int32')
output = Embedding(100, 40, trainable=True, mask_zero=True)(inputs)
output = Bidirectional(GRU(64, return_sequences=True))(output)
output = Dense(9, activation=None)(output)
crf = CRF(dtype='float32')
output = crf(output)
base_model = Model(inputs, output)
model = ModelWithCRFLoss(base_model)
model.compile(optimizer='adam')

x = [[5, 2, 3] * 3] * 10
y = [[1, 2, 3] * 3] * 10

model.fit(x=x, y=y, epochs=2, batch_size=2)
tf.keras.models.save_model(model, 'test/1')

Supoort for tensorflow mixed precision training

Currently these is a bug in tensorflow-addons.text.crf, which causes a dtype error when using miex precision. To correctly use mixed precison, you need to modify the line 488 of tensorflow_addons/text/crf.py to:

crf_fwd_cell = CrfDecodeForwardRnnCell(transition_params, dtype=inputs.dtype)

Example

import tensorflow as tf
from tensorflow.keras.layers import Input, Embedding, Bidirectional, GRU, Dense
from tensorflow.keras.models import Model
from tf2crf import CRF, ModelWithCRFLoss
from tensorflow.keras.mixed_precision import experimental as mixed_precision
policy = mixed_precision.Policy('mixed_float16')
mixed_precision.set_policy(policy)

inputs = Input(shape=(None,), dtype='int32')
output = Embedding(100, 40, trainable=True, mask_zero=True)(inputs)
output = Bidirectional(GRU(64, return_sequences=True))(output)
output = Dense(9, activation=None)(output)
crf = CRF(dtype='float32')
output = crf(output)
base_model = Model(inputs, output)
model = ModelWithCRFLoss(base_model)
model.compile(optimizer='adam')

x = [[5, 2, 3] * 3] * 10
y = [[1, 2, 3] * 3] * 10

model.fit(x=x, y=y, epochs=2, batch_size=2)
tf.keras.models.save_model(model, 'test/1')

How to save the model

Currently, Loading the model directly is not supported. So you need to load the model weights instead. For example:

tf.keras.models.save_model(model, '1')
model.load_weights('1/variables/variables')

or

model.save_weights('model.h5')
model.load_weights('model.h5')