Model	F1 macro	F1 macro weighted	Precision macro	Recall macro	Size
seara/rubert-tiny2-ru-go-emotions	0.33	0.48	0.51	0.29	29.2M
seara/rubert-base-cased-ru-go-emotions	0.36	0.49	0.52	0.31	178M
fyaronskiy/ruRoberta-large-ru-go-emotions default thresholds = 0.5	0.41	0.52	0.58	0.36	355M
fyaronskiy/ruRoberta-large-ru-go-emotions best thresholds	0.48	0.58	0.46	0.55	355M
fyaronskiy/deberta-v1-base-russian-go-emotions	0.48	0.57	0.46	0.54	125M

Summary

This is ruRoberta-large model finetuned on ru_go_emotions dataset for multilabel classification. Model can be used to extract all emotions from text or detect certain emotions. Thresholds are selected on validation set by maximizing f1 macro over all labels.

The quality of the model varies greatly across all classes (look at the table with metrics below). There are classes like amusement, gratitude, where the model shows high recognition quality, and classes that pose difficulties for the model - grief, relief, that do have much fewer examples in the training data.

Also ONNX version of model and INT8 quantized model is available. information about them is posted in the last sections.

Usage

Using model with Huggingface Transformers:

python

import torch
from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("fyaronskiy/ruRoberta-large-ru-go-emotions")
model = AutoModelForSequenceClassification.from_pretrained("fyaronskiy/ruRoberta-large-ru-go-emotions")

best_thresholds = [0.36734693877551017, 0.2857142857142857, 0.2857142857142857, 0.16326530612244897, 0.14285714285714285, 0.14285714285714285, 0.18367346938775508, 0.3469387755102041, 0.32653061224489793, 0.22448979591836732, 0.2040816326530612, 0.2857142857142857, 0.18367346938775508, 0.2857142857142857, 0.24489795918367346, 0.7142857142857142, 0.02040816326530612, 0.3061224489795918, 0.44897959183673464, 0.061224489795918366, 0.18367346938775508, 0.04081632653061224, 0.08163265306122448, 0.1020408163265306, 0.22448979591836732, 0.3877551020408163, 0.3469387755102041, 0.24489795918367346]
LABELS = ['admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral']
ID2LABEL = dict(enumerate(LABELS))

Here is how you can extract emotions contained in text:

python

def predict_emotions(text):
  inputs = tokenizer(text, truncation=True, add_special_tokens=True, max_length=128, return_tensors='pt')
  with torch.no_grad():
      logits = model(**inputs).logits
  probas = torch.sigmoid(logits).squeeze(dim=0)  
  class_binary_labels = (probas > torch.tensor(best_thresholds)).int()
  return [ID2LABEL[label_id] for label_id, value in enumerate(class_binary_labels) if value == 1]

print(predict_emotions('У вас отличный сервис и лучший кофе в городе, обожаю вашу кофейню!'))

#['admiration', 'love']

This is the way to get all emotions and their scores:

python

def predict(text):
    inputs = tokenizer(text, truncation=True, add_special_tokens=True, max_length=128, return_tensors='pt')
    with torch.no_grad():
        logits = model(**inputs).logits
    probas = torch.sigmoid(logits).squeeze(dim=0).tolist()
    probas = [round(proba, 3) for proba in probas]    
    
    labels2probas = dict(zip(LABELS, probas))
    probas_dict_sorted = dict(sorted(labels2probas.items(), key=lambda x: x[1], reverse=True))
    return probas_dict_sorted

print(predict('У вас отличный сервис и лучший кофе в городе, обожаю вашу кофейню!'))
'''{'admiration': 0.81,
 'love': 0.538,
 'joy': 0.041,
 'gratitude': 0.031,
 'approval': 0.026,
 'excitement': 0.023,
 'neutral': 0.009,
 'curiosity': 0.006,
 'amusement': 0.005,
 'desire': 0.005,
 'realization': 0.005,
 'caring': 0.004,
 'confusion': 0.004,
 'surprise': 0.004,
 'disappointment': 0.003,
 'disapproval': 0.003,
 'anger': 0.002,
 'annoyance': 0.002,
 'disgust': 0.002,
 'fear': 0.002,
 'grief': 0.002,
 'optimism': 0.002,
 'pride': 0.002,
 'relief': 0.002,
 'sadness': 0.002,
 'embarrassment': 0.001,
 'nervousness': 0.001,
 'remorse': 0.001}
'''

Eval results on test split of ru-go-emotions

	precision	recall	f1-score	support	threshold
admiration	0.63	0.75	0.69	504	0.37
amusement	0.76	0.91	0.83	264	0.29
anger	0.47	0.32	0.38	198	0.29
annoyance	0.33	0.39	0.36	320	0.16
approval	0.27	0.58	0.37	351	0.14
caring	0.32	0.59	0.41	135	0.14
confusion	0.41	0.52	0.46	153	0.18
curiosity	0.45	0.73	0.55	284	0.35
desire	0.54	0.31	0.40	83	0.33
disappointment	0.31	0.34	0.33	151	0.22
disapproval	0.31	0.57	0.40	267	0.20
disgust	0.44	0.40	0.42	123	0.29
embarrassment	0.48	0.38	0.42	37	0.18
excitement	0.29	0.43	0.34	103	0.29
fear	0.56	0.78	0.65	78	0.24
gratitude	0.95	0.85	0.89	352	0.71
grief	0.03	0.33	0.05	6	0.02
joy	0.48	0.58	0.53	161	0.31
love	0.73	0.84	0.78	238	0.45
nervousness	0.24	0.48	0.32	23	0.06
optimism	0.57	0.54	0.56	186	0.18
pride	0.67	0.38	0.48	16	0.04
realization	0.18	0.31	0.23	145	0.08
relief	0.30	0.27	0.29	11	0.10
remorse	0.53	0.84	0.65	56	0.22
sadness	0.56	0.53	0.55	156	0.39
surprise	0.55	0.57	0.56	141	0.35
neutral	0.59	0.79	0.68	1787	0.24
micro avg	0.50	0.66	0.57	6329
macro avg	0.46	0.55	0.48	6329
weighted avg	0.53	0.66	0.58	6329

ONNX and quantized versions of model

Full precision ONNX model (onnx/model.onnx) - 1.5x faster than Transformer model, quality is the same.

INT8 quantized model (onnx/model_quantized.onnx) - 2.5x faster than Transformer model, quality is almost the same.

In table below results of tests of inference of 5427 samples of test_set. I tested inference with batch_size 1 on Intel Xeon CPU with 2 vCPUs (Google Colab).

Model	Size	f1 macro	acceleration	Time of inference
Original model	1.4 GB	0.48	1x	44 min 55 sec
onnx.model	1.4 GB	0.48	1.5x	29 min 52 sec
model_quantized.onnx	0.36 GB	0.48	2.5x	18 min 10 sec

How to use ONNX versions

Loading full precision model

python

from optimum.onnxruntime import ORTModelForSequenceClassification
model_id = "fyaronskiy/ruRoberta-large-ru-go-emotions"
file_name = "onnx/model.onnx"
model = ORTModelForSequenceClassification.from_pretrained(model_id, file_name=file_name)
tokenizer = AutoTokenizer.from_pretrained(model_id)

INT8 quantized model

python

model_id = "fyaronskiy/ruRoberta-large-ru-go-emotions"
file_name = "onnx/model_quantized.onnx"

model = ORTModelForSequenceClassification.from_pretrained(model_id, file_name=file_name)
tokenizer = AutoTokenizer.from_pretrained(model_id)

After loading, using ONNX models for inference is the same as for usual Transformer model:

python

best_thresholds = [0.36734693877551017, 0.2857142857142857, 0.2857142857142857, 0.16326530612244897, 0.14285714285714285, 0.14285714285714285, 0.18367346938775508, 0.3469387755102041, 0.32653061224489793, 0.22448979591836732, 0.2040816326530612, 0.2857142857142857, 0.18367346938775508, 0.2857142857142857, 0.24489795918367346, 0.7142857142857142, 0.02040816326530612, 0.3061224489795918, 0.44897959183673464, 0.061224489795918366, 0.18367346938775508, 0.04081632653061224, 0.08163265306122448, 0.1020408163265306, 0.22448979591836732, 0.3877551020408163, 0.3469387755102041, 0.24489795918367346]
LABELS = ['admiration', 'amusement', 'anger', 'annoyance', 'approval', 'caring', 'confusion', 'curiosity', 'desire', 'disappointment', 'disapproval', 'disgust', 'embarrassment', 'excitement', 'fear', 'gratitude', 'grief', 'joy', 'love', 'nervousness', 'optimism', 'pride', 'realization', 'relief', 'remorse', 'sadness', 'surprise', 'neutral']
ID2LABEL = dict(enumerate(LABELS))

def predict_emotions(text):
  inputs = tokenizer(text, truncation=True, add_special_tokens=True, max_length=128, return_tensors='pt')
  with torch.no_grad():
      logits = model(**inputs).logits
  probas = torch.sigmoid(logits).squeeze(dim=0)  
  class_binary_labels = (probas > torch.tensor(best_thresholds)).int()
  return [ID2LABEL[label_id] for label_id, value in enumerate(class_binary_labels) if value == 1]

print(predict_emotions('У вас отличный сервис и лучший кофе в городе, обожаю вашу кофейню!'))
#['admiration', 'love']