app.py

import gradio as gr
import torch
import numpy as np

import pandas as pd
from tqdm import tqdm

from transformers import AutoTokenizer, AutoConfig, AutoModel, AutoModelForSequenceClassification


description_sentence = "<h3>Demo EmotioNL</h3>\nThis demo allows you to analyse the emotion in a sentence."
description2 = "<h3>Demo EmotioNL</h3>\nThis demo allows you to analyse the emotions in a dataset.\nThe data should be in tsv-format with two named columns: the first column (id) should contain the sentence IDs, and the second column (text) should contain the actual texts. Optionally, there is a third column named 'date', which specifies the date associated with the text (e.g., tweet date). This column is necessary when the options 'emotion distribution over time' and 'peaks' are selected."

inference_modelpath = "model/checkpoint-128"

"""
output_dir = "model"
model_config = {
    "model_weights": "pdelobelle/robbert-v2-dutch-base",
    "num_labels": 6,
    "max_length": 128,
    "device": "cpu"
}

## Tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_config["model_weights"])
model = AutoModelForSequenceClassification.from_pretrained(inference_modelpath)


# Function for encoding (tokenizing) data
def encode_data(data):
  text = data["text"]
  label = data["label"]
  encoded_input = tokenizer(
                text,
                add_special_tokens=True,
                max_length= model_config["max_length"],
                padding= "max_length",
                return_overflowing_tokens=True,
                truncation=True
            )
  encoded_input["labels"] = label
  return encoded_input


# Test arguments for Trainer
test_args = TrainingArguments(
    output_dir = output_dir,
    do_train = False,
    do_predict = True,
    per_device_eval_batch_size = 64,   
    dataloader_drop_last = False    
)
trainer = Trainer(
              model = model,
              args = test_args)


def inference_dataset(file_object):
  #input_file = open(file_object.name, 'r')
  input_file = file_object
  data_paths = {"train": input_file, "inference": input_file}
  dataset = load_dataset('csv', skiprows=1, data_files=data_paths, column_names = ['id', 'text', 'label'], delimiter='\t')
  encoded_dataset = dataset.map(encode_data, batched=True)
  encoded_dataset.set_format("torch")
  encoded_dataset["inference"] = encoded_dataset["inference"].remove_columns("label")
  # Run trainer in prediction mode
  prediction_output = trainer.predict(encoded_dataset["inference"])
  predictions = prediction_output[0]
  ids = dataset["inference"]["id"]
  texts = dataset["inference"]["text"]
  preds = np.argmax(predictions, axis=1)
  preds = [model.config.id2label[pred] for pred in preds]
  predictions_content = list(zip(ids, texts, preds))
  # write predictions to file
  output = "output.txt"
  f = open(output, 'w')
  f.write("id\ttext\tprediction\n")
  for line in predictions_content:
      f.write(str(line[0]) + '\t' + str(line[1]) + '\t' + str(line[2]) + '\n')
  f.close()
  return output
"""

def inference_dataset(file_object, option_list):
    tokenizer = AutoTokenizer.from_pretrained(inference_modelpath)
    model = AutoModelForSequenceClassification.from_pretrained(inference_modelpath)
    data_path = open(file_object, 'r')
    df = pd.read_csv(data_path, delimiter='\t', header=0, names=['id', 'text'])
    ids = df["id"].tolist()
    texts = df["text"].tolist()
    preds = []
    for text in tqdm(texts): # progressbar
        inputs = tokenizer(text, return_tensors="pt")
        with torch.no_grad(): # run model
            logits = model(**inputs).logits
        predicted_class_id = logits.argmax().item()
        prediction = model.config.id2label[predicted_class_id]
        preds.append(prediction)
    predictions_content = list(zip(ids, texts, preds))
    # write predictions to file
    output = "output.txt"
    f = open(output, 'w')
    f.write("id\ttext\tprediction\n")
    for line in predictions_content:
        f.write(str(line[0]) + '\t' + str(line[1]) + '\t' + str(line[2]) + '\n')
    f.close()
    output1 = output
    output2 = output3 = output4 = output5 = "This option was not selected."
    if "emotion frequencies" in option_list:
        output2 = "This option was selected."
    if "emotion distribution over time" in option_list:
        output3 = "This option was selected."
    if "peaks" in option_list:
        output4 = "This option was selected."
    if "topics" in option_list:
        output5 = "This option was selected."
    return [output1, output2, output3, output4, output5]

def what_happened(text, file_object, option_list):
    if file_object:
        output = "You uploaded a file."
        #if len(option_list) > 0:
            #output = output + "\nYou selected these options:\n- " + "\n- ".join(option_list)
    else:
        output = "Normally, this demo should analyse the emotions in this text:\n" + text
        if len(option_list) > 0:
            output = output + "\nYou can only select options when uploading a dataset."
    return output

def what_happened1(text):
    output = "Normally, this demo should analyse the emotions in this text:\n" + text
    return output

def what_happened2(file_object, option_list):
    #input_file = open(file_object.name, 'r')
    #lines = input_file.read()
    #input_file.close()
    #output_file = open('output.txt', 'w')
    #output_file.write(lines)
    #output_file.close()
    #output1 = 'output.txt'
    output1 = inference_dataset(file_object.name)
    output2 = output3 = output4 = output5 = "This option was not selected."
    if "emotion frequencies" in option_list:
        output2 = "This option was selected."
    if "emotion distribution over time" in option_list:
        output3 = "This option was selected."
    if "peaks" in option_list:
        output4 = "This option was selected."
    if "topics" in option_list:
        output5 = "This option was selected."
    return [output1, output2, output3, output4, output5]

def inference_sentence(text):
    tokenizer = AutoTokenizer.from_pretrained(inference_modelpath)
    model = AutoModelForSequenceClassification.from_pretrained(inference_modelpath)
    inputs = tokenizer(text, return_tensors="pt")
    with torch.no_grad(): # run model
        logits = model(**inputs).logits
        predicted_class_id = logits.argmax().item()
    output = model.config.id2label[predicted_class_id]
    return output
    

iface0 = gr.Interface(
            fn=what_happened,
            inputs=[
                gr.Textbox(
                    label= "Enter a sentence",
                    lines=1,
                    value="Your name"),
                gr.File(
                    label="Or upload a dataset"),
                gr.CheckboxGroup(
                    ["emotion frequencies", "emotion distribution over time", "peaks", "topics"],
                    label = "Select options")
                   ],
            outputs="text")

iface_sentence = gr.Interface(
            fn=inference_sentence,
            description = description_sentence,
            inputs = gr.Textbox(
                    label="Enter a sentence",
                    lines=1),
            outputs="text")

iface2 = gr.Interface(
            fn=inference_dataset,
            description = description2,
            inputs=[
                gr.File(
                    label="Upload a dataset"),
                gr.CheckboxGroup(
                    ["emotion frequencies", "emotion distribution over time", "peaks", "topics"],
                    label = "Select options")
                   ],
            #outputs=["text", "text", "text", "text", "text"])
            outputs = [
                #gr.Textbox(label="Output file"),
                "file",
                gr.Textbox(label="Emotion frequencies"),
                gr.Textbox(label="Emotion distribution over time"),
                gr.Textbox(label="Peaks"),
                gr.Textbox(label="Topics")
            ])

iface = gr.TabbedInterface([iface_sentence, iface2], ["Sentence", "Dataset"])

iface.launch()