Create new file

app.py

@@ -0,0 +1,70 @@
+import gradio as gr
+from gradio.mix import Parallel
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+import os
+from transformers import T5TokenizerFast, T5ForConditionalGeneration
+from transformers import PegasusForConditionalGeneration, PegasusTokenizer
+import pytorch_lightning as pl
+import torch
+import itertools
+import random
+import nltk
+from nltk.tokenize import sent_tokenize
+import requests
+import json
+nltk.download('punkt')
+from fastT5 import export_and_get_onnx_model
+
+
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+
+T5_tokenizer = AutoTokenizer.from_pretrained("jaimin/T5-Large")
+T5_model = export_and_get_onnx_model('jaimin/T5-large')
+
+
+
+def get_paraphrases(text, n_predictions=3, top_k=50, max_length=256, device="cpu"):
+    para = []
+    r = requests.post(
+        url="https://hf.space/embed/jaimin/CWI/+/api/predict",
+        json={"data": [text]},
+    )
+    response = r.json()
+    sentence = response["data"][0]
+    for sent in sent_tokenize(sentence):
+      text = "paraphrase: "+sent + " </s>"
+      encoding = T5_tokenizer.encode_plus(text, padding=True, return_tensors="pt", truncation=True)
+      input_ids, attention_masks = encoding["input_ids"].to(device), encoding["attention_mask"].to(device)
+      model_output = T5_model.generate(
+                                    input_ids=input_ids,attention_mask=attention_masks,
+                                    max_length = 512,
+                                    early_stopping=True,
+                                    num_beams=15,
+                                    num_beam_groups = 3,
+                                    num_return_sequences=n_predictions,
+                                    diversity_penalty = 0.70,
+                                    temperature=0.7)
+      outputs = []
+      for output in model_output:
+        generated_sent = T5_tokenizer.decode(
+            output, skip_special_tokens=True, clean_up_tokenization_spaces=True
+        )
+        if (
+                generated_sent.lower() != sentence.lower()
+                and generated_sent not in outputs
+        ):
+            outputs.append(generated_sent.replace('paraphrasedoutput:', ""))
+      para.append(outputs)
+    print(para)
+    a = list(itertools.product(*para))
+    random.shuffle(a)
+
+    l=[]
+    for i in range(len(a)):
+      l.append(" ".join(a[i]))
+
+    return l
+    
+iface = gr.Interface(fn=get_paraphrases, inputs=[gr.inputs.Textbox(lines=5)],outputs=["text"])
+iface.launch()
+