Initial commit

app.py

@@ -0,0 +1,66 @@
+import pickle
+import torch
+import numpy as np
+import gradio as gr
+from nltk import word_tokenize, sent_tokenize
+from scipy.stats import shapiro
+from transformers import GPT2LMHeadModel, GPT2TokenizerFast
+
+model = GPT2LMHeadModel.from_pretrained('gpt2-large').to('cuda')
+tokenizer: GPT2TokenizerFast = GPT2TokenizerFast.from_pretrained('gpt2-large')
+
+with open('model.pkl', 'rb') as f:
+    lr_model = pickle.load(f)
+
+def get_perplexity(text: str):
+    tokens = tokenizer(text, return_tensors='pt', truncation=True, return_offsets_mapping=True)
+    inputs = tokens.input_ids.to('cuda')
+    targets = inputs.clone()
+    with torch.no_grad():
+        outputs = model(inputs, labels=targets)
+        labels = targets.to(outputs.logits.device)
+        # Shift so that tokens < n predict n
+        shift_logits = outputs.logits[..., :-1, :].contiguous()
+        shift_labels = labels[..., 1:].contiguous()
+        perplexities = torch.nn.functional.cross_entropy(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1), reduce=False)
+        output = []
+        targets = targets.to('cpu')[0].tolist()
+        # tokens = tokenizer.convert_ids_to_tokens(targets)
+        offsets = tokens.offset_mapping[0].tolist()
+        print(perplexities.to('cpu').tolist())
+        perplexities = perplexities.to('cpu').numpy()
+        perplexities = perplexities / np.max(perplexities)
+        perplexities = perplexities.tolist()
+        print(perplexities)
+        # output.append((text[:offsets[0][1]], 0))
+        # for offset, p in zip(offsets[1:], perplexities):
+        #     output.append((text[offset[0]:offset[1]], p))
+        #     print(type(p))
+        output.append((text[:tokens.word_to_chars(0)[1]], 0))
+        for word_id, p in zip(tokens.word_ids()[1:], perplexities):
+            if word_id == len(output):
+                span = tokens.word_to_chars(word_id)
+                output.append((text[span[0]:span[1]], p))
+        return outputs.loss, output
+
+
+
+def score_text(text):
+    perplexity, word_perplexities = get_perplexity(text)
+    lengths = []
+    for sentence in sent_tokenize(text):
+        lengths.append(len(word_tokenize(sentence)))
+    scores = lr_model.predict_proba([[perplexity.item(), np.mean(lengths), np.std(lengths), shapiro(lengths).pvalue if len(lengths) > 2 else 0.5]])[0]
+
+    return {'Human': scores[0], 'AI': scores[1]}, word_perplexities
+
+sample_text = """
+The Saturn V is a type of rocket that was developed by NASA in the 1960s to support the Apollo program, which aimed to land humans on the Moon. 
+It remains the most powerful rocket ever built, and its five F-1 engines generated more than 7.5 million pounds of thrust at liftoff. 
+The Saturn V was used for all of the Apollo missions to the Moon, as well as the launch of the Skylab space station. 
+Despite its impressive capabilities, the Saturn V was only used for a brief period of time before being retired in 1973. 
+Nevertheless, it remains a landmark achievement in the history of space exploration and a symbol of human ingenuity and determination."""
+
+demo = gr.Interface(fn=score_text, inputs=[gr.Textbox(label="Text to score", lines=5, value=sample_text)], outputs=[gr.Label(), gr.HighlightedText()] )
+
+demo.launch()   

model.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fd5c62c983e46a73eadb4ce7f1d772799f3a2fc9f62c41d7d50beba93fc9140
+size 706

requirements.txt

@@ -0,0 +1,5 @@
+datasets
+transformers
+torch
+nltk
+scipy