import gradio as gr import os import torch import transformers import huggingface_hub import datetime import json import shutil import threading device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # To suppress the following warning: # huggingface/tokenizers: The current process just got forked, after parallelism has already been used. Disabling parallelism to avoid deadlocks... os.environ["TOKENIZERS_PARALLELISM"] = "false" HF_TOKEN_DOWNLOAD = os.environ['HF_TOKEN_DOWNLOAD'] HF_TOKEN_UPLOAD = os.environ['HF_TOKEN_UPLOAD'] MODE = os.environ['MODE'] # 'debug' or 'prod' MODEL_NAME = 'liujch1998/vera' DATASET_REPO_URL = "https://huggingface.co/datasets/liujch1998/cd-pi-dataset" DATA_DIR = 'data' DATA_FILENAME = 'data.jsonl' if MODE != 'debug' else 'data_debug.jsonl' DATA_PATH = os.path.join(DATA_DIR, DATA_FILENAME) class Interactive: def __init__(self): self.tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME, use_auth_token=HF_TOKEN_DOWNLOAD) if MODE == 'debug': return self.model = transformers.T5EncoderModel.from_pretrained(MODEL_NAME, use_auth_token=HF_TOKEN_DOWNLOAD, low_cpu_mem_usage=True, device_map='auto', torch_dtype='auto', offload_folder='offload') self.model.D = self.model.shared.embedding_dim self.linear = torch.nn.Linear(self.model.D, 1, dtype=self.model.dtype).to(device) self.linear.weight = torch.nn.Parameter(self.model.shared.weight[32099, :].unsqueeze(0)) # (1, D) self.linear.bias = torch.nn.Parameter(self.model.shared.weight[32098, 0].unsqueeze(0)) # (1) self.model.eval() self.t = self.model.shared.weight[32097, 0].item() def run(self, statement): if MODE == 'debug': return { 'timestamp': datetime.datetime.now().strftime('%Y%m%d-%H%M%S'), 'statement': statement, 'logit': 0.0, 'logit_calibrated': 0.0, 'score': 0.5, 'score_calibrated': 0.5, } input_ids = self.tokenizer.batch_encode_plus([statement], return_tensors='pt', padding='longest', truncation='longest_first', max_length=128).input_ids.to(device) with torch.no_grad(): output = self.model(input_ids) last_hidden_state = output.last_hidden_state.to(device) # (B=1, L, D) hidden = last_hidden_state[0, -1, :] # (D) logit = self.linear(hidden).squeeze(-1) # () logit_calibrated = logit / self.t score = logit.sigmoid() score_calibrated = logit_calibrated.sigmoid() return { 'timestamp': datetime.datetime.now().strftime('%Y%m%d-%H%M%S'), 'statement': statement, 'logit': logit.item(), 'logit_calibrated': logit_calibrated.item(), 'score': score.item(), 'score_calibrated': score_calibrated.item(), } def runs(self, statements): if MODE == 'debug': return [{ 'timestamp': datetime.datetime.now().strftime('%Y%m%d-%H%M%S'), 'statement': statement, 'logit': 0.0, 'logit_calibrated': 0.0, 'score': 0.5, 'score_calibrated': 0.5, } for statement in statements] tok = self.tokenizer.batch_encode_plus(statements, return_tensors='pt', padding='longest') input_ids = tok.input_ids.to(device) attention_mask = tok.attention_mask.to(device) with torch.no_grad(): output = self.model(input_ids=input_ids, attention_mask=attention_mask) last_indices = attention_mask.sum(dim=1, keepdim=True) - 1 # (B, 1) last_indices = last_indices.unsqueeze(-1).expand(-1, -1, self.model.D) # (B, 1, D) last_hidden_state = output.last_hidden_state.to(device) # (B, L, D) hidden = last_hidden_state.gather(dim=1, index=last_indices).squeeze(1) # (B, D) logits = self.linear(hidden).squeeze(-1) # (B) logits_calibrated = logits / self.t scores = logits.sigmoid() scores_calibrated = logits_calibrated.sigmoid() return [{ 'timestamp': datetime.datetime.now().strftime('%Y%m%d-%H%M%S'), 'statement': statement, 'logit': logit.item(), 'logit_calibrated': logit_calibrated.item(), 'score': score.item(), 'score_calibrated': score_calibrated.item(), } for statement, logit, logit_calibrated, score, score_calibrated in zip(statements, logits, logits_calibrated, scores, scores_calibrated)] interactive = Interactive() try: shutil.rmtree(DATA_DIR) except: pass global repo, lock repo = huggingface_hub.Repository( local_dir=DATA_DIR, clone_from=DATASET_REPO_URL, token=HF_TOKEN_UPLOAD, repo_type='dataset', ) repo.git_pull() lock = threading.Lock() # def predict(statement, do_save=True): # output_raw = interactive.run(statement) # output = { # 'True': output_raw['score_calibrated'], # 'False': 1 - output_raw['score_calibrated'], # } # if do_save: # with open(DATA_PATH, 'a') as f: # json.dump(output_raw, f, ensure_ascii=False) # f.write('\n') # commit_url = repo.push_to_hub() # print('Logged statement to dataset:') # print('Commit URL:', commit_url) # print(output_raw) # print() # return output, output_raw, gr.update(visible=False), gr.update(visible=True), gr.update(visible=True), gr.update(value='Please provide your feedback before trying out another statement.') # def record_feedback(output_raw, feedback, do_save=True): # if do_save: # output_raw.update({ 'feedback': feedback }) # with open(DATA_PATH, 'a') as f: # json.dump(output_raw, f, ensure_ascii=False) # f.write('\n') # commit_url = repo.push_to_hub() # print('Logged feedback to dataset:') # print('Commit URL:', commit_url) # print(output_raw) # print() # return gr.update(visible=True), gr.update(visible=False), gr.update(visible=False), gr.update(value='Thanks for your feedback! Now you can enter another statement.') # def record_feedback_agree(output_raw, do_save=True): # return record_feedback(output_raw, 'agree', do_save) # def record_feedback_disagree(output_raw, do_save=True): # return record_feedback(output_raw, 'disagree', do_save) def predict(statements, do_saves): global lock, interactive output_raws = interactive.runs(list(statements)) # statements is a tuple, but tokenizer takes a list outputs = [{ 'True': output_raw['score_calibrated'], 'False': 1 - output_raw['score_calibrated'], } for output_raw in output_raws] print(f'Logging statements to {DATA_FILENAME}:') lock.acquire() for output_raw, do_save in zip(output_raws, do_saves): if do_save: print(output_raw) with open(DATA_PATH, 'a') as f: json.dump(output_raw, f, ensure_ascii=False) f.write('\n') print() lock.release() return outputs, output_raws, \ [gr.update(visible=False) for _ in statements], \ [gr.update(visible=True) for _ in statements], \ [gr.update(visible=True) for _ in statements], \ [gr.update(visible=True) for _ in statements], \ [gr.update(visible=True) for _ in statements], \ [gr.update(value='Please share your feedback before trying out another statement.') for _ in statements] def record_feedback(output_raws, feedback, do_saves): global lock print(f'Logging feedbacks to {DATA_FILENAME}:') lock.acquire() for output_raw, do_save in zip(output_raws, do_saves): if do_save: output_raw.update({ 'feedback': feedback }) print(output_raw) with open(DATA_PATH, 'a') as f: json.dump(output_raw, f, ensure_ascii=False) f.write('\n') print() lock.release() return [gr.update(visible=True) for _ in output_raws], \ [gr.update(visible=False) for _ in output_raws], \ [gr.update(visible=False) for _ in output_raws], \ [gr.update(visible=False) for _ in output_raws], \ [gr.update(visible=False) for _ in output_raws], \ [gr.update(value='Thanks for sharing your feedback! You can now enter another statement.') for _ in output_raws] def record_feedback_agree(output_raws, do_saves): return record_feedback(output_raws, 'agree', do_saves) def record_feedback_disagree(output_raws, do_saves): return record_feedback(output_raws, 'disagree', do_saves) def record_feedback_uncertain(output_raws, do_saves): return record_feedback(output_raws, 'uncertain', do_saves) def record_feedback_outofscope(output_raws, do_saves): return record_feedback(output_raws, 'outofscope', do_saves) def push(): global repo, lock lock.acquire() if repo.is_repo_clean(): # print('No new data recorded, skipping git push ...') # print() pass else: try: commit_url = repo.push_to_hub() except Exception as e: print('Failed to push to git:', e) shutil.rmtree(DATA_DIR) repo = huggingface_hub.Repository( local_dir=DATA_DIR, clone_from=DATASET_REPO_URL, token=HF_TOKEN_UPLOAD, repo_type='dataset', ) repo.git_pull() lock.release() examples = [ # # openbookqa # 'If a person walks in the opposite direction of a compass arrow they are walking south.', # 'If a person walks in the opposite direction of a compass arrow they are walking north.', # arc_easy 'A pond is different from a lake because ponds are smaller and shallower.', 'A pond is different from a lake because ponds have moving water.', # arc_hard 'Hunting strategies are more likely to be learned rather than inherited.', 'A spotted coat is more likely to be learned rather than inherited.', # ai2_science_elementary 'Photosynthesis uses carbon from the air to make food for plants.', 'Respiration uses carbon from the air to make food for plants.', # ai2_science_middle 'The barometer measures atmospheric pressure.', 'The thermometer measures atmospheric pressure.', # commonsenseqa 'People aim to complete a job at work.', 'People aim to kill animals at work.', # qasc 'Climate is generally described in terms of local weather conditions.', 'Climate is generally described in terms of forests.', # physical_iqa 'ice box will turn into a cooler if you add water to it.', 'ice box will turn into a cooler if you add soda to it.', # social_iqa 'Kendall opened their mouth to speak and what came out shocked everyone. Kendall is a very aggressive and talkative person.', 'Kendall opened their mouth to speak and what came out shocked everyone. Kendall is a very quiet person.', # winogrande_xl 'Sarah was a much better surgeon than Maria so Maria always got the easier cases.', 'Sarah was a much better surgeon than Maria so Sarah always got the easier cases.', # com2sense_paired 'If you want a quick snack, getting one banana would be a good choice generally.', 'If you want a snack, getting twenty bananas would be a good choice generally.', # sciq 'Each specific polypeptide has a unique linear sequence of amino acids.', 'Each specific polypeptide has a unique linear sequence of fatty acids.', # quarel 'Tommy glided across the marble floor with ease, but slipped and fell on the wet floor because wet floor has more resistance.', 'Tommy glided across the marble floor with ease, but slipped and fell on the wet floor because marble floor has more resistance.', # quartz 'If less waters falls on an area of land it will cause less plants to grow in that area.', 'If less waters falls on an area of land it will cause more plants to grow in that area.', # cycic_mc 'In U.S. spring, Rob visits the financial district every day. In U.S. winter, Rob visits the park every day. Rob will go to the park on January 20.', 'In U.S. spring, Rob visits the financial district every day. In U.S. winter, Rob visits the park every day. Rob will go to the financial district on January 20.', # comve_a 'Summer in North America is great for swimming, boating, and fishing.', 'Summer in North America is great for skiing, snowshoeing, and making a snowman.', # csqa2 'Gas is always capable of turning into liquid under high pressure.', 'Cotton candy is sometimes made out of cotton.', # symkd_anno 'James visits a famous landmark. As a result, James learns about the world.', 'Cliff and Andrew enter the castle. But before, Cliff needed to have been a student at the school.', # gengen_anno 'Generally, bar patrons are capable of taking care of their own drinks.', 'Generally, ocean currents have little influence over storm intensity.', # 'If A sits next to B and B sits next to C, then A must sit next to C.', # 'If A sits next to B and B sits next to C, then A might not sit next to C.', ] # input_statement = gr.Dropdown(choices=examples, label='Statement:') # input_model = gr.Textbox(label='Commonsense statement verification model:', value=MODEL_NAME, interactive=False) # output = gr.outputs.Label(num_top_classes=2) # description = '''This is a demo for Vera, a commonsense statement verification model. Under development. # ⚠️ Data Collection: by default, we are collecting the inputs entered in this app to further improve and evaluate the model. Do not share any personal or sensitive information while using the app!''' # gr.Interface( # fn=predict, # inputs=[input_statement, input_model], # outputs=output, # title="Vera", # description=description, # ).launch() with gr.Blocks() as demo: with gr.Column(): gr.Markdown( '''# Vera Vera is a commonsense statement verification model. See our paper at: [https://arxiv.org/abs/2305.03695](https://arxiv.org/abs/2305.03695). Type a commonsense statement in the box below and click the submit button to see Vera's prediction on its correctness. You can try both correct and incorrect statements. If you are looking for inspiration, try the examples at the bottom of the page! We'd love your feedback! Please indicate whether you agree or disagree with Vera's prediction (and don't mind the percentage numbers). If you're unsure or the statement doesn't have a certain correctness label, please select "Uncertain". If your input is actually not a statement about commonsense, please select "I don't think this is a statement about commonsense". **Intended Use**: Vera is a research prototype and may make mistakes. Do not use for making critical decisions. It is intended to predict the correctness of commonsense statements, and may be unreliable when taking input out of this scope. **DO NOT input encyclopedic facts.** Vera is trained on **English** data only, please do not input statements in other languages. **Data Collection**: By default, we are collecting the inputs entered in this app to further improve and evaluate the model. Do not share any personal or sensitive information while using the app! You can opt out of this data collection by removing the checkbox below: ''' ) with gr.Row(): with gr.Column(scale=3): do_save = gr.Checkbox( value=True, label="Store data", info="You agree to the storage of your input for research and development purposes:") statement = gr.Textbox(placeholder='Enter a commonsense statement here, or select an example from below', label='Statement', interactive=True) submit = gr.Button(value='Submit', variant='primary', visible=True) with gr.Column(scale=2): output = gr.Label(num_top_classes=2, interactive=False) output_raw = gr.JSON(visible=False) with gr.Row(): feedback_agree = gr.Button(value='👍 Agree', variant='secondary', visible=False) feedback_uncertain = gr.Button(value='🤔 Uncertain', variant='secondary', visible=False) feedback_disagree = gr.Button(value='👎 Disagree', variant='secondary', visible=False) feedback_outofscope = gr.Button(value='🚫 I don\'t think this a statement about commonsense', variant='secondary', visible=False) feedback_ack = gr.Markdown(value='', visible=True, interactive=False) gr.Markdown('\n---\n') with gr.Row(): gr.Examples( examples=examples, fn=predict, inputs=[statement], outputs=[output, output_raw, statement, submit, feedback_agree, feedback_disagree, feedback_ack], examples_per_page=100, cache_examples=False, run_on_click=False, # If we want this to be True, I suspect we need to enable the statement.submit() ) submit.click(predict, inputs=[statement, do_save], outputs=[output, output_raw, submit, feedback_agree, feedback_uncertain, feedback_disagree, feedback_outofscope, feedback_ack], batch=True, max_batch_size=16) # statement.submit(predict, inputs=[statement], outputs=[output, output_raw]) feedback_agree.click(record_feedback_agree, inputs=[output_raw, do_save], outputs=[submit, feedback_agree, feedback_uncertain, feedback_disagree, feedback_outofscope, feedback_ack], batch=True, max_batch_size=16) feedback_uncertain.click(record_feedback_uncertain, inputs=[output_raw, do_save], outputs=[submit, feedback_agree, feedback_uncertain, feedback_disagree, feedback_outofscope, feedback_ack], batch=True, max_batch_size=16) feedback_disagree.click(record_feedback_disagree, inputs=[output_raw, do_save], outputs=[submit, feedback_agree, feedback_uncertain, feedback_disagree, feedback_outofscope, feedback_ack], batch=True, max_batch_size=16) feedback_outofscope.click(record_feedback_outofscope, inputs=[output_raw, do_save], outputs=[submit, feedback_agree, feedback_uncertain, feedback_disagree, feedback_outofscope, feedback_ack], batch=True, max_batch_size=16) demo.load(push, inputs=None, outputs=None, every=60) # Push to git every 60 seconds demo.queue(concurrency_count=1).launch(debug=True)