import json from pathlib import Path import gradio as gr import torch from torch.nn import functional as F from torch.utils.data import DataLoader from common import setup_cpu from models import build_tokenizer, build_model from models.meta_optimizer import AttnOptimWrapper from tasks import load_task from tasks.loader import TokenizedForMCRightPad DISPLAY_MAPPING = { "sst2": {"positive": "Pos", "negative": "Neg"}, } @torch.no_grad() def do_infer_probs(model, exemplar_attn_kv, exemplar_attn_mask, batched_choices_input): batched_choices_logprobs = [] for batched_one_choice_input in batched_choices_input: ( batch_input_ids, batch_attention_mask, batch_choice_start, batch_choice_end, ) = batched_one_choice_input bs = len(batch_input_ids) merged_attn_mask = torch.cat((exemplar_attn_mask.expand(bs, -1), batch_attention_mask), dim=1) # [B, #Heads, Length, Hidden] expand_exemplar_attn_kv = [[layer_k.expand((bs, -1, -1, -1)), layer_v.expand((bs, -1, -1, -1))] for layer_k, layer_v in exemplar_attn_kv] batched_logits = model( input_ids=batch_input_ids, # [B, L'] attention_mask=merged_attn_mask, # [B, L + L'] past_key_values=expand_exemplar_attn_kv, # num_layers * 2 * [B, num_heads, L, H] ).logits batched_output = F.log_softmax(batched_logits, dim=-1) # [B, L', Vocab] batched_one_choice_logprobs = [] for input_ids, choice_start, choice_end, lm_logprobs in zip(batch_input_ids, batch_choice_start, batch_choice_end, batched_output): choice_tokens = input_ids[choice_start:choice_end].unsqueeze(1) # [L, 1] choice_logprobs = lm_logprobs[choice_start - 1 : choice_end - 1] # [L, Vocab] extracted = torch.gather(choice_logprobs, -1, choice_tokens).squeeze(-1) choice_length = choice_end - choice_start lm_log_p = torch.sum(extracted).item() norm_lm_log_p = (lm_log_p / choice_length).item() choice_lm_info = {"lm_log_p": lm_log_p, "norm_lm_log_p": norm_lm_log_p} batched_one_choice_logprobs.append(choice_lm_info) batched_choices_logprobs.append(batched_one_choice_logprobs) return batched_choices_logprobs @torch.no_grad() def process_once(dataset_name, exemplar_str, forward_steps, raw_data): setup_cpu(seed=seed) TaskHandler = load_task(dataset_name) task_agent = TaskHandler(prompt_version) processed_data = task_agent.dataset_preprocess(raw_data) dataset = TokenizedForMCRightPad(processed_data, tokenizer, task_agent.multiple_choice_promptify) exemplar_input_ids, exemplar_attn_mask = dataset.tokenize_demonstration(exemplar_str) loader = DataLoader(dataset, shuffle=False, drop_last=False, batch_size=1) meta_optim = AttnOptimWrapper(model, model_name, step_size=step_size, momentum=momentum) meta_optim.init() for _ in range(forward_steps): exemplar_kv = meta_optim.step(exemplar_input_ids) generated_info = [] # question * [choice0_prob, choice1_prob] for batch_input in loader: batch_output = do_infer_probs(model, exemplar_kv, exemplar_attn_mask.unsqueeze(0), batch_input) # [batch_of_choice0, batch_of_choice1, ...] zipped_logprobs = list(zip(*batch_output)) # batch * (choice0, choice1, ...) generated_info.extend(zipped_logprobs) all_predicted = [] num_correct = 0 for idx, (data, choice_info) in enumerate(zip(processed_data, generated_info)): merged_choice_info = task_agent.merge_choice_info(choice_info) merged_predictions_idx = task_agent.choice_info_to_predictions(merged_choice_info)["lm_log_p"] predicted = task_agent.CHOICES[merged_predictions_idx] ground_truth = task_agent.CHOICES[data["answer_idx"]] res = f"{DISPLAY_MAPPING[dataset_name][predicted]}" if predicted == ground_truth: res += " β " num_correct += 1 else: res += " β" all_predicted.append(res) all_predicted.append(f"{100*num_correct / len(all_predicted):.2f}%") return all_predicted def transpose(l): return list(map(list, zip(*l))) def button_pressed(prev_state): dataset_name = prev_state["dataset_name"] exemplar_str = prev_state["exemplar_str"] forward_steps = prev_state["step"] + 2 raw_data = prev_state["raw_data"] prev_table_data = prev_state["table_data"] current_output = process_once(dataset_name, exemplar_str, forward_steps, raw_data) t_prev = transpose(prev_table_data) if forward_steps == 1: t_prev.append(["**ICL**"] + current_output) else: t_prev.append([f"**Step={forward_steps}**"] + current_output) updated_table_data = transpose(t_prev) ret = [ { "dataset_name": dataset_name, "exemplar_str": exemplar_str, "raw_data": raw_data, "step": forward_steps, "table_data": updated_table_data, }, f"Click here to train LLM ! Now Step: {forward_steps}", updated_table_data, ] return ret if __name__ == "__main__": dataset_name = "sst2" seed = 0 prompt_version = "default" kv_iter = 10 model_name, model_size = "opt", "125m" step_size, momentum = 0.01, 0.9 setup_cpu(seed=seed) tokenizer = build_tokenizer(model_name, model_size, padding_side="right") model = build_model(model_name, model_size, False) torch.autograd.set_grad_enabled(False) print(f"Dataset: {dataset_name}") task_root = Path("example_sets").joinpath(dataset_name) with task_root.joinpath("demos.txt").open("r") as f: demos = f.read() with task_root.joinpath("sample.pkl").open("r") as f: raw_data = json.load(f) icl_result = process_once(dataset_name, demos, 1, raw_data) text = """We utilize a Large Language Model (LLM) to perform in-context learning (ICL) for sentiment classification of movie reviews. Taking the following two labeled examples as demonstrations, we predict the sentiment of the subsequent test input. Directly employing ICL results in lower prediction accuracy. However, in our proposed approach, **Deep-Thinking**, we repeatedly apply **Forward Tuning**, leading to improved accuracy of the model.""" css = """ #the-table { overflow: auto; } #the-table > div:nth-child(2) { margin: auto; width: fit-content; } #the-table > div > div > div > table { width: auto; margin: 0; white-space: normal; } #the-table > div > div > div > table > thead {display: none} #the-table > div > div > div > table > tbody > tr:last-child {background-color: beige} #the-table > div > div > div > table > tbody > tr:first-child {background-color: lightgray} #the-table > div > div > div > table > tbody > tr > td {padding: 0 2px;} #the-table > div > div > div > table > tbody > tr > td:first-child {min-width: 300px;} #the-table > div > div > div > table > tbody > tr > td:not(:first-child) {white-space: nowrap; } #the-text { font-size: large; } #main-button { max-width: 500px; margin: 0 auto; } """ title = "π€ Iterative Forward Tuning Boosts In-context Learning in Language Models" demo = gr.Blocks(css=css, title="π€Deep-Thinking") with demo: gr.Markdown(f"