from transformers import GenerationConfig, LlamaForCausalLM, LlamaTokenizer import transformers import numpy as np import torch import os import math import json import ipdb # os.environ["CUDA_VISIBLE_DEVICES"] = "3" import argparse parse = argparse.ArgumentParser() parse.add_argument("--n_fewshot", type=str, default="1024") parse.add_argument("--filename", type=str, default="", help="the filename of the generation results under the ../../results folder") parse.add_argument("--dataset", type=str, default="games") args = parse.parse_args() if torch.cuda.is_available(): device = "cuda" else: device = "cpu" try: if torch.backends.mps.is_available(): device = "mps" except: # noqa: E722 pass base_model = "/your/path/to/llama/hf/model/" tokenizer = LlamaTokenizer.from_pretrained(base_model) model = LlamaForCausalLM.from_pretrained( base_model, torch_dtype=torch.float16, device_map="auto", ) model.half() # seems to fix bugs for some users. model.config.pad_token_id = tokenizer.pad_token_id = 0 # unk model.config.bos_token_id = 1 model.config.eos_token_id = 2 # games movies = np.load(f'../../../data/{args.dataset}/title_maps.npy', allow_pickle=True).item() movie_names = list(movies['seqid2title'].values()) movie_ids = [_ for _ in range(len(movie_names))] movie_dict = dict(zip(movie_names, movie_ids)) result_dict = dict() # ipdb.set_trace() tokenizer.padding_side = "left" def batch(list, batch_size=1): chunk_size = (len(list) - 1) // batch_size + 1 for i in range(chunk_size): yield list[batch_size * i: batch_size * (i + 1)] item_embeddings = [] from tqdm import tqdm model.eval() with torch.no_grad(): for i, name in tqdm(enumerate(batch(movie_names, 4))): input = tokenizer(name, return_tensors="pt", padding=True).to(device) input_ids = input.input_ids attention_mask = input.attention_mask outputs = model(input_ids, attention_mask=attention_mask, output_hidden_states=True) hidden_states = outputs.hidden_states item_embeddings.append(hidden_states[-1][:, -1, :].detach().cpu()) item_embeddings = torch.cat(item_embeddings, dim=0).cuda() path = [f'../../results/{args.filename}.json'] for p in path: result_dict[p] = { "NDCG": [], "HR": [], } model.config.pad_token_id = tokenizer.pad_token_id = 0 # unk model.config.bos_token_id = 1 model.config.eos_token_id = 2 model.eval() f = open(p, 'r') import json test_data = json.load(f) f.close() text = [_["predict"].strip("\"") for _ in test_data] tokenizer.padding_side = "left" def batch(list, batch_size=1): chunk_size = (len(list) - 1) // batch_size + 1 for i in range(chunk_size): yield list[batch_size * i: batch_size * (i + 1)] predict_embeddings = [] from tqdm import tqdm with torch.no_grad(): for i, batch_input in tqdm(enumerate(batch(text, 2))): input = tokenizer(batch_input, return_tensors="pt", padding=True).to(device) input_ids = input.input_ids attention_mask = input.attention_mask outputs = model(input_ids, attention_mask=attention_mask, output_hidden_states=True) hidden_states = outputs.hidden_states predict_embeddings.append(hidden_states[-1][:, -1, :].detach().cpu()) predict_embeddings = torch.cat(predict_embeddings, dim=0).cuda() dist = torch.cdist(predict_embeddings, item_embeddings, p=2) rank = dist.detach().cpu() rank = rank.argsort(dim = -1).argsort(dim = -1) # get the rank list topk_list = [5, 10, 20, 50] NDCG = [] for topk in topk_list: # S = 0 ndcg_sum = 0 for i in range(len(test_data)): idcg = 0 ndcg = 0 S = 0 for idx, t in enumerate(test_data[i]['output']): t = t.strip("\"") t_id = movie_dict[t] rankId = rank[i][t_id].item() if rankId < topk: S = S + (1 / math.log2(rankId + 2)) # todo: math.log2 # idcg if idx+1 <= topk: idcg += (1 / math.log2(idx + 2)) # ndcg ndcg += S / idcg ndcg_sum += ndcg NDCG.append(round(ndcg_sum / len(test_data),4)) RECALL = [] for topk in topk_list: S = 0 for i in range(len(test_data)): s = 0 for t in test_data[i]['output']: t = t.strip("\"") t_id = movie_dict[t] rankId = rank[i][t_id].item() if rankId < topk: s = s + 1 S += s / len(test_data[i]['output']) RECALL.append(round(S / len(test_data), 4)) print(RECALL) print(NDCG) print('_' * 100) result_dict[p]["NDCG"] = NDCG result_dict[p]["RECALL"] = RECALL f = open(f'./results/{args.filename}.json', 'w') json.dump(result_dict, f, indent=4)