Spaces:
Runtime error
Runtime error
| import os | |
| import csv | |
| import json | |
| import torch | |
| import argparse | |
| import pandas as pd | |
| import torch.nn as nn | |
| from tqdm import tqdm | |
| from collections import defaultdict | |
| from transformers.models.llama.tokenization_llama import LlamaTokenizer | |
| from torch.utils.data import DataLoader | |
| from mplug_owl_video.modeling_mplug_owl import MplugOwlForConditionalGeneration | |
| from mplug_owl_video.processing_mplug_owl import MplugOwlImageProcessor, MplugOwlProcessor | |
| from peft import LoraConfig, get_peft_model | |
| from data_utils.xgpt3_dataset import MultiModalDataset | |
| from utils import batchify | |
| softmax = nn.Softmax(dim=2) | |
| def get_entail(logits, input_ids, tokenizer): | |
| logits = softmax(logits) | |
| token_id_yes = tokenizer.encode('Yes', add_special_tokens = False)[0] | |
| token_id_no = tokenizer.encode('No', add_special_tokens = False)[0] | |
| entailment = [] | |
| for j in range(len(logits)): | |
| for i in range(len(input_ids[j])): | |
| if input_ids[j][i] == tokenizer.pad_token_id: # pad token if the answer is not present | |
| i = i - 1 | |
| break | |
| elif i == len(input_ids[j]) - 1: | |
| break | |
| score = logits[j][i][token_id_yes] / (logits[j][i][token_id_yes] + logits[j][i][token_id_no]) | |
| entailment.append(score) | |
| entailment = torch.stack(entailment) | |
| return entailment | |
| def get_scores(model, tokenizer, dataloader): | |
| with torch.no_grad(): | |
| for index, inputs in tqdm(enumerate(dataloader)): | |
| for k, v in inputs.items(): | |
| if torch.is_tensor(v): | |
| if v.dtype == torch.float: | |
| inputs[k] = v.bfloat16() | |
| inputs[k] = inputs[k].to(model.device) | |
| outputs = model(pixel_values = inputs['pixel_values'], video_pixel_values = inputs['video_pixel_values'], labels = None, \ | |
| num_images = inputs['num_images'], num_videos = inputs['num_videos'], input_ids = inputs['input_ids'], non_padding_mask = inputs['non_padding_mask'], \ | |
| non_media_mask = inputs['non_media_mask'], prompt_mask = inputs['prompt_mask']) | |
| logits = outputs['logits'] | |
| entail_scores = get_entail(logits, inputs['input_ids'], tokenizer) | |
| return entail_scores[0].item() |