做不出来，去开船题

心态稳，才是真的稳，稳则无敌~ 杰克李

transformer torch版本 contributed by Harvard University

恭喜室友找到实习！Envy you but not a bit jealous!

Position Encoding source form CSDN

更改链接：

难得有卡，跑一个bert二分类 。tmd

Examples of inputs and outputs obtained from our 80B parameter model,Flamingo 异常检测，如果不在文件夹，就下载。os.popen()可以用python执行bash指令。很方便；

问author一个问题： chain of Thought: mumu警官

输入：GPT风格的样式。[SEP]

input image:已处理好，等待输入【和dialog history对应】。 decoder也要单独处理，[u1,u2,u3,….] 缺失的图片，直接传None. 困难。如何衔接？如何正确回传参数。我建议在flamingo-mini上做出来。 学到了查看类属性的方法：DataLoader.__dict__

('I just got back from a family reunion.', 'okay[SEP]oh okay', None)
('I just got back from a family reunion.[CLS]okay[SEP]oh okay', 'It was fun.[SEP]It was held at a restaurant.', None)
('I just got back from a family reunion.[CLS]okay[SEP]oh okay[CLS]It was fun.[SEP]It was held at a restaurant.', 'what are you doing now[SEP]okay enjoy', None)
('I just got back from a family reunion.[CLS]okay[SEP]oh okay[CLS]It was fun.[SEP]It was held at a restaurant.[CLS]what are you doing now[SEP]okay enjoy', "I'm just relaxing.[SEP]I took some photos of the reunion.", None)
("I just got back from a family reunion.[CLS]okay[SEP]oh okay[CLS]It was fun.[SEP]It was held at a restaurant.[CLS]what are you doing now[SEP]okay enjoy[CLS]I'm just relaxing.[SEP]I took some photos of the reunion.", 'okay', None)
("I just got back from a family reunion.[CLS]okay[SEP]oh okay[CLS]It was fun.[SEP]It was held at a restaurant.[CLS]what are you doing now[SEP]okay enjoy[CLS]I'm just relaxing.[SEP]I took some photos of the reunion.[CLS]okay", 'oh', None)
("I just got back from a family reunion.[CLS]okay[SEP]oh okay[CLS]It was fun.[SEP]It was held at a restaurant.[CLS]what are you doing now[SEP]okay enjoy[CLS]I'm just relaxing.[SEP]I took some photos of the reunion.[CLS]okay[CLS]Here's one with my aunt Hallie.[SEP]The photo has your aunt Hallie. Objects in the photo: Woman, Wine glass, Clothing, Face[CLS]oh", None, ['The photo has your aunt Hallie. Objects in the photo: Woman, Wine glass, Clothing, Face', 'train/175798036bc27bfb', 'https://c4.staticflickr.com/5/4123/4791319436_99d6d6bf81_o.jpg'])

结构化思考能力： 布鲁克的《批判思维》 麦肯锡的《金字塔原理》 自在不成人，成人不自在。

心灵受到挑战。yk Don’t u think it is carzy?

Dataloader中如果需要变长，需要将序列换成对应的向量： 参考一下S.O. 开始 项目解释一下板子。

---

# -*- encoding: utf-8 -*-
'''
@File    :   utils.py
@Date    :   10-28-2022
@Time    :   00:30:13
@Author  :   lizeyujack@sjtu.edu.cn 
'''
{%raw%}
def merge_img(img):
    '''
    @File    :   utils.py
    @Date    :   10-29-2022
    @Time    :   13:22:20
    @Author  :   lizeyujack@sjtu.edu.cn 
    '''
    batch_size = len([i for i in img])
    padded_img = torch.zeros((batch_size,1,3,224,224))
    for i,img in enumerate(img):
        if img is not None and torch.is_tensor(img):
            padded_img[i] = img
    print(padded_img)
    return padded_img

    return [i for i in img]

def merge_label(sequences):
    lengths = [seq.size()[1] for seq in sequences]
    max_len = 1 if max(lengths) == 0 else max(lengths)
    padded_seqs = torch.ones(len(sequences), max_len).long()
    
    for i, seq in enumerate(sequences):
        end = lengths[i]
        print('end\t',end)
        # print('end\t',end)
        print('seq',seq.shape)
        print('padded_seqs',padded_seqs.shape)
        padded_seqs[i,:end] = seq[:end]
    return padded_seqs, lengths
data.sort(key=lambda x: len(x['dialog_history_input']), reverse=True)
item_info = {}
for key in data[0].keys():
    item_info[key] = [d[key] for d in data]
print(item_info)
# print()
dialog_history_input, dialog_history_input_len = merge(item_info['dialog_history_input'],False)
print(dialog_history_input)
dialog_history_atten_mask,dialog_history_atten_mask_len = merge(item_info['dialog_history_atten_mask'],False)
response_input, response_input_len = merge_label(item_info['response_input'])
response_atten_masks, response_atten_masks_len = merge(item_info['response_atten_mask'],False)
share_images = merge_img(item_info['share_images'])

dialog_history_input = dialog_history_input.contiguous().cuda()
dialog_history_atten_mask = dialog_history_atten_mask.contiguous().cuda()
response_input = response_input.contiguous().cuda()
# share_images = share_images.contiguous().cuda()
data_info = {}
for k in item_info.keys():
    try:
        data_info[k] = locals()[k]
    except:
        data_info[k] = item_info[k]
data_info['dialog_history_input_len'] = dialog_history_input_len
data_info['dialog_history_atten_mask_len'] = dialog_history_atten_mask_len
data_info['response_input_len'] = response_input_len
data_info['response_atten_masks_len'] = response_atten_masks_len
# data_info['share_images_len'] = share_images_len

# print(data_info)
return data_info

Twitter  facebook  Email