Luffy

Results 38 issues of Luffy

![image](https://user-images.githubusercontent.com/49063302/112124869-c9967680-8bfd-11eb-8c83-827fddc57322.png)

`mean, logstd = tf.split(axis=len(flat.shape)-1, num_or_size_splits=2, value=flat)` Why is this problem?

windows下使用 tf.keras,tf== 2.4.0 运行 task_iflytek_gradient_penalty 的例子 当调用 search_layer 函数遇到错误:'Tensor' object has no attribute '_keras_history' 后来发现是 _keras_history 是keras tensor特有的属性,tf tensor 没有 于是改成如下的方式: ![image](https://user-images.githubusercontent.com/49063302/149277034-eab89075-d2e5-4d40-b06e-be05026837d3.png) 能跑起来,但请问苏神这种方法是否也可行,感谢苏神

苏神中午好! 最近clone了您的代码, 准备从零开始感受一下BERT和 如今如此众多基于BERT变化或改进的模型 之间千丝万缕般的联系和惊为天人的设计思想。 然而进行的并不顺利,所以才来向您请教! 关于BERT的输入问题,BERT论文中提及应该是 token,segment,position。 token 是词句映射输入 segment 标记不同句子,是做nsp任务所需要 position 应该是解决做selfattention时没有词句位置信息的问题 但是在bert4keras中只有前两者输入,并没有position的信息,后来看了源码才知道是在内部做了处理,所以不需要。 关于 PositionEmbedding 的实现部分没有看的很懂,以及 merge_mode 的参数,还希望苏神能解惑一下,谢谢苏神!!!

My training images: ![image](https://user-images.githubusercontent.com/49063302/195975438-22f12473-39fc-4a19-abea-410e7bcfd593.png) My regularization images, 100 sheets: ![image](https://user-images.githubusercontent.com/49063302/195975571-dc702657-04e4-4e63-b1c8-f0d03948a8a1.png) Results obtained from prompt "a man wears a sks hat in winter": ![image](https://user-images.githubusercontent.com/49063302/195975371-3c7d67ca-632f-41b3-8658-84163161447a.png) It seems that there is a gap...

My training data ![IMG_0855 HEIC](https://user-images.githubusercontent.com/49063302/189869487-c0f098f1-c74d-4859-bac9-0842970881be.JPG) ![IMG_0856 HEIC](https://user-images.githubusercontent.com/49063302/189869528-e4e40a0a-ab34-4112-9610-3ba098cc04cc.JPG) ![IMG_0857 HEIC](https://user-images.githubusercontent.com/49063302/189869551-e220b655-53b2-4783-8702-7c038a93595d.JPG) ![IMG_0858 HEIC](https://user-images.githubusercontent.com/49063302/189869566-704ce0e4-1ac2-4a8c-8542-d6ce6528199d.JPG) ![IMG_0859 HEIC](https://user-images.githubusercontent.com/49063302/189869583-a08f997d-105e-44c3-9bdf-c17b0cfbd398.JPG) logs ouput images --- Epoch 2 --- ![samples_scaled_gs-000500_e-000002_b-000000](https://user-images.githubusercontent.com/49063302/189870577-0cf6c410-b1e3-4e61-8f62-0e1c3fd6ccdf.jpg) --- Epoch 10 --- ![samples_scaled_gs-002500_e-000010_b-000000](https://user-images.githubusercontent.com/49063302/189870756-5f636cf4-365b-4806-9b73-57237ca60f78.jpg) --- Epoch 20 ---...

对应 task_sequence_labeling_ner_crf.py,有个地方没看懂,想请教一下苏神: ![image](https://user-images.githubusercontent.com/49063302/185287541-9caad0fd-41d1-4e91-8b4e-467c2cece5e6.png) 经过观察,categories 是一个['PER','LOC',‘ORG’] 三类别的List 不太理解,这个label为什么是这样设计的: `labels[start] = categories.index(label) * 2 + 1` `labels[start + 1:end + 1] = categories.index(label) * 2 + 2` 谢谢苏神!!!

After finishing the second picture, original dog lose a lot of detail ![overture-creations-5sI6fQgYIuo](https://user-images.githubusercontent.com/49063302/198818786-aa93dde2-ba2c-4b6e-a8ab-ddfef048afbc.png) ![outpaint_20221027_160103](https://user-images.githubusercontent.com/49063302/198818768-d54bf17a-d379-49e5-a0d0-34cf82e9ee5e.png)