pszemraj commited on
Commit
8e679c6
1 Parent(s): 08f3fad

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +187 -1
README.md CHANGED
@@ -1,3 +1,189 @@
1
- # README - long-t5-tglobal-base-16384-booksum-V12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2
  - this README was added because there wasn't one
3
  - created 2022-09-09_20-12-57
 
1
+ ---
2
+ tags:
3
+ - summarization
4
+ - summary
5
+ - booksum
6
+ - long-document
7
+ - long-form
8
+ license:
9
+ - apache-2.0
10
+ - bsd-3-clause
11
+ datasets:
12
+ - kmfoda/booksum
13
+ metrics:
14
+ - rouge
15
+ widget:
16
+ - text: large earthquakes along a given fault segment do not occur at random intervals
17
+ because it takes time to accumulate the strain energy for the rupture. The rates
18
+ at which tectonic plates move and accumulate strain at their boundaries are approximately
19
+ uniform. Therefore, in first approximation, one may expect that large ruptures
20
+ of the same fault segment will occur at approximately constant time intervals.
21
+ If subsequent main shocks have different amounts of slip across the fault, then
22
+ the recurrence time may vary, and the basic idea of periodic mainshocks must be
23
+ modified. For great plate boundary ruptures the length and slip often vary by
24
+ a factor of 2. Along the southern segment of the San Andreas fault the recurrence
25
+ interval is 145 years with variations of several decades. The smaller the standard
26
+ deviation of the average recurrence interval, the more specific could be the long
27
+ term prediction of a future mainshock.
28
+ example_title: earthquakes
29
+ - text: " A typical feed-forward neural field algorithm. Spatiotemporal coordinates\
30
+ \ are fed into a neural network that predicts values in the reconstructed domain.\
31
+ \ Then, this domain is mapped to the sensor domain where sensor measurements are\
32
+ \ available as supervision. Class and Section Problems Addressed Generalization\
33
+ \ (Section 2) Inverse problems, ill-posed problems, editability; symmetries. Hybrid\
34
+ \ Representations (Section 3) Computation & memory efficiency, representation\
35
+ \ capacity, editability: Forward Maps (Section 4) Inverse problems Network Architecture\
36
+ \ (Section 5) Spectral bias, integration & derivatives. Manipulating Neural Fields\
37
+ \ (Section 6) Edit ability, constraints, regularization. Table 2: The five classes\
38
+ \ of techniques in the neural field toolbox each addresses problems that arise\
39
+ \ in learning, inference, and control. (Section 3). We can supervise reconstruction\
40
+ \ via differentiable forward maps that transform Or project our domain (e.g, 3D\
41
+ \ reconstruction via 2D images; Section 4) With appropriate network architecture\
42
+ \ choices, we can overcome neural network spectral biases (blurriness) and efficiently\
43
+ \ compute derivatives and integrals (Section 5). Finally, we can manipulate neural\
44
+ \ fields to add constraints and regularizations, and to achieve editable representations\
45
+ \ (Section 6). Collectively, these classes constitute a 'toolbox' of techniques\
46
+ \ to help solve problems with neural fields There are three components in a conditional\
47
+ \ neural field: (1) An encoder or inference function \u20AC that outputs the conditioning\
48
+ \ latent variable 2 given an observation 0 E(0) =2. 2 is typically a low-dimensional\
49
+ \ vector, and is often referred to aS a latent code Or feature code_ (2) A mapping\
50
+ \ function 4 between Z and neural field parameters O: Y(z) = O; (3) The neural\
51
+ \ field itself $. The encoder \u20AC finds the most probable z given the observations\
52
+ \ O: argmaxz P(2/0). The decoder maximizes the inverse conditional probability\
53
+ \ to find the most probable 0 given Z: arg- max P(Olz). We discuss different encoding\
54
+ \ schemes with different optimality guarantees (Section 2.1.1), both global and\
55
+ \ local conditioning (Section 2.1.2), and different mapping functions Y (Section\
56
+ \ 2.1.3) 2. Generalization Suppose we wish to estimate a plausible 3D surface\
57
+ \ shape given a partial or noisy point cloud. We need a suitable prior over the\
58
+ \ sur- face in its reconstruction domain to generalize to the partial observations.\
59
+ \ A neural network expresses a prior via the function space of its architecture\
60
+ \ and parameters 0, and generalization is influenced by the inductive bias of\
61
+ \ this function space (Section 5)."
62
+ example_title: scientific paper
63
+ - text: 'Is a else or outside the cob and tree written being of early client rope
64
+ and you have is for good reasons. On to the ocean in Orange for time. By''s the
65
+ aggregate we can bed it yet. Why this please pick up on a sort is do and also
66
+ M Getoi''s nerocos and do rain become you to let so is his brother is made in
67
+ use and Mjulia''s''s the lay major is aging Masastup coin present sea only of
68
+ Oosii rooms set to you We do er do we easy this private oliiishs lonthen might
69
+ be okay. Good afternoon everybody. Welcome to this lecture of Computational Statistics.
70
+ As you can see, I''m not socially my name is Michael Zelinger. I''m one of the
71
+ task for this class and you might have already seen me in the first lecture where
72
+ I made a quick appearance. I''m also going to give the tortillas in the last third
73
+ of this course. So to give you a little bit about me, I''m a old student here
74
+ with better Bulman and my research centres on casual inference applied to biomedical
75
+ disasters, so that could be genomics or that could be hospital data. If any of
76
+ you is interested in writing a bachelor thesis, a semester paper may be mastathesis
77
+ about this topic feel for reach out to me. you have my name on models and my email
78
+ address you can find in the directory I''d Be very happy to talk about it. you
79
+ do not need to be sure about it, we can just have a chat. So with that said, let''s
80
+ get on with the lecture. There''s an exciting topic today I''m going to start
81
+ by sharing some slides with you and later on during the lecture we''ll move to
82
+ the paper. So bear with me for a few seconds. Well, the projector is starting
83
+ up. Okay, so let''s get started. Today''s topic is a very important one. It''s
84
+ about a technique which really forms one of the fundamentals of data science,
85
+ machine learning, and any sort of modern statistics. It''s called cross validation.
86
+ I know you really want to understand this topic I Want you to understand this
87
+ and frankly, nobody''s gonna leave Professor Mineshousen''s class without understanding
88
+ cross validation. So to set the stage for this, I Want to introduce you to the
89
+ validation problem in computational statistics. So the problem is the following:
90
+ You trained a model on available data. You fitted your model, but you know the
91
+ training data you got could always have been different and some data from the
92
+ environment. Maybe it''s a random process. You do not really know what it is,
93
+ but you know that somebody else who gets a different batch of data from the same
94
+ environment they would get slightly different training data and you do not care
95
+ that your method performs as well. On this training data. you want to to perform
96
+ well on other data that you have not seen other data from the same environment.
97
+ So in other words, the validation problem is you want to quantify the performance
98
+ of your model on data that you have not seen. So how is this even possible? How
99
+ could you possibly measure the performance on data that you do not know The solution
100
+ to? This is the following realization is that given that you have a bunch of data,
101
+ you were in charge. You get to control how much that your model sees. It works
102
+ in the following way: You can hide data firms model. Let''s say you have a training
103
+ data set which is a bunch of doubtless so X eyes are the features those are typically
104
+ hide and national vector. It''s got more than one dimension for sure. And the
105
+ why why eyes. Those are the labels for supervised learning. As you''ve seen before,
106
+ it''s the same set up as we have in regression. And so you have this training
107
+ data and now you choose that you only use some of those data to fit your model.
108
+ You''re not going to use everything, you only use some of it the other part you
109
+ hide from your model. And then you can use this hidden data to do validation from
110
+ the point of you of your model. This hidden data is complete by unseen. In other
111
+ words, we solve our problem of validation.'
112
+ example_title: transcribed audio - lecture
113
+ - text: "Transformer-based models have shown to be very useful for many NLP tasks.\
114
+ \ However, a major limitation of transformers-based models is its O(n^2)O(n 2)\
115
+ \ time & memory complexity (where nn is sequence length). Hence, it's computationally\
116
+ \ very expensive to apply transformer-based models on long sequences n > 512n>512.\
117
+ \ Several recent papers, e.g. Longformer, Performer, Reformer, Clustered attention\
118
+ \ try to remedy this problem by approximating the full attention matrix. You can\
119
+ \ checkout \U0001F917's recent blog post in case you are unfamiliar with these\
120
+ \ models.\nBigBird (introduced in paper) is one of such recent models to address\
121
+ \ this issue. BigBird relies on block sparse attention instead of normal attention\
122
+ \ (i.e. BERT's attention) and can handle sequences up to a length of 4096 at a\
123
+ \ much lower computational cost compared to BERT. It has achieved SOTA on various\
124
+ \ tasks involving very long sequences such as long documents summarization, question-answering\
125
+ \ with long contexts.\nBigBird RoBERTa-like model is now available in \U0001F917\
126
+ Transformers. The goal of this post is to give the reader an in-depth understanding\
127
+ \ of big bird implementation & ease one's life in using BigBird with \U0001F917\
128
+ Transformers. But, before going into more depth, it is important to remember that\
129
+ \ the BigBird's attention is an approximation of BERT's full attention and therefore\
130
+ \ does not strive to be better than BERT's full attention, but rather to be more\
131
+ \ efficient. It simply allows to apply transformer-based models to much longer\
132
+ \ sequences since BERT's quadratic memory requirement quickly becomes unbearable.\
133
+ \ Simply put, if we would have \u221E compute & \u221E time, BERT's attention\
134
+ \ would be preferred over block sparse attention (which we are going to discuss\
135
+ \ in this post).\nIf you wonder why we need more compute when working with longer\
136
+ \ sequences, this blog post is just right for you!\nSome of the main questions\
137
+ \ one might have when working with standard BERT-like attention include:\nDo all\
138
+ \ tokens really have to attend to all other tokens? Why not compute attention\
139
+ \ only over important tokens? How to decide what tokens are important? How to\
140
+ \ attend to just a few tokens in a very efficient way? In this blog post, we will\
141
+ \ try to answer those questions.\nWhat tokens should be attended to? We will give\
142
+ \ a practical example of how attention works by considering the sentence 'BigBird\
143
+ \ is now available in HuggingFace for extractive question answering'. In BERT-like\
144
+ \ attention, every word would simply attend to all other tokens.\nLet's think\
145
+ \ about a sensible choice of key tokens that a queried token actually only should\
146
+ \ attend to by writing some pseudo-code. Will will assume that the token available\
147
+ \ is queried and build a sensible list of key tokens to attend to.\n>>> # let's\
148
+ \ consider following sentence as an example >>> example = ['BigBird', 'is', 'now',\
149
+ \ 'available', 'in', 'HuggingFace', 'for', 'extractive', 'question', 'answering']\n\
150
+ >>> # further let's assume, we're trying to understand the representation of 'available'\
151
+ \ i.e. >>> query_token = 'available' >>> # We will initialize an empty `set` and\
152
+ \ fill up the tokens of our interest as we proceed in this section. >>> key_tokens\
153
+ \ = [] # => currently 'available' token doesn't have anything to attend Nearby\
154
+ \ tokens should be important because, in a sentence (sequence of words), the current\
155
+ \ word is highly dependent on neighboring past & future tokens. This intuition\
156
+ \ is the idea behind the concept of sliding attention."
157
+ example_title: bigbird blog intro
158
+ - text: "To be fair, you have to have a very high IQ to understand Rick and Morty.\
159
+ \ The humour is extremely subtle, and without a solid grasp of theoretical physics\
160
+ \ most of the jokes will go over a typical viewer's head. There's also Rick's\
161
+ \ nihilistic outlook, which is deftly woven into his characterisation- his personal\
162
+ \ philosophy draws heavily from Narodnaya Volya literature, for instance. The\
163
+ \ fans understand this stuff; they have the intellectual capacity to truly appreciate\
164
+ \ the depths of these jokes, to realise that they're not just funny- they say\
165
+ \ something deep about LIFE. As a consequence people who dislike Rick & Morty\
166
+ \ truly ARE idiots- of course they wouldn't appreciate, for instance, the humour\
167
+ \ in Rick's existential catchphrase 'Wubba Lubba Dub Dub,' which itself is a cryptic\
168
+ \ reference to Turgenev's Russian epic Fathers and Sons. I'm smirking right now\
169
+ \ just imagining one of those addlepated simpletons scratching their heads in\
170
+ \ confusion as Dan Harmon's genius wit unfolds itself on their television screens.\
171
+ \ What fools.. how I pity them. \U0001F602\nAnd yes, by the way, i DO have a Rick\
172
+ \ & Morty tattoo. And no, you cannot see it. It's for the ladies' eyes only- and\
173
+ \ even then they have to demonstrate that they're within 5 IQ points of my own\
174
+ \ (preferably lower) beforehand. Nothin personnel kid \U0001F60E"
175
+ example_title: Richard & Mortimer
176
+ parameters:
177
+ max_length: 64
178
+ min_length: 8
179
+ no_repeat_ngram_size: 3
180
+ early_stopping: true
181
+ repetition_penalty: 3.5
182
+ length_penalty: 0.3
183
+ encoder_no_repeat_ngram_size: 3
184
+ num_beams: 4
185
+ ---
186
+ # long-t5-tglobal-base-16384-booksum-V12
187
+
188
  - this README was added because there wasn't one
189
  - created 2022-09-09_20-12-57