profile
viewpoint
If you are wondering where the data of this site comes from, please visit https://api.github.com/users/cbockman/events. GitMemory does not store any data, but only uses NGINX to cache data for a period of time. The idea behind GitMemory is simply to give users a better reading experience.

cbockman/beam 0

Mirror of Apache Beam

cbockman/Bidirectional-LSTM-CRF-for-Clinical-Concept-Extraction 0

Bidirectional LSTM-CRF for Clinical Concept Extraction using i2b2-2010 data

cbockman/ctakes-neural 0

Utilities and interfaces for working with neural networks, especially through ClearTK and Keras

cbockman/fold 0

Deep learning with dynamic computation graphs in TensorFlow

cbockman/forseti-security 0

Forseti Security

cbockman/incubator-airflow 0

Apache Airflow (Incubating)

cbockman/k8s 0

Tools for ML/Tensorflow on Kubernetes.

cbockman/ner-lstm 0

Named Entity Recognition using multilayered bidirectional LSTM

PR opened medicode/tensor2tensor

Reviewers
Rotate Circle CI keys 05/13/2021

Rotating CircleCI Service Account keys

+1 -1

0 comment

1 changed file

pr created time in 4 days

create barnchmedicode/tensor2tensor

branch : millingab-patch-3

created branch time in 4 days

create barnchmedicode/tensor2tensor

branch : pawel/sota

created branch time in 6 days

issue commentgoogle-research/bert

Use cross entropy loss function in the classification task.

NLL loss is just the same as the cross-entropy loss. So actually the model uses cross-entropy.

Realvincentyuan

comment created time in 8 days

issue commentgoogle-research/bert

how the model reflect 'bidirectional'?

@hsm207 It's a great explanation for why is cls leart sentence representation, thanks a lot.

HaodaY

comment created time in 8 days

PR closed medicode/tensor2tensor

Reviewers
Dummy
+1 -1

1 comment

1 changed file

fathom-max

pr closed time in 11 days

pull request commentmedicode/tensor2tensor

Dummy

t2t

fathom-max

comment created time in 11 days

PR opened medicode/tensor2tensor

Dummy
+1 -1

0 comment

1 changed file

pr created time in 11 days

create barnchmedicode/tensor2tensor

branch : dummy-2021-05-05

created branch time in 11 days

PR closed medicode/docker-airflow

Airflow 2 upgrade

https://app.asana.com/0/1188112030623330/1199965798729370/f

+44 -36

0 comment

3 changed files

fathom-pawel

pr closed time in 12 days

create barnchmedicode/tensor2tensor

branch : key-rotation-2021-05-05

created branch time in 12 days

PR opened medicode/tensor2tensor

Update config.yml
+1 -1

0 comment

1 changed file

pr created time in 12 days

push eventmedicode/docker-airflow

Pawel Sienkowski

commit sha 5ed37d8e8eabd3f18b9dfca81acb054b6c5043f3

Adjustment

view details

push time in 14 days

push eventmedicode/docker-airflow

Pawel Sienkowski

commit sha 2b264fc43a2cea0c7b7b6bacd8c24d33320285bc

Update entrypoint

view details

push time in 14 days

issue commentgoogle-research/bert

run run_classifier on colab with TPU got "^C" after the first checkpoint.

Hey,

I got the same output. Did you get the solution ?

TYTYTYTYTYTYTYTYTY

comment created time in 15 days

issue openedgoogle-research/bert

How do BERT-Base, Multilingual Cased and BERT-Base, Uncased have the same number of parameters with different vocabulary sizes?

The readme states that both BERT-Base, Multilingual Cased and BERT-Base Uncased have 110 million parameters.

The following returns 109,482,240 for BERT-Base, Uncased print(sum([param.nelement() for param in model.parameters()]))

However, for BERT-Base, Multilingual Cased, it returns 177,853,440

This difference can be accounted for by the number of embeddings. For BERT-Base, Uncased, the number of embeddings is 30522 and for BERT-Base, Multilingual Cased, the number of embeddings is 119547. Each have a size of 768. 109,482,240 - 30,522 * 768 = 86,041,344 177,853,440 - 119,547 * 768 = 86,041,344

If the embedding parameters are trainable, I think the number of parameters for BERT-Base, Multilingual Cased should be 178 million due to its larger vocabulary size. Is there something here that I am overlooking?

created time in 17 days

create barnchmedicode/tensor2tensor

branch : feature/space-punc-tokens

created branch time in 19 days

issue openedgoogle-research/bert

For news classification long text tasks, BERT fine-tune, loss does not drop, training does not move, fixed classification to one category

For news classification long text tasks, BERT fine-tune, loss does not drop, training does not move, fixed classification to one category,why?

I am very anxious to seek help from Daniel

created time in 20 days

issue commentgoogle-research/bert

How is the number of BERT model parameters calculated?

I think the attention head number is chosen such that H / A = 64 for all models, where H is the hidden size and A is the number of attention heads

aslicedbread

comment created time in 20 days

pull request commentmedicode/tensor2tensor

2021-04-26 SA key rotation

Screen Shot 2021-04-26 at 9 59 53 PM

fathom-matthew

comment created time in 21 days

PR opened medicode/tensor2tensor

Reviewers
2021-04-26 SA key rotation

ASANA TASK LINK: Engineering Rotates Keys: Rotate CircleCI SA keys

DESIGN DOC LINK: Rotate CircleCI GCP Service Account Keys Playbook

DETAILS

Early rotation of CircleCI SA keys

RISK & IMPACT ASSESSMENT: Minimal (standard security procedure)

+1 -1

0 comment

1 changed file

pr created time in 21 days

create barnchmedicode/tensor2tensor

branch : 2021-04-26-sa-key-rotation

created branch time in 21 days

issue commentgoogle-research/bert

请问获得训练好的词向量的值? How to get the pretrained embedding_table?

After get the token_ to_ orig_ Map, how to get the embedding of each word? For example, if a word is divided into three segments, can we directly calculate the average of the three segments?

WHQ1111

comment created time in 21 days

issue commentgoogle-research/bert

The [CLS] token

Is this still valid for RoBERTa ? What does RoBERTa do with this token?

hamzaleroi

comment created time in 22 days

issue commentgoogle-research/bert

The [CLS] token

During pre-training the [CLS] representation is fed into binary classifier which predicts whether the sentence pair <A,B> are concurrent sentences or not. 50% of the time <A,B> are actual next sentences in a large corpus and 50% of the time B is a sentence chosen randomly from other document in a large corpus and therefore isn't a valid continuation A. for eg. <Delhi is capital of India. Penguins don't fly>.

hamzaleroi

comment created time in 22 days

issue commentgoogle-research/bert

Is this piece of code in function 'transformer_model' useless?

I have the same question too.

Traeyee

comment created time in 22 days

issue openedgoogle-research/bert

The [CLS] token

During pre-training, is the [CLS] token ignored when calculating the loss? And if that is the case, doesn't that mean that my model will predict a random token instead of [CLS]?

created time in 22 days

pull request commentgoogle-research/bert

PR to resolve issue #1140 - AttributeError: module 'tensorflow._api.v2.train' has no attribute 'Optimizer'

Changes:

  1. Provided python 2 support for _run_split_on_punc (tokenization.py)
  2. Provided tf2 options in modeling.py and modeling_test.py
abhilash1910

comment created time in 23 days

pull request commentmedicode/tensor2tensor

Rotate SA Key 2021 04 21

Screen Shot 2021-04-21 at 3 16 58 PM Screen Shot 2021-04-21 at 3 17 03 PM

fathom-zh59412

comment created time in a month

PR opened medicode/tensor2tensor

Reviewers
rotate SA kay 2021 04 21
+1 -1

0 comment

1 changed file

pr created time in a month