Meeting Notes 2020/05/08
===

###### tags: `Meeting`

:::info
**Agenda**
1. Plan with HackMD self-hosted instance @mb706?
2. Tasks from last meeting `15min` [name=All]
3. [[Pad for collection of ideas]](https://hackmd.io/XM4QM-yLTY2QWpuFVTjGEw?both)`10min` [name=Philipp]
4. [[Benchmark paper]](https://hackmd.io/@genomenet/BJJfhLZ5I)`10min` [name=Philipp]
5. Weekend: [[Don't Starve Together]](https://www.klei.com/games/dont-starve-together)? `5min` [name=Philipp]

**Participants**
- Julia (JM)
- Martin (MB)
- René (RM)
- Philipp (PM)
:::

:closed_book: Tasks from last meeting
--

- [x] OMM error caused by Tensorboard ggplot confusion matrix and remove this part from *deepG* / RM
- [x] explore _stateful_ LSTM / RM
- [x] move the new negative dataset for CRISPR detection to luna and remove overlap / PM
- [x] Stateful LSTM
- [x] sensible values for batch size and step size
- [x] add callback to reset graph?
- [x] Wavenet
- [x] Hyperparameter tuning
- [x] Stop runs that are not getting better? [TB grid](http://bioinf026:8097/#)

:open_book: New tasks
--

- [ ] post example code that would work on CRISPR binary task for training and inference and sent to JM / RM

## Notes
<!-- Other important details discussed during the meeting can be entered here. -->

### Neg. dataset for CRISPR prediction

$ bowtie2 -p 50 -x all_hmp1-II_reads -f reads_neg.fasta -S alignment.sam --un reads_neg_filtered.fasta
68818787 reads; of these:
  68818787 (100.00%) were unpaired; of these:
    67992983 (98.80%) aligned 0 times
    189394 (0.28%) aligned exactly 1 time
    636410 (0.92%) aligned >1 times
1.20% overall alignment rate

Filtered dataset is here: `/net/scratch/nobackup/pmuench/reads_neg_filtered.fasta` and should contain the 67,992,983 sequences.

### _stateful_ RNN

- `step.size` should be (`lenth(sequence)-1)` or `length(sequence)`?

#### "Batch size 1"

+------------------------------+
|abkajlfjioajfeijaflijfladkjfdx|
+------------------------------+

 Batch 1    Batch 2    Batch 3
+----------+----------+----------+
|abkajlfjio|ajfeijafli|jfladkjfdx|
+----------+----------+----------+

#### "Batch size 2"

seq 1
+------------------------------+
|abkajlfjioajfeijaflijfladkjfdx|
+------------------------------+

seq 2
+------------------------------+
|ewjio;jaijfkaj;fkejak;jdfdjfds|
+------------------------------+

 Batch 1    Batch 2    Batch 3
+----------+----------+----------+
|abkajlfjio|ajfeijafli|jfladkjfdx|
+          +          +          +
|ewjio;jaij|fkaj;fkeja|k;jdfdjfds|
+----------+----------+----------+

#### "Batch size 2" alternative?

+------------------------------+
|abkajlfjioajfeijaflijfladkjfdx|
+------------------------------+

 Batch 1    Batch 2    Batch 3
+----------+----------+----------+
|abkajlfjio|ajfeijafli|jfladkjfdx|
+          +          +          +
|ajlfjioajf|eijaflijfl|adkjfdx...| # shift by x
+----------+----------+----------+

(I think this is not a good idea)

- use small batch size (1-2) and big network sizes

> add callback to reset graph?
- best solution would be to reset as soon as a new file will be used
- maybe just reset after x batches?
- or not reseat it at all

### Wavenet

- try out networks with more parameters (e.g. 10M)
- in TB wave_bifo26_*
- most influenze `initial_kernel` size / `initial_filters`