thomwolf
4bc4c69af9
finetuning any BERT model - fixes #455
2019-04-11 16:57:59 +02:00
jeonsworld
60005f464d
Update pregenerate_training_data.py
...
If the value of rand_end is returned from the randint function, the value of sampled_doc_index that matches current_idx is returned from searchsorted.
example:
cumsum_max = {int64} 30
doc_cumsum = {ndarray} [ 5 7 11 19 30]
doc_lengths = {list} <class 'list'>: [5, 2, 4, 8, 11]
if current_idx = 1,
rand_start = 7
rand_end = 35
sentence_index = randint(7, 35) % cumsum_max
if randint return 35, sentence_index becomes 5.
if sentence_index is 5, np.searchsorted returns 1 equal to current_index.
2019-03-30 14:50:17 +09:00
thomwolf
361aff6de5
typos
2019-03-27 11:54:59 +01:00
thomwolf
cea8ba1d59
adjusted formating and some wording in the readme
2019-03-27 11:53:44 +01:00
Matthew Carrigan
24e67fbf75
Minor README update
2019-03-25 12:33:30 +00:00
Matthew Carrigan
8d1d1ffde2
Corrected the displayed loss when gradient_accumulation_steps > 1
2019-03-25 12:15:19 +00:00
Matthew Carrigan
abb7d1ff6d
Added proper context management to ensure cleanup happens in the right
...
order.
2019-03-21 17:50:03 +00:00
Matthew Carrigan
06a30cfdf3
Added a --reduce_memory option to the training script to keep training
...
data on disc as a memmap rather than in memory
2019-03-21 17:04:12 +00:00
Matthew Carrigan
7d1ae644ef
Added a --reduce_memory option to the training script to keep training
...
data on disc as a memmap rather than in memory
2019-03-21 17:02:18 +00:00
Matthew Carrigan
2bba7f810e
Added a --reduce_memory option to shelve docs to disc instead of keeping them in memory.
2019-03-21 16:50:16 +00:00
Matthew Carrigan
8733ffcb5e
Removing a couple of other old unnecessary comments
2019-03-21 14:09:57 +00:00
Matthew Carrigan
8a861048dd
Fixed up the notes on a possible future low-memory path
2019-03-21 14:08:39 +00:00
Matthew Carrigan
a8a577ba93
Reduced memory usage for pregenerating the data a lot by writing it
...
out on the fly without shuffling - the Sampler in the finetuning script
will shuffle for us.
2019-03-21 14:05:52 +00:00
Matthew Carrigan
0ae59e662d
Reduced memory usage for pregenerating the data a lot by writing it
...
out on the fly without shuffling - the Sampler in the finetuning script
will shuffle for us.
2019-03-21 14:04:17 +00:00
Matthew Carrigan
6a9038ba53
Removed an old irrelevant comment
2019-03-21 13:36:41 +00:00
Matthew Carrigan
29a392fbcf
Small README changes
2019-03-20 17:35:17 +00:00
Matthew Carrigan
832b2b0058
Adding README
2019-03-20 17:31:49 +00:00
Matthew Carrigan
934d3f4d2f
Syncing up argument names between the scripts
2019-03-20 17:23:23 +00:00
Matthew Carrigan
f19ba35b2b
Move old finetuning script into the new folder
2019-03-20 16:47:06 +00:00
Matthew Carrigan
7de5c6aa5e
PEP8 and formatting cleanups
2019-03-20 16:44:04 +00:00
Matthew Carrigan
1798e98e5a
Added final TODOs
2019-03-20 16:42:37 +00:00
Matthew Carrigan
c64c2fc4c2
Fixed embarrassing indentation problem
2019-03-20 15:42:57 +00:00
Matthew Carrigan
0540d360f2
Fixed logging
2019-03-20 15:36:51 +00:00
Matthew Carrigan
976554a472
First commit of the new LM finetuning
2019-03-20 14:23:51 +00:00
