Upload 2 files
Browse files- config.yaml +88 -0
- train_vdr.log +0 -0
config.yaml
ADDED
@@ -0,0 +1,88 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
biencoder:
|
2 |
+
alias: vdr
|
3 |
+
sequence_length: 512
|
4 |
+
dropout: 0.1
|
5 |
+
shared_encoder: false
|
6 |
+
semiparametric: true
|
7 |
+
device: null
|
8 |
+
encoder_q:
|
9 |
+
model_id: bert-base-uncased
|
10 |
+
max_seq_len: 256
|
11 |
+
pretrained: true
|
12 |
+
norm: false
|
13 |
+
shift_vocab_num: 1000
|
14 |
+
prefix: encoder_q.
|
15 |
+
encoder_p:
|
16 |
+
model_id: bert-base-uncased
|
17 |
+
max_seq_len: 256
|
18 |
+
pretrained: true
|
19 |
+
norm: false
|
20 |
+
shift_vocab_num: 1000
|
21 |
+
prefix: encoder_p.
|
22 |
+
datasets:
|
23 |
+
nq_train:
|
24 |
+
_target_: src.utils.data.BiEncoderDataset
|
25 |
+
file: ${HOME}/data/train/nq-train.jsonl
|
26 |
+
trivia_train:
|
27 |
+
_target_: src.utils.data.BiEncoderDataset
|
28 |
+
file: ${HOME}/data/train/trivia-train.jsonl
|
29 |
+
webq_train:
|
30 |
+
_target_: src.utils.data.BiEncoderDataset
|
31 |
+
file: ${HOME}/data/train/webq-train.jsonl
|
32 |
+
dl:
|
33 |
+
_target_: src.utils.data.BiEncoderDataset
|
34 |
+
file: ${HOME}/data/train/dl_10m.jsonl
|
35 |
+
cm:
|
36 |
+
_target_: src.utils.data.BiEncoderDataset
|
37 |
+
file: ${HOME}/data/train/cm_10m.jsonl
|
38 |
+
marco_dev:
|
39 |
+
_target_: src.utils.data.BiEncoderDataset
|
40 |
+
file: ${HOME}/data/train/msmarco-dev.jsonl
|
41 |
+
marco_dev_30neg:
|
42 |
+
_target_: src.utils.data.BiEncoderDataset
|
43 |
+
file: ${HOME}/data/train/msmarco-dev-30neg.jsonl
|
44 |
+
marco_train_bm25:
|
45 |
+
_target_: src.utils.data.BiEncoderDataset
|
46 |
+
file: ${HOME}/data/train/marco_bm25_20230119.jsonl
|
47 |
+
marco_train_1000:
|
48 |
+
_target_: src.utils.data.BiEncoderDataset
|
49 |
+
file: ${HOME}/data/train/marco_train_1000.jsonl
|
50 |
+
HOME: /export/data/jzhoubu/workspace/VDR-dense
|
51 |
+
train_datasets:
|
52 |
+
- marco_train_bm25
|
53 |
+
dev_datasets:
|
54 |
+
- marco_dev_30neg
|
55 |
+
datastore: null
|
56 |
+
output_dir: null
|
57 |
+
train_sampling_rates: null
|
58 |
+
batch_size: 64
|
59 |
+
num_train_epochs: 20
|
60 |
+
num_warmup_epochs: 1
|
61 |
+
num_save_epochs: 1
|
62 |
+
num_eval_epochs: 1
|
63 |
+
hard_negatives: 1
|
64 |
+
other_negatives: 0
|
65 |
+
ret_negatives: 0
|
66 |
+
train_insert_title: false
|
67 |
+
valid_insert_title: false
|
68 |
+
adam_eps: 1.0e-08
|
69 |
+
adam_betas: (0.9, 0.999)
|
70 |
+
learning_rate: 2.0e-05
|
71 |
+
max_grad_norm: 2.0
|
72 |
+
log_batch_step: 100
|
73 |
+
train_rolling_loss_step: 100
|
74 |
+
weight_decay: 0.0
|
75 |
+
sym_loss: true
|
76 |
+
do_lower_case: true
|
77 |
+
seed: 12345
|
78 |
+
checkpoint_file_name: vdr
|
79 |
+
save_every_epoch: true
|
80 |
+
model_file: null
|
81 |
+
local_rank: -1
|
82 |
+
local-rank: 2
|
83 |
+
device: null
|
84 |
+
distributed_world_size: null
|
85 |
+
distributed_port: null
|
86 |
+
no_cuda: false
|
87 |
+
n_gpu: null
|
88 |
+
fp16: true
|
train_vdr.log
ADDED
The diff for this file is too large to render.
See raw diff
|
|