lorenlugosch
commited on
Commit
•
a77c09c
1
Parent(s):
aca14ab
add checkpoints
Browse files- .gitattributes +6 -0
- CKPT.yaml +4 -0
- brain.ckpt +3 -0
- counter.ckpt +3 -0
- dataloader-TRAIN.ckpt +3 -0
- hyperparams.yaml +177 -0
- model.ckpt +3 -0
- optimizer.ckpt +3 -0
- scheduler.ckpt +3 -0
.gitattributes
CHANGED
@@ -14,3 +14,9 @@
|
|
14 |
*.pb filter=lfs diff=lfs merge=lfs -text
|
15 |
*.pt filter=lfs diff=lfs merge=lfs -text
|
16 |
*.pth filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
14 |
*.pb filter=lfs diff=lfs merge=lfs -text
|
15 |
*.pt filter=lfs diff=lfs merge=lfs -text
|
16 |
*.pth filter=lfs diff=lfs merge=lfs -text
|
17 |
+
brain.ckpt filter=lfs diff=lfs merge=lfs -text
|
18 |
+
counter.ckpt filter=lfs diff=lfs merge=lfs -text
|
19 |
+
dataloader-TRAIN.ckpt filter=lfs diff=lfs merge=lfs -text
|
20 |
+
model.ckpt filter=lfs diff=lfs merge=lfs -text
|
21 |
+
optimizer.ckpt filter=lfs diff=lfs merge=lfs -text
|
22 |
+
scheduler.ckpt filter=lfs diff=lfs merge=lfs -text
|
CKPT.yaml
ADDED
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# yamllint disable
|
2 |
+
SER: 10.33210332103321
|
3 |
+
end-of-epoch: true
|
4 |
+
unixtime: 1616380444.089852
|
brain.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:d9e24193f36931b7f57932532efbdcf64971f42732383ba6808825f77db258f6
|
3 |
+
size 28
|
counter.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:811786ad1ae74adfdd20dd0372abaaebc6246e343aebd01da0bfc4c02bf0106c
|
3 |
+
size 2
|
dataloader-TRAIN.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:454f63ac30c8322997ef025edff6abd23e0dbe7b8a3d5126a894e4a168c1b59b
|
3 |
+
size 3
|
hyperparams.yaml
ADDED
@@ -0,0 +1,177 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Generated 2021-03-21 from:
|
2 |
+
# /home/mila/l/lugoschl/code/fork/speechbrain/recipes/timers-and-such/direct/hparams/train.yaml
|
3 |
+
# yamllint disable
|
4 |
+
# ############################################################################
|
5 |
+
# Model: Direct SLU
|
6 |
+
# Encoder: Pre-trained ASR encoder -> LSTM
|
7 |
+
# Decoder: GRU + beamsearch
|
8 |
+
# Tokens: BPE with unigram
|
9 |
+
# losses: NLL
|
10 |
+
# Training: Timers and Such
|
11 |
+
# Authors: Loren Lugosch, Mirco Ravanelli 2020
|
12 |
+
# ############################################################################
|
13 |
+
|
14 |
+
# Seed needs to be set at top of yaml, before objects with parameters are made
|
15 |
+
seed: 4
|
16 |
+
__set_seed: !apply:torch.manual_seed [4]
|
17 |
+
experiment: train-real-only
|
18 |
+
output_folder: results/train-real-only/4
|
19 |
+
save_folder: results/train-real-only/4/save
|
20 |
+
train_log: results/train-real-only/4/train_log.txt
|
21 |
+
|
22 |
+
# Data files
|
23 |
+
data_folder: /localscratch/timers-and-such/
|
24 |
+
# e.g, /localscratch/timers-and-such
|
25 |
+
data_folder_rirs: /localscratch/timers-and-such/
|
26 |
+
train_splits: [train-real]
|
27 |
+
csv_train: results/train-real-only/4/train-type=direct.csv
|
28 |
+
csv_dev_real: results/train-real-only/4/dev-real-type=direct.csv
|
29 |
+
csv_dev_synth: results/train-real-only/4/dev-synth-type=direct.csv
|
30 |
+
csv_test_real: results/train-real-only/4/test-real-type=direct.csv
|
31 |
+
csv_test_synth: results/train-real-only/4/test-synth-type=direct.csv
|
32 |
+
csv_all_real: results/train-real-only/4/all-real-type=direct.csv
|
33 |
+
tokenizer_file: /home/mila/l/lugoschl/code/speechbrain/recipes/timers-and-such/Tokenizer/results/tokenizer_bpe51/51_unigram.model
|
34 |
+
skip_prep: false
|
35 |
+
ckpt_interval_minutes: 15 # save checkpoint every N min
|
36 |
+
test_on_all_real: false
|
37 |
+
|
38 |
+
# Training parameters
|
39 |
+
number_of_epochs: 50
|
40 |
+
batch_size: 16
|
41 |
+
lr: 0.0003
|
42 |
+
token_type: unigram # ["unigram", "bpe", "char"]
|
43 |
+
sorting: random
|
44 |
+
|
45 |
+
# Model parameters
|
46 |
+
sample_rate: 16000
|
47 |
+
emb_size: 128
|
48 |
+
dec_neurons: 512
|
49 |
+
output_neurons: 51 # index(eos/bos) = 0
|
50 |
+
ASR_encoder_dim: 512
|
51 |
+
encoder_dim: 256
|
52 |
+
|
53 |
+
# Decoding parameters
|
54 |
+
bos_index: 0
|
55 |
+
eos_index: 0
|
56 |
+
min_decode_ratio: 0.0
|
57 |
+
max_decode_ratio: 10.0
|
58 |
+
slu_beam_size: 80
|
59 |
+
eos_threshold: 1.5
|
60 |
+
temperature: 1.25
|
61 |
+
|
62 |
+
dataloader_opts:
|
63 |
+
batch_size: 16
|
64 |
+
shuffle: true
|
65 |
+
|
66 |
+
epoch_counter: &id009 !new:speechbrain.utils.epoch_loop.EpochCounter
|
67 |
+
|
68 |
+
limit: 50
|
69 |
+
|
70 |
+
# Models
|
71 |
+
asr_model: !apply:speechbrain.pretrained.EncoderDecoderASR.from_hparams
|
72 |
+
source: speechbrain/asr-crdnn-rnnlm-librispeech
|
73 |
+
run_opts: {device: cuda:0}
|
74 |
+
|
75 |
+
slu_enc: &id001 !new:speechbrain.nnet.containers.Sequential
|
76 |
+
input_shape: [null, null, 512]
|
77 |
+
lstm: !new:speechbrain.nnet.RNN.LSTM
|
78 |
+
input_size: 512
|
79 |
+
bidirectional: true
|
80 |
+
hidden_size: 256
|
81 |
+
num_layers: 2
|
82 |
+
linear: !new:speechbrain.nnet.linear.Linear
|
83 |
+
input_size: 512
|
84 |
+
n_neurons: 256
|
85 |
+
|
86 |
+
output_emb: &id002 !new:speechbrain.nnet.embedding.Embedding
|
87 |
+
num_embeddings: 51
|
88 |
+
embedding_dim: 128
|
89 |
+
|
90 |
+
dec: &id003 !new:speechbrain.nnet.RNN.AttentionalRNNDecoder
|
91 |
+
enc_dim: 256
|
92 |
+
input_size: 128
|
93 |
+
rnn_type: gru
|
94 |
+
attn_type: keyvalue
|
95 |
+
hidden_size: 512
|
96 |
+
attn_dim: 512
|
97 |
+
num_layers: 3
|
98 |
+
scaling: 1.0
|
99 |
+
dropout: 0.0
|
100 |
+
|
101 |
+
seq_lin: &id004 !new:speechbrain.nnet.linear.Linear
|
102 |
+
input_size: 512
|
103 |
+
n_neurons: 51
|
104 |
+
|
105 |
+
env_corrupt: &id005 !new:speechbrain.lobes.augment.EnvCorrupt
|
106 |
+
|
107 |
+
openrir_folder: /localscratch/timers-and-such/
|
108 |
+
babble_prob: 0.0
|
109 |
+
reverb_prob: 0.0
|
110 |
+
noise_prob: 1.0
|
111 |
+
noise_snr_low: 0
|
112 |
+
noise_snr_high: 15
|
113 |
+
|
114 |
+
modules:
|
115 |
+
slu_enc: *id001
|
116 |
+
output_emb: *id002
|
117 |
+
dec: *id003
|
118 |
+
seq_lin: *id004
|
119 |
+
env_corrupt: *id005
|
120 |
+
model: &id007 !new:torch.nn.ModuleList
|
121 |
+
- [*id001, *id002, *id003, *id004]
|
122 |
+
tokenizer: &id006 !new:sentencepiece.SentencePieceProcessor
|
123 |
+
|
124 |
+
pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
|
125 |
+
collect_in: results/train-real-only/4/save/TAS_tokenizer
|
126 |
+
loadables:
|
127 |
+
tokenizer: *id006
|
128 |
+
paths:
|
129 |
+
tokenizer: /home/mila/l/lugoschl/code/speechbrain/recipes/timers-and-such/Tokenizer/results/tokenizer_bpe51/51_unigram.model
|
130 |
+
|
131 |
+
beam_searcher: !new:speechbrain.decoders.S2SRNNBeamSearcher
|
132 |
+
embedding: *id002
|
133 |
+
decoder: *id003
|
134 |
+
linear: *id004
|
135 |
+
bos_index: 0
|
136 |
+
eos_index: 0
|
137 |
+
min_decode_ratio: 0.0
|
138 |
+
max_decode_ratio: 10.0
|
139 |
+
beam_size: 80
|
140 |
+
eos_threshold: 1.5
|
141 |
+
temperature: 1.25
|
142 |
+
using_max_attn_shift: false
|
143 |
+
max_attn_shift: 30
|
144 |
+
coverage_penalty: 0.
|
145 |
+
|
146 |
+
opt_class: !name:torch.optim.Adam
|
147 |
+
lr: 0.0003
|
148 |
+
|
149 |
+
lr_annealing: &id008 !new:speechbrain.nnet.schedulers.NewBobScheduler
|
150 |
+
initial_value: 0.0003
|
151 |
+
improvement_threshold: 0.0025
|
152 |
+
annealing_factor: 0.8
|
153 |
+
patient: 0
|
154 |
+
|
155 |
+
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
|
156 |
+
checkpoints_dir: results/train-real-only/4/save
|
157 |
+
recoverables:
|
158 |
+
model: *id007
|
159 |
+
scheduler: *id008
|
160 |
+
counter: *id009
|
161 |
+
augmentation: !new:speechbrain.lobes.augment.TimeDomainSpecAugment
|
162 |
+
sample_rate: 16000
|
163 |
+
speeds: [95, 100, 105]
|
164 |
+
|
165 |
+
log_softmax: !new:speechbrain.nnet.activations.Softmax
|
166 |
+
apply_log: true
|
167 |
+
|
168 |
+
seq_cost: !name:speechbrain.nnet.losses.nll_loss
|
169 |
+
label_smoothing: 0.1
|
170 |
+
|
171 |
+
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
|
172 |
+
save_file: results/train-real-only/4/train_log.txt
|
173 |
+
|
174 |
+
error_rate_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
|
175 |
+
|
176 |
+
cer_computer: !name:speechbrain.utils.metric_stats.ErrorRateStats
|
177 |
+
split_tokens: true
|
model.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:dc92f122a12758d82a4f74a86b1532de433796e04680a2769142586dc3ea0a8e
|
3 |
+
size 37183449
|
optimizer.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:cd0b106b455e3b89c605817e8ae7078c52c398110d2c7dc16cbb42fea6a2bbda
|
3 |
+
size 74369617
|
scheduler.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:9160465bf22dad633122f6a3254d8097be25e76ef754b2723c42ecec891a7329
|
3 |
+
size 879
|