File size: 1,640 Bytes
d28af7f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 |
dataset:
video_processor: ShardedVideoRetriVideoProcessor
bert_name: bert-base-uncased
meta_processor: ShardedHow2VideoRetriMetaProcessor
train_path: data/how2/how2_s3d_train.lst
val_path: data/how2/how2_s3d_val.lst
vfeat_dir: data/feat/feat_how2_s3d_shard_small
text_processor: ShardedVideoRetriTextProcessor
tfeat_dir: data/feat/feat_how2_s3d_shard_small/raw_caption_dedup.bert-base-uncased.
aligner: VideoRetriOverlappedAligner
subsampling: 1
sampled_min_len: 8
sampled_max_len: 64
max_video_len: 32
max_len: 96
lazy_vfeat_mask: true
mfm_probability: 0.15
mlm_probability: 0.15
mm_prob: 0.5
sampled_video_min_len: 3
sampled_video_max_len: 32
num_video_per_batch: 32
clip_per_video: 16
fairseq:
common:
tensorboard_logdir: run
log_interval: 1000
fp16: true
dataset:
num_workers: 4
batch_size: 1
optimization:
lr:
- 5.0e-05
clip_norm: 2.0
optimizer: adam
adam_betas: (0.9, 0.98)
lr_scheduler: polynomial_decay
total_num_update: 1000000
warmup_updates: 1000
weight_decay: 0.0
ddp_backend: no_c10d
max_epoch: 25
checkpoint:
save_dir: runs/retri/videoclip
save_interval_updates: 1024
keep_interval_updates: 2
keep_last_epochs: 30
task_type: sweep_big
slurm_config: big
eval:
save_path: runs/retri/videoclip
model:
model_cls: MMFusionSeparate
mm_encoder_cls: null
video_encoder_cls: MMBertForEncoder
text_encoder_cls: BertModel
num_hidden_video_layers: 6
loss:
loss_cls: MMContraLoss
task: VideoRetriTask
retri_epoch: 1
vectorpool_cls: VideoVectorPool
retriever_cls: VectorRetriever
num_cands: 64
|