Storage
We use rocksdb to save the data. You should create the storage object fisrt, then you can create embedding objects.
When creating storage object, you should input data-dir and ttl. Data-dir is the path to save data. Ttl is time to live, which is supported by rocksdb.
condition
conditions are wrap in parameter object.
expire_days
If the last update time of the key is less then expire_days ago, this key is ignored.
configure parameters:
expire_days: int type
group
If the group is setted (0 <= group < 256), we will only dump keys have the same group.
configure parameters:
group: int type
min_count
If the update number of key is less then min_count, this key is ignored.
configure parameters:
min_count: int type
dump
When model training finishes, you may dump the weights of keys with some conditions.
file format
First part stores the dim and count of each group, totally 256 groups. Second part stores all the key and weigh.
first part
| type | size | length | description |
|---|---|---|---|
| int32 | 4bit | 256 | dim of 256 group, default 0 |
| size_t | 8bit | 256 | key count of 256 group, default 0 |
second part
| type | size | length | description |
|---|---|---|---|
| int64_t | 8bit | 1 | key value |
| int32 | 4bit | 1 | group of the key |
| float | 4bit | dim of this group | weight of the key |
Example
import damo
# first param: data dir
# second param: ttl second
storage = damo.PyStorage("/tmp/data_dir", 86400*100)
cond = damo.Parameters()
cond.insert("expire_days", 100)
cond.insert("min_count", 3)
cond.insert("group", 0)
storage.dump("/tmp/weight.dat", cond)