BB3 3B memory module querying - object has no attribute

[krisstud]

According to the BB3 3B documentation, it should be possible to work with memories through the agent object here using .set_memory, .get_memory, get_available_memories() etc.

This probably some silly configuration error, but i can't find it. Using the following code i get AttributeError: 'ComboFidGoldDocumentAgent' object has no attribute 'get_available_memories'. May this be related to loading the seeker agent instead of r2c2_bb3 agent?

The code

from parlai.core.agents import create_agent

opt_dict = {'model':'projects.bb3.agents.r2c2_bb3_agent:BB3SubSearchAgent','interactive_mode': True, 'init_opt' : 'gen/r2c2_bb3.opt', 'model_file':'/path/ParlAI/data/models/bb3/bb3_3B/model'}
bb3_agent = create_agent(opt_dict) 
bb3_agent.get_available_memories()

I'm aware that i can pass personas using observation like so:

agent.observe({'text': 'your persona: You are a female\nyour persona: your name is jenny', 'episode_done': False})
agent.observe({'text': 'Whats your name?', 'episode_done': False})
response = agent.act()

Background information is that if the model is shown context longer than the truncation limit (even using several agent.observation that in themselves do not exceed it) it does not seem to store long-term memories. I would like to query the memory directly to confirm this. I saw a while back that another user had similar problems with long-term storage of memories in BB2 (can't find the link) and solved it using 'memory_key':'full_text', which isn't available in BB3 afaik.

[klshuster]

if you're initializing directly, you may need to manually specify all of the opt values in the gen/r2c2_bb3.opt. this includes using a different --model argument

[klshuster]

With BB3 that is the best way to do it

[krisstud]

I ended up creating a custom safe_interactive script that also takes the previous context, calling it using --safety none, creating a custom batch_act that takes also takes the label from the context as input and uses that instead of reply generation, a custom batch_act_decision that never searches, and finally looping over the context using observe, custom batch_act and self_observe. This "hack" works and generates relevant memory decisions and memories before entering parley, but..

i run out of cuda memory when the total memory count exceeds ~15-16. I see from nvidia-smi that each memory added can use up to 1-2GB of GPU memory. What to do? My memory generation related opts are as follows:

"memory_generator_control_token": "__generate-memory__",
  "mgm_beam_block_ngram": 3,
  "mgm_beam_min_length": 5,
  "mgm_beam_size": 3,
  "mgm_inference": "beam",
  "mgm_history_size": 1,
  "mgm_model": "projects.bb3.agents.r2c2_bb3_agent:BB3SubSearchAgent",

decision:

"mdm_beam_block_ngram": -1,
 "mdm_beam_min_length": 1,
 "mdm_beam_size": 1,
 "mdm_history_size": -1,
 "mdm_inference": "greedy",
 "mdm_model": "projects.bb3.agents.r2c2_bb3_agent:BB3SubSearchAgent",
 "memory_decision": "compute",
 "memory_decision_control_token": "__is-memory-required__",
 "memory_decision_do_access_reply": "__do-access-memory__",
 "memory_decision_dont_access_reply": "__do-not-access-memory__",
 "memory_decision_use_memories": true,

[klshuster]

You may be running out of memory due to the context length growing. If you type [DONE] after each interaction it will reset the conversation history

[krisstud]

You may be running out of memory due to the context length growing. If you type [DONE] after each interaction it will reset the conversation history

Appended that to every turn. Same result. Can i somehow set a limit (maximum) on the number of memories?

[klshuster]

Unfortunately not in the current code