SocialAISchool / param_tree_demo.py
grg's picture
Cleaned old git history
be5548b
import streamlit as st
import copy
import streamlit.components.v1 as components
import streamlit.caching as caching
import time
import argparse
import numpy as np
import gym
import gym_minigrid
from gym_minigrid.wrappers import *
from gym_minigrid.window import Window
import matplotlib.pyplot as plt
from gym_minigrid.social_ai_envs.socialaigrammar import SocialAIGrammar, SocialAIActions, SocialAIActionSpace
default_params = {
"Pointing": 0,
"Emulation": 1,
"Language_grounding": 2,
"Pragmatic_frame_complexity": 1,
}
class InteractiveACL:
def choose(self, node, chosen_parameters):
options = [n.label for n in node.children]
box_name = f'{node.label} ({node.id})'
ret = st.sidebar.selectbox(
box_name,
options,
index=default_params.get(node.label, 0)
)
for ind, (c, c_lab) in enumerate(zip(node.children, options)):
if c_lab == ret:
return c
def get_info(self):
return {}
@st.cache(allow_output_mutation=True, suppress_st_warning=True)
def load_env():
env = gym.make("SocialAI-SocialAIParamEnv-v1")
env.curriculum=InteractiveACL()
return env
st.title("SocialAI interactive demo")
env = load_env()
st.subheader("Primitive actions")
# moving buttons
columns = st.columns([1]*(len(SocialAIActions)+1))
action_names = [a.name for a in list(SocialAIActions)] + ["no_op"]
# keys = ["Left arrow", "Right arrow", "Up arrow", "t", "q", "Shift"]
keys = ["a", "d", "w", "t", "q", "Shift"]
# actions = [st.button(a.name) for a in list(SocialAIActions)] + [st.button("none")]
actions = []
for a_name, col, key in zip(action_names, columns, keys):
with col:
actions.append(st.button(a_name+f" ({key})", help=f"Shortcut: {key}"))
st.subheader("Speaking actions")
# talking buttons
t, w, b = st.columns([1, 1, 1])
changes = [False, False]
with t:
templ = st.selectbox("Template", options=SocialAIGrammar.templates, index=1)
with w:
word = st.selectbox("Word", options=SocialAIGrammar.things, index=0)
speak = st.button("Speak (s)", help="Shortcut s")
# utterance change detection
utt_changed = False
if "template" in st.session_state:
utt_changed = st.session_state.template != templ
if "word" in st.session_state:
utt_changed = utt_changed or st.session_state.word != word
st.session_state["template"] = templ
st.session_state["word"] = word
st.sidebar.subheader("Select the parameters:")
play = st.button("Play (Enter)", help="Generate the env. Shortcut: Enter")
components.html(
"""
<script>
const doc = window.parent.document;
buttons = Array.from(doc.querySelectorAll('button[kind=primary]'));
const left_button = buttons.find(el => el.innerText === 'left (a)');
const right_button = buttons.find(el => el.innerText === 'right (d)');
const forward_button = buttons.find(el => el.innerText === 'forward (w)');
const toggle_button = buttons.find(el => el.innerText === 'toggle (t)');
const none_button = buttons.find(el => el.innerText === 'no_op (Shift)');
const done_button = buttons.find(el => el.innerText === 'done (q)');
const play_button = buttons.find(el => el.innerText === 'Play (Enter)');
const speak_button = buttons.find(el => el.innerText === 'Speak (s)');
doc.addEventListener('keydown', function(e) {
switch (e.keyCode) {
case 65: // (65 = a )
left_button.click();
break;
case 68: // (68 = d )
right_button.click();
break;
case 87: // (87 = w )
forward_button.click();
break;
case 84: // (84 = t)
toggle_button.click();
break;
case 16: // (16 = shift)
none_button.click();
break;
case 81: // (81 = q)
done_button.click();
break;
case 13: // (13 = enter)
play_button.click();
break;
case 83: // (83 = s)
speak_button.click();
break;
}
});
</script>
""",
height=0,
width=0,
)
# no action
done_ind = len(actions) - 2
actions[done_ind] = False
# was agent controlled
no_action = not any(actions) and not speak
done = False
info = None
if not no_action or play or utt_changed:
# agent is controlled
if any(actions):
p_act = np.argmax(actions)
if p_act == len(actions) - 1:
p_act = np.nan
action = [p_act, np.nan, np.nan]
elif speak:
templ_ind = SocialAIGrammar.templates.index(templ)
word_ind = SocialAIGrammar.things.index(word)
action = [np.nan, templ_ind, word_ind]
else:
action = None
if action:
obs, reward, done, info = env.step(action)
env.render(mode='human')
st.pyplot(env.window.fig)
# if done or no_action:
if done or (no_action and not play and not utt_changed):
env.reset()
else:
env.parameter_tree.sample_env_params(ACL=env.curriculum)
with st.expander("Parametric tree", True):
# draw tree
current_param_labels = env.current_env.parameters if env.current_env.parameters else {}
folded_nodes = [
"Information_seeking",
"Collaboration",
"OthersPerceptionInference"
]
# print(current_param_labels["Env_type"])
folded_nodes.remove(current_param_labels["Env_type"])
env.parameter_tree.draw_tree(
filename="viz/streamlit_temp_tree",
ignore_labels=["Num_of_colors"],
selected_parameters=current_param_labels,
folded_nodes=folded_nodes,
# save=False
)
# st.graphviz_chart(env.parameter_tree.tree)
st.image("viz/streamlit_temp_tree.png")
# if not no_action or play or utt_changed:
# # agent is controlled
# if any(actions):
# p_act = np.argmax(actions)
# if p_act == len(actions) - 1:
# p_act = np.nan
#
# action = [p_act, np.nan, np.nan]
#
# elif speak:
# templ_ind = SocialAIGrammar.templates.index(templ)
# word_ind = SocialAIGrammar.things.index(word)
# action = [np.nan, templ_ind, word_ind]
#
# else:
# action = None
#
# if action:
# obs, reward, done, info = env.step(action)
#
# env.render(mode='human')
# st.pyplot(env.window.fig)