def sample(self, batch, mode, word_based=True):
"""used for actor in ELF and visually evaulating model
return
inst: [batch, max_sentence_len], even inst is one-hot
inst_len: [batch]
"""
output = self.rl_forward(batch, mode)
samples = self.sampler.sample(output["cont_pi"], output["inst_pi"],
batch["prev_inst_idx"])
log_prob_reply = {
"samples": samples,
"probs": {
self.sampler.cont_prob_key: output["cont_pi"],
self.sampler.prob_key: output["inst_pi"],
},
"value": output["v"],
}
reply = {
"cont": samples["cont"].unsqueeze(1),
"cont_pi": output["cont_pi"],
"inst": samples["inst"].unsqueeze(1),
"inst_pi": output["inst_pi"],
}
# convert format needed by executor
samples = []
lengths = []
raws = []
for idx in reply["inst"]:
inst = self.inst_dict.get_inst(int(idx.item()))
tokens, length = self.inst_dict.parse(inst, True)
samples.append(tokens)
lengths.append(length)
raw = convert_to_raw_instruction(inst, self.max_raw_chars)
raws.append(convert_to_raw_instruction(inst, self.max_raw_chars))
device = reply["cont"].device
if word_based:
# for word based
inst = torch.LongTensor(samples).to(device)
else:
inst = reply["inst"]
inst_len = torch.LongTensor(lengths).to(device)
reply["raw_inst"] = torch.LongTensor(raws).to(device)
return inst, inst_len, reply["cont"], reply, log_prob_reply
def _get_human_instruction(self, batch):
assert_eq(batch['prev_inst'].size(0), 1)
device = batch['prev_inst'].device
inst = input('Please input your instruction\n')
# inst = 'build peasant'
import pdb
pdb.set_trace()
inst_idx = torch.zeros((1, )).long().to(device)
inst_idx[0] = self.executor.inst_dict.get_inst_idx(inst)
inst_cont = torch.zeros((1, )).long().to(device)
if len(inst) == 0:
# inst = batch['prev_inst']
inst = self.prev_inst
inst_cont[0] = 1
self.prev_inst = inst
raw_inst = convert_to_raw_instruction(inst, self.max_raw_chars)
inst, inst_len = self.executor.inst_dict.parse(inst, True)
inst = torch.LongTensor(inst).unsqueeze(0).to(device)
inst_len = torch.LongTensor([inst_len]).to(device)
raw_inst = torch.LongTensor([raw_inst]).to(device)
reply = {
'inst': inst_idx.unsqueeze(1),
'inst_pi':
torch.ones(1, self.num_insts).to(device) / self.num_insts,
'cont': inst_cont.unsqueeze(1),
'cont_pi': torch.ones(1, 2).to(device) / 2,
'raw_inst': raw_inst
}
return inst, inst_len, inst_cont, reply
def sample(self, batch, mode, word_based=True, agent_mask=None):
"""used for actor in ELF and visually evaulating model
return
inst: [batch, max_sentence_len], even inst is one-hot
inst_len: [batch]
"""
output = self.rl_forward(batch, mode, agent_mask=agent_mask)
samples = self.sampler.sample(output['cont_pi'], output['inst_pi'],
batch['prev_inst_idx'])
reply = {
'cont': samples['cont'].unsqueeze(1),
'cont_pi': output['cont_pi'],
'inst': samples['inst'].unsqueeze(1),
'inst_pi': output['inst_pi'],
}
# convert format needed by executor
samples = []
lengths = []
raws = []
for idx in reply['inst']:
inst = self.inst_dict.get_inst(int(idx.item()))
tokens, length = self.inst_dict.parse(inst, True)
samples.append(tokens)
lengths.append(length)
raw = convert_to_raw_instruction(inst, self.max_raw_chars)
raws.append(convert_to_raw_instruction(inst, self.max_raw_chars))
device = reply['cont'].device
if word_based:
# for word based
inst = torch.LongTensor(samples).to(device)
else:
inst = reply['inst']
inst_len = torch.LongTensor(lengths).to(device)
reply['raw_inst'] = torch.LongTensor(raws).to(device)
return inst, inst_len, reply['cont'], reply