Gro2 env: support random goal from teacher task

[Jialn]

Add a random goal support, goal's name (string) was encoded by hash and remapped to range (-1, +1)

Some encoded codes for example: goalname:hammer, encode to: -0.2531430738150149 goalname:bookshelf, encode to: 0.43231639941069755 goalname:cube_20k, encode to: 0.7890891626763448 goalname:car_wheel, encode to: 0.25891776869694905 goalname:table, encode to: 0.33379767746182765 goalname:beer, encode to: -0.5148801851630069 goalname:plastic_cup, encode to: 0.1974983606824552

[emailweixu]

I think a more general approach would be better: the Task yield the utterance, then teacher change string of word sequence to integer sequence with fixed length, which can be used for models.

Encoding using a random number is not good for two reasons: 1. it cannot be extend to future scenario of sentences, 2. using a fixed random number to represent the goal may make the learning more difficult.

[Jialn]

I think a more general approach would be better: the Task yield the utterance, then teacher change string of word sequence to integer sequence with fixed length, which can be used for models.

Encoding using a random number is not good for two reasons: 1. it cannot be extend to future scenario of sentences, 2. using a fixed random number to represent the goal may make the learning more difficult.

Ok, I will close this PR and re-implement this with "with_sentence=True", as a minimal example of future scenario of full sentences.

[Jialn]

I think a more general approach would be better: the Task yield the utterance, then teacher change string of word sequence to integer sequence with fixed length, which can be used for models.

Encoding using a random number is not good for two reasons: 1. it cannot be extend to future scenario of sentences, 2. using a fixed random number to represent the goal may make the learning more difficult.

Which way is better to encode the word sequence to integer sequence? Shall we use one-hot encoding?

[emailweixu]

Just a sequence of integer is fine. The encoding is handled by the model.