IntGen's training and test

[wuxiaofei01]

Hello author, when I read your paper, I found that IntGen is divided into 4 columns. Do you train on 4 objects separately and then test on 4 objects? So is it divided into 4 independent experiments?

DATASET:
  TRAIN:
    TYPE: OIShape
    DATA_SPLIT: train
    DATA_ROOT: ./data
    DATA_MODE: intent
    OBJ_CATES: ["trigger_sprayer"]
    INTENT_MODE: ["use", "hold"]

  VAL:
    TYPE: OIShape
    DATA_SPLIT: val
    DATA_ROOT: ./data
    DATA_MODE: intent
    OBJ_CATES: ["trigger_sprayer"]
    INTENT_MODE: ["use", "hold"]

  TEST:
    TYPE: OIShape
    DATA_SPLIT: test
    DATA_ROOT: ./data
    DATA_MODE: intent
    OBJ_CATES: ["trigger_sprayer"]
    INTENT_MODE: "all"

[wuxiaofei01]

The metric only has physical constraints, but not whether to complete the "hold, use" posture.

[wuxiaofei01]

intgen，maybe have 4 categories？ but i just find 2 categories in files （“use” && “hold”） maybe i just train my mode in use and hold category？

[wuxiaofei01]

intgen，maybe have 4 categories？ but i just find 2 categories in files （“use” && “hold”） maybe i just train my mode in use and hold category？

in Oakink paper，there is a sentence "In the IntGen task, we select three intents: use, holdand hand-out, map the intents' word string to a real-valued word vector" what is the "hand-out"?

[lixiny]

1. Initially, we trained all category together and then fine-tuned on each category separately. The code provided here instead trained the model on different categories independently.
2. The code shared here demonstrates how we trained the initial intGen model. The original data loader differed significantly from the current data loading toolkit. I do not intend to replicate the model on the early, in-progression dataset. However, you can modify the config file to include a third intent, such as 'hand-out'. The network design and dataset readily support such modifications.