hans

def viz_example(ex):
    toks = [v[idx] for idx in ex.features.feature["input_ids"].int64_list.value if idx > 0]
    masked_words = [v[idx] for idx in ex.features.feature["masked_lm_ids"].int64_list.value if idx > 0]

    masked_dict = dict(zip(ex.features.feature["masked_lm_positions"].int64_list.value, masked_words))
    toks_out = ["%10s" % tok for tok in toks]
    mask_out = ["%10s" % masked_dict[i] if i in masked_dict else " " * 10 for i in range(len(toks))]

    for block_start in range(0, len(toks), 10):
        print(" ".join(toks_out[block_start:block_start+10]))

hans

hans

hans

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "import matplotlib.pyplot as plt\n",

hans

// WebPPL code for a proof-of-concept model of
// question-guided visual search.

var images = [0, 1]
// deterministic p(description | image)
var descriptions = [
  {blocked_intent: true, missing_object: false},
  {blocked_intent: false, missing_object: true}
]

hans

from argparse import ArgumentParser
from collections import namedtuple
from pathlib import Path
import re

import pandas as pd


UTT_RE = re.compile(r"\*([A-Z]+):\s*(.+)\s*\x15(\d+)_(\d+)\x15$")
TAG_RE = re.compile(r"([a-z:]+)\|(\w+)")

hans

"""
Calculate statistics on verb-argument pairings given a parsed corpus
of CoNLL-U-formatted files. Part-of-speech tags and dependency heads+labels
are required.
"""

from collections import Counter
from pathlib import Path
import re
import sys

hans

\begin{tikzpicture}
\node (example) [draw,rectangle,align=center] at (0, 0) {Learning instance\\$w_1, \dots, w_N \to A$};
\node (dec-learn) [draw,rectangle,align=center] at (8, -3) {\textbf{Learn:}\\Infer meanings for $w_1, \dots, w_N$};
\node (lexicon) [draw,rectangle,align=center] at (0, -3) {Lexicon $\Lambda$\\$\text{raise} \to \texttt{S/N/PP} \lambda c\,x.\, c \land dir(e) = up \land move(e,x)$\\$\text{drop} \to \texttt{S/N/PP} \lambda c\,x.\, c \land dir(e) = down \land move(e,x)$\\$\text{onto} \to \texttt{PP} \lambda x.contact(x,e.patient)$\\$\ldots$};
\node (dec-compress) [draw,rectangle,align=center] at (0, -6) {\textbf{Compress:}\\Abstract away regularities in the lexicon};
\node (dec-bridge) [draw,rectangle,align=center] at (-8, -3) {\textbf{Bridge:}\\Derive new syntactic sub-categories};

\begin{scope}[shift={(-6,-7.2)}]
\Tree [.$\lambda$ [.$c$ ] [.$x$ ] [.$\land$ [.$c$ ] [.$=$ [.$dir(e)$ ] [.$up$ ] ] [.$move$ [.$e$ ] [.$x$ ] ] ] ]
\begin{scope}[xshift=130pt]

hans

var ps = Categorical({vs: [0.2, 0.4, 0.6, 0.8]})
var ps_skew_left = Categorical({vs: [0.1, 0.2, 0.3, 0.8]})

var vs = Categorical({vs: [2, 4, 6, 8]})
var vs_skew_left = Categorical({vs: [1, 2, 3, 10]})

var Model = function(name, f) {
    Object.defineProperty(f, 'name', {value: name})
    return f;
};

hans

First stab at decoding P01 data onto InferSent embeddings (dim 4096).
This is promising performance, since I haven't touched either the imaging data or the encodings since they were produced. (The regression is ~20,000 dim -> 4096 dim – definitely can improve with some preprocessing.)
Compare with original results from the problem set later on in this gist.

INFO:__main__:Loaded encodings of size (384, 4096).
INFO:__main__:Loaded subject data/P01 data.
INFO:__main__:Trained classifier for subject data/P01.
Fold  0: min  4.0       mean  96.5      med  41.0       max  350.0
Fold  1: min  0.0       mean  131.9     med  68.0       max  323.0
Fold  2: min  2.0       mean  144.3     med  129.0      max  366.0