jcrousse

import cupy as cp
import time
import asyncio


async def predict(N, power):
    compute_stream = cp.cuda.stream.Stream(non_blocking=True)
    compute_stream.use()
    d_mat = cp.random.randn(N * N, dtype=cp.float64).reshape(N, N)
    d_ret = d_mat

jcrousse

from prefixspan import PrefixSpan

from data_sources.data_generator import ExamplesGenerator, get_multiple_patterns

VOCAB_SIZE = 1000
SEQ_LEN = 250
multiple_patterns = get_multiple_patterns(10)

NUM_EXAMPLES = 200
MIN_FREQ = 25

jcrousse

def get_learned_scores(**kwargs):
    """
    scores each sentence, then multiply by score before next sequence layer.
    :Keyword Arguments:
        * sent_len (int) Sentence length
        * embedding_size (int) word embedding length
        * seq_len (int) length of overall sequence, equal to number of sentences x number of words per sentence
        * pre_embedded (bool) True if input is already vectors of word embeddings, false if tokens to be embedded
    :param :  (int)
    """

jcrousse

def get_learned_scores(**kwargs):
    """
    scores each sentence, then multiply by score before next sequence layer.
    :Keyword Arguments:
        * sent_len (int) Sentence length
        * embedding_size (int) word embedding length
        * seq_len (int) length of overall sequence, equal to number of sentences x number of words per sentence
        * pre_embedded (bool) True if input is already vectors of word embeddings, false if tokens to be embedded
        * concat_outputs (bool) True for a model with two similar outputs (2 level sequence model), False for 
        a single output attention model (weighted average of sentences) 

jcrousse

def write_to_html(sentences, highlight_vals, filename, low_val=(255, 255, 255), high_val=(77, 145, 255),
                  out_dir=OUT_FOLDER):
    scaled_hl = [e * (1 / max(highlight_vals)) for e in highlight_vals]
    with open(Path(out_dir) / filename, 'w') as f:
        for sent, score in zip(sentences, scaled_hl):
            color_vals = [int(low*(1-score) + high*score) for low, high in zip(low_val, high_val)]
            f.write(f"<span style=\"background-color: rgb({color_vals[0]},{color_vals[1]},{color_vals[2]})\">"
                    f"{sent}</span>\n")

jcrousse

def get_learned_scores(**kwargs):
    """
    scores each sentence, then multiply by score before next sequence layer
    """
    sent_len = kwargs.get('sent_len')
    embed_size = kwargs.get('embedding_size')
    seq_len = kwargs.get("seq_len")
    pre_embedded = kwargs.get("pre_embedded", False)
    assert seq_len % sent_len == 0, "sequence length must be a multiple of sentence length"
    sent_per_obs = seq_len // sent_len