def analyze_semantics(args,
model,
vocabulary,
seq_to_mutate,
escape_seqs,
min_pos=None,
max_pos=None,
prob_cutoff=0.,
beta=1.,
comb_batch=None,
plot_acquisition=True,
plot_namespace=None,
verbose=True):
if plot_acquisition:
dirname = ('target/{}/semantics/cache'.format(args.namespace))
mkdir_p(dirname)
if plot_namespace is None:
plot_namespace = args.namespace
y_pred = predict_sequence_prob(args,
seq_to_mutate,
vocabulary,
model,
verbose=verbose)
if min_pos is None:
min_pos = 0
if max_pos is None:
max_pos = len(seq_to_mutate) - 1
word_pos_prob = {}
for i in range(min_pos, max_pos + 1):
for word in vocabulary:
if seq_to_mutate[i] == word:
continue
word_idx = vocabulary[word]
prob = y_pred[i + 1, word_idx]
word_pos_prob[(word, i)] = prob
prob_seqs = {seq_to_mutate: [{'word': None, 'pos': None}]}
seq_prob = {}
for (word, pos), prob in word_pos_prob.items():
mutable = seq_to_mutate[:pos] + word + seq_to_mutate[pos + 1:]
seq_prob[mutable] = prob
if prob >= prob_cutoff:
prob_seqs[mutable] = [{'word': word, 'pos': pos}]
seqs = np.array([str(seq) for seq in sorted(seq_prob.keys())])
if plot_acquisition:
ofname = dirname + '/{}_mutations.txt'.format(args.namespace)
with open(ofname, 'w') as of:
of.write('orig\tmutant\n')
for seq in seqs:
try:
didx = [c1 != c2
for c1, c2 in zip(seq_to_mutate, seq)].index(True)
of.write('{}\t{}\t{}\n'.format(didx, seq_to_mutate[didx],
seq[didx]))
except ValueError:
of.write('NA\n')
base_embedding = embed_seqs(args,
model, {seq_to_mutate: [{}]},
vocabulary,
use_cache=False,
verbose=False)[seq_to_mutate][0]['embedding']
if comb_batch is None:
comb_batch = len(seqs)
n_batches = math.ceil(float(len(seqs)) / comb_batch)
seq_change = {}
for batchi in range(n_batches):
start = batchi * comb_batch
end = (batchi + 1) * comb_batch
prob_seqs_batch = {
seq: prob_seqs[seq]
for seq in seqs[start:end] if seq != seq_to_mutate
}
prob_seqs_batch = embed_seqs(args,
model,
prob_seqs_batch,
vocabulary,
use_cache=False,
verbose=False)
for mut_seq in prob_seqs_batch:
meta = prob_seqs_batch[mut_seq][0]
sem_change = abs(base_embedding - meta['embedding']).sum()
seq_change[mut_seq] = sem_change
cache_fname = dirname + ('/analyze_semantics_{}_{}_{}.txt'.format(
plot_namespace, args.model_name, args.dim))
probs, changes = [], []
with open(cache_fname, 'w') as of:
fields = [
'pos', 'wt', 'mut', 'prob', 'change', 'is_viable', 'is_escape'
]
of.write('\t'.join(fields) + '\n')
for seq in seqs:
prob = seq_prob[seq]
change = seq_change[seq]
mut = prob_seqs[seq][0]['word']
pos = prob_seqs[seq][0]['pos']
orig = seq_to_mutate[pos]
is_viable = seq in escape_seqs
is_escape = ((seq in escape_seqs)
and (sum([m['significant']
for m in escape_seqs[seq]]) > 0))
fields = [pos, orig, mut, prob, change, is_viable, is_escape]
of.write('\t'.join([str(field) for field in fields]) + '\n')
probs.append(prob)
changes.append(change)
if plot_acquisition:
from cached_semantics import cached_escape
cached_escape(cache_fname,
beta,
plot=plot_acquisition,
namespace=plot_namespace)
return seqs, np.array(probs), np.array(changes)
def fb_semantics(model,
repr_layers,
alphabet,
seq_to_mutate,
escape_seqs,
min_pos=None,
max_pos=None,
prob_cutoff=0.,
beta=1.,
comb_batch=None,
plot_acquisition=True,
namespace='fb',
plot_namespace=None,
verbose=True):
if plot_acquisition:
dirname = ('target/{}/semantics/cache'.format(namespace))
mkdir_p(dirname)
if plot_namespace is None:
plot_namespace = namespace
y_pred = predict_sequence_prob_fb(seq_to_mutate,
alphabet,
model,
repr_layers,
verbose=verbose)
if min_pos is None:
min_pos = 0
if max_pos is None:
max_pos = len(seq_to_mutate) - 1
word2idx = {
word: alphabet.all_toks.index(word)
for word in alphabet.all_toks
}
word_pos_prob = {}
for i in range(min_pos, max_pos + 1):
for word in alphabet.all_toks:
if '<' in word:
continue
if seq_to_mutate[i] == word:
continue
prob = y_pred[i + 1, word2idx[word]]
word_pos_prob[(word, i)] = 10**prob
prob_seqs = {seq_to_mutate: [{'word': None, 'pos': None}]}
seq_prob = {}
for (word, pos), prob in word_pos_prob.items():
mutable = seq_to_mutate[:pos] + word + seq_to_mutate[pos + 1:]
seq_prob[mutable] = prob
if prob >= prob_cutoff:
prob_seqs[mutable] = [{'word': word, 'pos': pos}]
seqs = np.array([str(seq) for seq in sorted(seq_prob.keys())])
base_embedding = embed_seqs_fb(
model, [seq_to_mutate],
repr_layers,
alphabet,
use_cache=False,
verbose=False)[seq_to_mutate][0]['embedding']
if comb_batch is None:
comb_batch = len(seqs)
n_batches = math.ceil(float(len(seqs)) / comb_batch)
seq_change = {}
for batchi in range(n_batches):
start = batchi * comb_batch
end = (batchi + 1) * comb_batch
tprint('Analyzing sequences {} to {}...'.format(start, end))
prob_seqs_batch = [
seq for seq in seqs[start:end] if seq != seq_to_mutate
]
prob_seqs_batch = embed_seqs_fb(model,
prob_seqs_batch,
repr_layers,
alphabet,
use_cache=False,
verbose=False)
for mut_seq in prob_seqs_batch:
meta = prob_seqs_batch[mut_seq][0]
sem_change = abs(base_embedding - meta['embedding']).sum()
seq_change[mut_seq] = sem_change
cache_fname = dirname + ('/analyze_semantics_{}_{}.txt'.format(
plot_namespace, model.model_version))
probs, changes = [], []
with open(cache_fname, 'w') as of:
fields = [
'pos', 'wt', 'mut', 'prob', 'change', 'is_viable', 'is_escape'
]
of.write('\t'.join(fields) + '\n')
for seq in seqs:
prob = seq_prob[seq]
change = seq_change[seq]
mut = prob_seqs[seq][0]['word']
pos = prob_seqs[seq][0]['pos']
orig = seq_to_mutate[pos]
is_viable = seq in escape_seqs
is_escape = ((seq in escape_seqs)
and (sum([m['significant']
for m in escape_seqs[seq]]) > 0))
fields = [pos, orig, mut, prob, change, is_viable, is_escape]
of.write('\t'.join([str(field) for field in fields]) + '\n')
probs.append(prob)
changes.append(change)
if plot_acquisition:
from cached_semantics import cached_escape
cached_escape(cache_fname,
beta,
plot=plot_acquisition,
namespace=plot_namespace)
return seqs, np.array(probs), np.array(changes)