def benchmark_correlations(best_layer=True):
data = []
# Pereira internal
Pereira_experiment2_scores, Pereira_experiment3_scores = collect_Pereira_experiment_scores(
best_layer=best_layer)
Pereira_experiment2_scores = Pereira_experiment2_scores['score'].values
Pereira_experiment3_scores = Pereira_experiment3_scores['score'].values
correlation_Pereira, p_Pereira = pearsonr(Pereira_experiment2_scores,
Pereira_experiment3_scores)
data.append(
dict(benchmark1='Pereira Exp. 2',
benchmark2='Pereira Exp. 3',
r=correlation_Pereira,
p=p_Pereira))
# cross-benchmark
benchmarks = ('Pereira2018-encoding', 'Blank2014fROI-encoding',
'Fedorenko2016-encoding')
for benchmark1, benchmark2 in itertools.combinations(benchmarks, 2):
benchmark1_scores = collect_scores(benchmark=benchmark1,
models=all_models)
benchmark2_scores = collect_scores(benchmark=benchmark2,
models=all_models)
benchmark1_scores = average_adjacent(benchmark1_scores).dropna()
benchmark2_scores = average_adjacent(benchmark2_scores).dropna()
if best_layer:
benchmark1_scores = choose_best_scores(benchmark1_scores)
benchmark2_scores = choose_best_scores(benchmark2_scores)
benchmark1_scores, benchmark2_scores = align_scores(
benchmark1_scores,
benchmark2_scores,
identifier_set=('model', ) if best_layer else ('model', 'layer'))
benchmark1_scores, benchmark2_scores = benchmark1_scores[
'score'].values, benchmark2_scores['score'].values
r, p = pearsonr(benchmark1_scores, benchmark2_scores)
data.append(
dict(benchmark1=benchmark1, benchmark2=benchmark2, r=r, p=p))
data = pd.DataFrame(data)
# plot
fig, ax = pyplot.subplots(figsize=(3, 4))
x = np.arange(len(data))
ax.bar(x, data['r'])
ax.set_xticks(x)
ax.set_xticklabels([
f"{benchmark1[:5]} / {benchmark2[:5]}" for benchmark1, benchmark2 in
zip(data['benchmark1'].values, data['benchmark2'].values)
],
rotation=90)
for _x, r, p in zip(x, data['r'].values, data['p'].values):
ax.text(_x,
r + .05,
significance_stars(p) if p < .05 else 'n.s.',
fontsize=12,
horizontalalignment='center',
verticalalignment='center')
savefig(
fig,
Path(__file__).parent / 'benchmark-correlations' +
('-best' if best_layer else '-layers'))
def shortcomings(model):
benchmarks = [
'Futrell2018-encoding', 'Futrell2018sentences-encoding',
'Futrell2018stories-encoding'
]
scores = [
collect_scores(models=[model], benchmark=benchmark)
for benchmark in benchmarks
]
scores = reduce(lambda left, right: pd.concat([left, right]), scores)
scores = average_adjacent(scores).dropna()
scores = choose_best_scores(scores)
fig, ax = pyplot.subplots(figsize=(3.5, 5))
x = np.arange(len(benchmarks))
ax.bar(x,
height=scores['score'],
yerr=scores['error'],
color=model_colors[model],
edgecolor='none',
ecolor='gray',
error_kw=dict(elinewidth=1, alpha=.5))
ax.set_xticks(x)
ax.set_xticklabels(['words', 'sentences', 'stories'], rotation=45)
ax.yaxis.set_major_locator(MultipleLocator(base=0.2))
ax.yaxis.set_major_formatter(score_formatter)
ax.set_ylim([0, 1.2])
ax.set_ylabel('Normalized Predictivity')
ax.set_title('Futrell2018 variations')
savefig(fig, Path(__file__).parent / "bars-generalization")
def random_embedding():
models = ['gpt2-xl', 'gpt2-xl-untrained', 'random-embedding']
benchmarks = [
'Pereira2018-encoding', 'Fedorenko2016v3-encoding',
'Blank2014fROI-encoding', 'Futrell2018-encoding'
]
scores = [
collect_scores(benchmark=benchmark, models=models)
for benchmark in benchmarks
]
scores = [
average_adjacent(benchmark_scores) for benchmark_scores in scores
]
scores = [
choose_best_scores(benchmark_scores).dropna()
for benchmark_scores in scores
]
scores = reduce(lambda left, right: pd.concat([left, right]), scores)
fig, ax = pyplot.subplots(figsize=(5, 4))
colors = {
'gpt2-xl': model_colors['gpt2-xl'],
'gpt2-xl-untrained': '#284343',
'random-embedding': '#C3CCCC'
}
offsets = {0: -.2, 1: 0, 2: +.2}
width = 0.5 / 3
text_kwargs = dict(fontdict=dict(fontsize=7), color='white')
base_x = np.arange(len(benchmarks))
for i, model in enumerate(models):
model_scores = scores[scores['model'] == model]
x = base_x + offsets[i]
ax.bar(x,
height=model_scores['score'],
yerr=model_scores['error'],
width=width,
align='center',
color=colors[model],
edgecolor='none',
ecolor='gray',
error_kw=dict(elinewidth=1, alpha=.5))
for xpos in x:
text = ax.text(x=xpos + .6 * width / 2,
y=.05,
s=model_label_replace[model],
rotation=90,
rotation_mode='anchor',
**text_kwargs)
text.set_path_effects(
[patheffects.withStroke(linewidth=0.75, foreground='black')])
ax.set_xticks(base_x)
ax.set_xticklabels(
[benchmark_label_replace[benchmark] for benchmark in benchmarks],
fontsize=9)
ax.yaxis.set_major_locator(MultipleLocator(base=0.2))
ax.yaxis.set_major_formatter(score_formatter)
ax.set_ylim([0, 1.2])
ax.set_ylabel('Normalized Predictivity')
savefig(fig, Path(__file__).parent / "bars-random_embedding")
def retrieve_scores(benchmark, models=all_models):
scores = collect_scores(benchmark, models)
scores = average_adjacent(
scores) # average each model+layer's score per experiment and atlas
scores = scores.fillna(0) # nan scores are 0
scores = choose_best_scores(scores)
nan = scores[scores.isna().any(1)]
if len(nan) > 0:
_logger.warning(f"Dropping nan rows: {nan}")
scores = scores.dropna()
return scores
def Fedorenko2016(best_layer=True):
scores_lang = collect_scores(benchmark='Fedorenko2016v3-encoding',
models=models,
normalize=True)
scores_nonlang = collect_scores(
benchmark='Fedorenko2016v3nonlang-encoding',
models=models,
normalize=True)
scores_lang, scores_nonlang = average_adjacent(
scores_lang).dropna(), average_adjacent(scores_nonlang).dropna()
if best_layer:
scores_lang, scores_nonlang = choose_best_scores(
scores_lang), choose_best_scores(scores_nonlang)
scores_lang, scores_nonlang = align_scores(
scores_lang,
scores_nonlang,
identifier_set=['model'] if best_layer else ['model', 'layer'])
diffs = scores_lang['score'] - scores_nonlang['score']
print(f"median drop {np.nanmedian(diffs)}+-{np.std(diffs)}")
mults = scores_lang['score'] / scores_nonlang['score']
print(f"median multiplicative drop {np.nanmedian(mults)}+-{np.std(mults)}")
def layer_preference_single(
model='gpt2-xl',
benchmarks=('Pereira2018-encoding', 'Fedorenko2016v3-encoding',
'Blank2014fROI-encoding'),
smoothing=False,
):
data = [
collect_scores(benchmark=benchmark, models=[model])
for benchmark in benchmarks
]
data = [average_adjacent(d) for d in data]
fig, axes = pyplot.subplots(figsize=(15, 6),
nrows=1,
ncols=len(benchmarks),
sharey=True)
for benchmark_iter, (ax, benchmark_name, benchmark_data) in enumerate(
zip(axes.flatten(), benchmarks, data)):
ax.set_title(benchmark_name)
num_layers = len(
benchmark_data['layer']) # assume layers are correctly ordered
relative_position = np.arange(num_layers) / (num_layers - 1)
y, error = benchmark_data['score'], benchmark_data['error']
if smoothing:
window_size = int(len(y) * 2 / 3)
if window_size % 2 == 0: # if even
window_size += 1 # make odd (required for filter)
y = savgol_filter(y, window_size, 3)
shaded_errorbar(x=relative_position,
y=y,
error=error,
label=model,
ax=ax,
alpha=0.4,
color=model_colors[model],
linewidth=7.0 if model == 'gpt2-xl' else 1.0,
shaded_kwargs=dict(alpha=0.2,
color=model_colors[model]))
if benchmark_iter > 0:
ax.set_yticklabels([])
else:
ax.set_ylabel('score')
ax.set_ylim([0, 1.2])
# xlabel
fig.text(0.5, 0.01, 'relative layer position', ha='center')
# save
fig.tight_layout()
savefig(fig, Path(__file__).parent / f'layer_ordering-{model}')
def layer_preference(benchmark='Pereira2018-encoding'):
models = [model for model in all_models if len(model_layers[model]) > 1
] # need at least 2 layers to plot
data = collect_scores(benchmark=benchmark, models=models)
data = average_adjacent(data)
groups = copy.deepcopy(model_groups)
groups['other'] = [
model for model in models
if not any(model in group for group in groups.values())
]
_logger.debug(f"Non-assigned models: {groups['other']}")
fig, axes = pyplot.subplots(figsize=(20, 6),
nrows=1,
ncols=len(groups),
sharey=True)
for model_group_iter, (ax, (group_name, models)) in enumerate(
zip(axes.flatten(), groups.items())):
ax.set_title(group_name)
for model in models:
group = data[data['model'] == model]
num_layers = len(
group['layer']) # assume layers are correctly ordered
relative_position = np.arange(num_layers) / (num_layers - 1)
shaded_errorbar(x=relative_position,
y=group['score'],
error=group['error'],
label=model,
ax=ax,
alpha=0.4,
color=model_colors[model],
shaded_kwargs=dict(alpha=0.2,
color=model_colors[model]))
if model_group_iter > 0:
ax.set_yticklabels([])
else:
ax.set_ylabel('score')
ax.set_ylim([0, 1.2])
# xlabel
fig.text(0.5, 0.01, 'relative layer position', ha='center')
# save
fig.tight_layout()
savefig(fig, Path(__file__).parent / f'layer_ordering-{benchmark}')
def first_last_layer_scores(benchmarks=('Pereira2018-encoding',
'Fedorenko2016v3-encoding',
'Blank2014fROI-encoding')):
models = all_models
data = [
collect_scores(benchmark=benchmark, models=models)
for benchmark in benchmarks
]
data = [average_adjacent(d) for d in data]
@matplotlib.ticker.FuncFormatter
def score_formatter(score, pos):
if score < 0 or score > 1:
return ""
return f"{score:.2f}"
fig, axes = pyplot.subplots(figsize=(15, 15),
nrows=len(benchmarks),
ncols=1,
sharey=False)
width = 0.5
for benchmark_iter, (benchmark,
benchmark_data) in enumerate(zip(benchmarks, data)):
ax = axes[benchmark_iter]
for model_iter, model in enumerate(models):
model_data = benchmark_data[benchmark_data['model'] == model]
best_score = model_data['score'].max()
best_score_error = model_data[model_data['score'] ==
best_score]['error']
ax.errorbar(x=model_iter,
y=best_score,
yerr=best_score_error,
marker='.',
color='black',
label='best layer' if model_iter == len(models) -
1 else None)
if len(model_data) > 1:
first_score, first_score_error = model_data['score'].values[
0], model_data['error'].values[0]
ax.errorbar(x=model_iter - 0.2 * width,
y=first_score,
yerr=first_score_error,
marker='.',
color='lightgray',
label='first layer' if model_iter == len(models) -
1 else None)
last_score, last_score_error = model_data['score'].values[
-1], model_data['error'].values[-1]
ax.errorbar(x=model_iter + 0.2 * width,
y=last_score,
yerr=last_score_error,
marker='.',
color='gray',
label='last layer' if model_iter == len(models) -
1 else None)
if benchmark_iter < len(benchmarks) - 1:
ax.set_xticks([])
else:
ax.set_xticks(np.arange(len(models)))
ax.set_xticklabels(
[model_label_replace[model] for model in models], rotation=90)
if benchmark_iter == 0:
ax.legend()
ax.set_ylabel('Normalized Predictivity')
ax.set_title(benchmark_label_replace[benchmark])
ax.yaxis.set_major_formatter(score_formatter)
savefig(fig, Path(__file__).parent / 'first_layer_layer_scores')
def _model_scores(benchmark, models):
scores = collect_scores(benchmark=benchmark, models=models, normalize=True)
scores = average_adjacent(scores).dropna()
scores = choose_best_scores(scores)
return scores