-
Notifications
You must be signed in to change notification settings - Fork 5
/
example.py
299 lines (244 loc) · 8.98 KB
/
example.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
# # Using smat for new models and datasets
#
import jax
import jax.numpy as jnp
from numpy import random
import flax
# ## Get dataset and define dataloader
#
# First we need to get the SST-2 data. We will use HF's dataset since it is already
# very compatible with the smat library.
# +
from datasets import load_dataset
def load_data(
split: str,
seed: int = 0,
):
hf_dataset = load_dataset("glue", "sst2")
data = hf_dataset["validation" if split == "valid" else split]
data = list(data)
rng = random.RandomState(seed)
rng.shuffle(data)
return data
train_data = load_data("train")
valid_data = load_data("valid")
num_classes = 2
# -
# We also need to define the dataloader, which is going to create batches that
# are going to be consumed by the model
# +
from typing import List, Dict
def dataloader(
dataset: List[Dict[str, int]],
tokenizer,
batch_size: int,
shuffle: bool = True,
max_len: int = 128,
):
idxs = list(range(len(dataset)))
if shuffle:
random.shuffle(idxs)
for i in range(0, len(idxs), batch_size):
batch_inputs, batch_outputs, batch_idxs = [], [], []
for j in range(batch_size):
if i + j >= len(idxs):
break
sample = dataset[idxs[i + j]]
batch_idxs.append(idxs[i + j])
batch_inputs.append(sample["sentence"])
batch_outputs.append(jnp.array(sample["label"]))
batch_inputs = dict(
tokenizer(
batch_inputs,
padding="max_length",
truncation=True,
return_tensors="jax",
max_length=max_len,
)
)
batch_outputs = jnp.stack(batch_outputs)
yield batch_inputs, batch_outputs
# -
# ## Define Wrapped Model
#
# Next we need to define a wrapper class for our BERT model.
# This is not strictly necessary, but it makes integration
# with the smat library much easier
from transformers import FlaxBertForSequenceClassification, BertConfig
from transformers.models.bert.modeling_flax_bert import FlaxBertForSequenceClassificationModule
from smat.models import WrappedModel, register_model
@register_model("bert")
class WrappedBert(WrappedModel):
num_labels: int
config: BertConfig
def setup(self):
self.bert_module = FlaxBertForSequenceClassificationModule(
config=self.config,
)
self.scalarmix = ScalarMix()
def __call__(
self,
input_ids,
attention_mask=None,
token_type_ids=None,
position_ids=None,
deterministic: bool = True,
):
if token_type_ids is None:
token_type_ids = jnp.ones_like(input_ids)
if position_ids is None:
position_ids = jnp.arange(jnp.atleast_2d(input_ids).shape[-1])
if attention_mask is None:
attention_mask = jnp.ones_like(input_ids)
outputs, hidden_states, attentions = self.bert_module(
input_ids,
attention_mask=attention_mask,
token_type_ids=token_type_ids,
position_ids=position_ids,
output_hidden_states=True,
output_attentions=True,
unnorm_attention=True,
deterministic=deterministic,
return_dict=False,
)
state = {"hidden_states": hidden_states, "attentions": attentions}
return outputs, state
@classmethod
def initialize_new_model(
cls,
key,
inputs,
num_classes,
identifier="bert-base-uncased",
**kwargs,
):
model = FlaxBertForSequenceClassification.from_pretrained(
identifier,
num_labels=num_classes,
)
classifier = cls(
num_labels=num_classes,
config=model.config,
)
params = classifier.init(key, **inputs)
# replace original parameters with pretrained parameters
params = params.unfreeze()
assert "bert_module" in params["params"]
params["params"]["bert_module"] = model.params
params = flax.core.freeze(params)
return classifier, params
# we also need to get the tokenizer for the BERT model
# +
from transformers import BertTokenizer
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
# -
# ## Getting the trained model (the teacher)
# To get the teacher model, we take a simple approach and use an existing model
# on the hugginface hub. We also also need an example input
# for jax's shape inference, for which we will just use the first batch in the dataset
# +
from smat.utils import PRNGSequence
keyseq = PRNGSequence(0)
example_batchx = next(dataloader(train_data, tokenizer, batch_size=1))[0]
model, params = WrappedBert.initialize_new_model(
next(keyseq),
example_batchx,
num_classes=num_classes,
identifier="textattack/bert-base-uncased-SST-2"
)
# -
# ## Training an explainer
#
# Finally, we are ready to train an explainer model. `smat.compact` provides
# a convenient wrapper that trains a student and an explainer using
# the SMAT framework. Some options to consider for this function are the number
# of samples used to train the student model. In this example we are going to use
# 10% of the training data for a single epoch, and will make convinient use
# of the default bert model defined by the `register_model` decorator. **NOTE**: the hyperparameters weren't tuned, so you might obtain better explainers by playing with those.
# +
from smat.compact import train_explainer
from functools import partial
explainer, expl_params = train_explainer(
task_type="classification",
teacher_model=model,
teacher_params=params,
dataloader=partial(dataloader, tokenizer=tokenizer),
train_dataset=train_data,
valid_dataset=valid_data,
num_examples=0.1,
student_model="bert",
max_epochs=5,
)
# -
# **WARNING**: smat currently requires that the student shares the same
# vocabulary as the teacher. Notice that a single dataloader with the same tokenizer
# is used for both the teacher and the student.
def predict_and_explain(sample):
y, state = model.apply(params, **sample)
explanation = explainer.apply(expl_params, sample, state)[0]
return y, explanation
text = "Though everything might be literate and smart , it never took off and always seemed static ."
batch = dict(tokenizer([text],return_tensors="jax"))
y_pred, expl = predict_and_explain(batch)
print(y_pred.argmax())
print(expl)
# ## Visualizing the head coefficients
#
# We will recover the learned head coefficients and reapply the sparsemax transformation
# to see the actual weight that each head received.
# +
import numpy as np
from matplotlib import pyplot as plt
from entmax_jax.activations import sparsemax
# 12 layers * 12 heads
head_coeffs = sparsemax(expl_params['params']['head_coeffs']).reshape(12, 12)
head_coeffs = np.asarray(head_coeffs)
fig, ax = plt.subplots(figsize=(4, 4))
ax.set_xticks(list(range(12)))
ax.set_yticks(list(range(12)))
ax.set_xlabel('Head')
ax.set_ylabel('Layer')
ax.set_title('Head coefficients')
ax.imshow(head_coeffs, cmap='Greens')
# -
# As expected, the explainer learns that some heads are more relevant than others.
# Sparsemax help to filter out irrelevant heads by assigning zero probability (this
# becomes more evidenced as training progresses).
# ## Visualizing the explanation
#
# First we will define a helper function to print HTML stuff.
def show_explanation(tokens, expl_scores, method='smat'):
from IPython.display import display, HTML
span_style_pos = 'color: black; background-color: rgba(3, 252, 94, {}); display:inline-block;'
span_style_neg = 'color: black; background-color: rgba(252, 161, 3, {}); display:inline-block;'
template_pos = '<span style="'+span_style_pos+'"> {} </span>'
template_neg = '<span style="'+span_style_neg+'"> {} </span>'
colored_string = ''
f = lambda w: w.replace('<', 'ᐸ').replace('>', 'ᐳ')
for token, score in zip(tokens, expl_scores):
if score >= 0:
colored_string += template_pos.format(score, f(token))
else:
colored_string += template_neg.format(-score, f(token))
html_text = '<div style="width:100%;">{}: {}</div>'.format(method, colored_string)
display(HTML(html_text))
# Now we can try just call the `show_explanation` function to see
# our explanations as highlights. Note that we can try many different
# strategies for normalizing explainability scores.
# +
# normalization strategies
sum_norm = lambda v: v / v.sum()
minmax_norm = lambda v: (v - v.min()) / (v.max() - v.min())
std_norm = lambda v: (v - v.mean()) / v.std()
abs_norm = lambda v: v / np.max(np.abs(v))
# recover data
input_ids = np.asarray(batch['input_ids'])[0]
input_tokens = tokenizer.convert_ids_to_tokens(input_ids)
pred_expl = np.asarray(expl)[0]
pred_label = y_pred.argmax()
# remove <s> and </s> and show the explanation as highlights
show_explanation(
tokens=input_tokens[1:-1],
expl_scores=minmax_norm(pred_expl[1:-1]),
method='SMAT ({})'.format(pred_label)
)