-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel_intervention.py
299 lines (243 loc) · 11.8 KB
/
model_intervention.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
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from transformer_lens import HookedTransformer, utils
import copy
from datasets import load_dataset
import pandas as pd
from tqdm import tqdm
import torch.nn.functional as F
import re
torch.set_grad_enabled(False)
def _extract_prefix_before_layer_number(name):
"""
Extracts the prefix before the layer number in a parameter name.
Args:
name (str): The name of the parameter.
Returns:
str: The prefix before the layer number.
"""
parts = name.split('.')
prefix_parts = []
for part in parts:
if re.match(r'^\d+$', part): # Check if the part is a number
break
prefix_parts.append(part)
return '.'.join(prefix_parts)
def _extract_layer_prefixes(model):
"""
Extracts the unique layer prefixes from a model's parameters.
Args:
model (torch.nn.Module): The model from which to extract layer prefixes.
Returns:
str: The unique prefix for the layers.
"""
for name, param in model.named_parameters():
prefix = _extract_prefix_before_layer_number(name)
if 'weight' not in prefix and 'bias' not in prefix:
return prefix
class ModelExperiment:
"""
Class to perform experiments on a transformer model such as layer swapping and ablation.
Attributes:
model_name (str): The name of the model to use.
device (str): The device to use for computation.
model (AutoModelForCausalLM): The original transformer model.
tokenizer (AutoTokenizer): The tokenizer associated with the model.
original_state_dict (OrderedDict): The original state dictionary of the model.
hooked_model (HookedTransformer): The hooked version of the model for intervention experiments.
"""
def __init__(self, model_name="openai-community/gpt2", device='cuda:1'):
self.model_name = model_name
self.device = device
self.model = AutoModelForCausalLM.from_pretrained(model_name)
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
self.original_state_dict = copy.deepcopy(self.model.state_dict())
self.hooked_model = HookedTransformer.from_pretrained(self.model_name.split("/")[-1], device=self.device, tokenizer=self.tokenizer, hf_model=self.model, center_unembed=True, center_writing_weights=True, fold_ln=True)
self.hooked_model.eval()
def reset_to_original_model(self):
"""
Resets the model to its original state.
"""
self.model.load_state_dict(self.original_state_dict)
self.hooked_model = HookedTransformer.from_pretrained(self.model_name.split("/")[-1], device=self.device, tokenizer=self.tokenizer, hf_model=self.model, center_unembed=True, center_writing_weights=True, fold_ln=True)
self.hooked_model.eval()
def swap_model_layers(self, layer_a, layer_b):
"""
Swaps two layers in the model.
Args:
layer_a (int): The index of the first layer.
layer_b (int): The index of the second layer.
"""
def swap_state_dict_layers(model_state_dict, layer_a, layer_b):
prefix = _extract_layer_prefixes(self.model)
layer_a_pattern = f"{prefix}.{layer_a}."
layer_b_pattern = f"{prefix}.{layer_b}."
result_dict = copy.deepcopy(model_state_dict)
temp_storage = {}
for key in model_state_dict.keys():
if layer_a_pattern in key:
new_key = key.replace(layer_a_pattern, layer_b_pattern)
temp_storage[new_key] = model_state_dict[key]
elif layer_b_pattern in key:
new_key = key.replace(layer_b_pattern, layer_a_pattern)
temp_storage[new_key] = model_state_dict[key]
for temp_key, value in temp_storage.items():
result_dict[temp_key] = value
return result_dict
new_dict = swap_state_dict_layers(self.original_state_dict, layer_a, layer_b)
self.model.load_state_dict(new_dict)
self.hooked_model = HookedTransformer.from_pretrained(self.model_name.split("/")[-1], device=self.device, tokenizer=self.tokenizer, hf_model=self.model, center_unembed=True, center_writing_weights=True, fold_ln=True)
self.hooked_model.eval()
def ablate_model_layer(self, text, layer):
"""
Ablates (sets to zero) the outputs of a specific layer.
Args:
text (str): The input text to the model.
layer (int): The index of the layer to ablate.
Returns:
tuple: The logits and loss after ablation.
"""
def zero_out_layer_hook(value, hook):
value[:, :, :] = 0.
return value
tokens = self.hooked_model.to_tokens(text, prepend_bos=True)
logits = self.hooked_model.run_with_hooks(
tokens,
return_type="logits",
fwd_hooks=[
(utils.get_act_name("attn_out", layer), zero_out_layer_hook),
(utils.get_act_name("mlp_out", layer), zero_out_layer_hook)
]
)
loss = self.hooked_model.loss_fn(logits, tokens, per_token=True)
self.hooked_model.reset_hooks()
return logits, loss
def compute_logits_and_loss(self, input):
"""
Computes the logits and loss for a given input.
Args:
input (str): The input text to the model.
Returns:
tuple: The logits and loss for the input text.
"""
logits = self.hooked_model(input)
loss = self.hooked_model.loss_fn(logits, self.hooked_model.to_tokens(input, prepend_bos=True), per_token=True)
return logits, loss
class ModelMetrics:
"""
Class to compute various metrics for model evaluation.
"""
@staticmethod
def compute_kl_divergence(logit_p: torch.Tensor, logit_q: torch.Tensor, dim: int = -1) -> torch.Tensor:
"""
Computes the Kullback-Leibler (KL) divergence between two sets of logits.
Args:
logit_p (torch.Tensor): The first set of logits.
logit_q (torch.Tensor): The second set of logits.
dim (int): The dimension over which to compute the KL divergence.
Returns:
torch.Tensor: The KL divergence.
"""
log_p = logit_p.log_softmax(dim)
log_q = logit_q.log_softmax(dim)
return torch.sum(log_p.exp() * (log_p - log_q), dim)
@staticmethod
def compute_base2_entropy(logits):
"""
Computes the base-2 entropy of the logits.
Args:
logits (torch.Tensor): The logits from the model.
Returns:
torch.Tensor: The entropy.
"""
probs = F.softmax(logits, dim=-1)
probs = torch.clamp(probs, min=1e-9)
log_probs = torch.log(probs) / torch.log(torch.tensor(2.0))
entropy = -torch.sum(probs * log_probs, dim=-1)
return entropy
def run_layer_intervention_experiment(model, intervention_type='swap', num_samples=1388):
"""
Runs a layer intervention experiment by swapping or ablating layers in the model.
Args:
model (ModelExperiment): The model to run the experiment on.
intervention_type (str): The type of intervention ('swap' or 'ablate').
num_samples (int): The number of samples to use in the experiment.
Returns:
pd.DataFrame: A dataframe containing the results of the experiment.
"""
torch.cuda.empty_cache()
dataset = load_dataset("EleutherAI/the_pile_deduplicated", split='train', streaming=True)
kl_divs = []
losses = []
layers = []
token_normal = []
token_intervened = []
token = []
entropy_normal = []
entropy_intervened = []
loss_normals = []
loss_interveneds = []
metrics = ModelMetrics()
normal_model = model
intervened_model = ModelExperiment(model_name=model.model_name, device=model.device)
n_layers = normal_model.hooked_model.cfg.n_layers
layer_range = range(n_layers - 1) if intervention_type == 'swap' else range(n_layers)
for i in tqdm(layer_range):
if intervention_type == 'swap':
intervened_model.swap_model_layers(i, i+1)
for sample in dataset.take(num_samples):
logits_normal, loss_normal = normal_model.compute_logits_and_loss(sample["text"])
if intervention_type == 'swap':
logits_intervened, loss_intervened = intervened_model.compute_logits_and_loss(sample["text"])
else: # ablation
logits_intervened, loss_intervened = normal_model.ablate_model_layer(sample["text"], i)
token.extend([normal_model.hooked_model.to_string(tkn) for tkn in normal_model.hooked_model.to_tokens(sample["text"], prepend_bos=True)[0, 1:]])
logits_normal = logits_normal[0, 1:].cpu()
logits_intervened = logits_intervened[0, 1:].cpu()
loss_normal = loss_normal.cpu()
loss_intervened = loss_intervened.cpu()
loss_normals.extend(loss_normal.squeeze(0).tolist())
loss_interveneds.extend(loss_intervened.squeeze(0).tolist())
kl_div = metrics.compute_kl_divergence(logits_normal, logits_intervened, dim=-1)
kl_divs.extend(kl_div.tolist())
losses.extend((loss_normal - loss_intervened).squeeze(0).tolist())
layers.extend([i] * loss_normal.shape[1])
entropy_normal.extend(metrics.compute_base2_entropy(logits_normal).tolist())
entropy_intervened.extend(metrics.compute_base2_entropy(logits_intervened).tolist())
token_normal.extend([normal_model.hooked_model.to_string(tkn) for tkn in logits_normal.argmax(-1).tolist()])
token_intervened.extend([intervened_model.hooked_model.to_string(tkn) for tkn in logits_intervened.argmax(-1).tolist()])
if intervention_type == 'swap':
intervened_model.reset_to_original_model()
results = pd.DataFrame({
'Token': token,
'KL Divergence': kl_divs,
'Layer Intervened':[s for s in layers],
'Loss Difference': losses,
'Token Normal': token_normal,
'Token Intervened': token_intervened,
'Loss Normal': loss_normals,
'Loss Intervened': loss_interveneds,
'Entropy Normal': entropy_normal,
'Entropy Intervened': entropy_intervened,
})
return results
if __name__ == "__main__":
#For Phi Models do: (See Github Issue)
# hf_model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32, trust_remote_code=True)
# tokenizer = AutoTokenizer.from_pretrained(model_name, add_bos_token = True, use_fast=False, trust_remote_code=True)
#Put / before name, or use HuggingFace model names {family}/{model_name}
#Make sure transformer_lens supports the model https://transformerlensorg.github.io/TransformerLens/generated/model_properties_table.html
#Consider fixing the number of tokens that the model processes or remove the end of sentence token for analysis
model = ModelExperiment(model_name="EleutherAI/pythia-410m-deduped", device='cuda:1')
print("Running swap experiment...")
swap_results = run_layer_intervention_experiment(model, intervention_type='swap', num_samples=10)
swap_results.to_csv("swap_experiment_results.csv", index=False)
print(swap_results.head())
#Save
#swap_results.to_csv("swap_experiment_results.csv", index
print("\nRunning ablation experiment...")
ablation_results = run_layer_intervention_experiment(model, intervention_type='ablate', num_samples=10)
ablation_results.to_csv("ablation_experiment_results.csv", index=False)
print(ablation_results.head())
#Save
#ablation_results.to_csv("ablation_experiment_results.csv", index=False)