-
Notifications
You must be signed in to change notification settings - Fork 912
/
torch_deep_neural_classifier_iit.py
197 lines (158 loc) · 7.08 KB
/
torch_deep_neural_classifier_iit.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
from collections import defaultdict
import copy
import numpy as np
import torch
import torch.nn as nn
import torch.utils.data
from torch_deep_neural_classifier import TorchDeepNeuralClassifier
__author__ = "Atticus Geiger"
__version__ = "CS224u, Stanford, Spring 2022"
class IITModel(torch.nn.Module):
def __init__(self, model, layers, id_to_coords,device):
super().__init__()
self.model = model
self.layers = layers
self.id_to_coords = defaultdict(lambda: defaultdict(list))
for k, vals in id_to_coords.items():
for d in vals:
layer = d['layer']
self.id_to_coords[k][layer].append(d)
self.device = device
def no_IIT_forward(self, X):
return self.model(X)
def forward(self, X):
base = X[:,0,:].squeeze(1).type(torch.FloatTensor).to(self.device)
coord_ids = X[:,1,:].squeeze(1).type(torch.FloatTensor).to(self.device)
sources = X[:,2:,:].to(self.device)
sources = [sources[:,j,:].squeeze(1).type(torch.FloatTensor).to(self.device)
for j in range(sources.shape[1])]
gets = self.id_to_coords[int(coord_ids.flatten()[0])]
sets = copy.deepcopy(gets)
self.activation = dict()
for layer in gets:
for i, get in enumerate(gets[layer]):
handlers = self._gets_sets(gets ={layer: [get]},sets = None)
source_logits = self.no_IIT_forward(sources[i])
for handler in handlers:
handler.remove()
sets[layer][i]["intervention"] = self.activation[f'{get["layer"]}-{get["start"]}-{get["end"]}']
base_logits = self.no_IIT_forward(base)
handlers = self._gets_sets(gets = None, sets = sets)
counterfactual_logits = self.no_IIT_forward(base)
for handler in handlers:
handler.remove()
return counterfactual_logits, base_logits
def make_hook(self, gets, sets, layer):
def hook(model, input, output):
layer_gets, layer_sets = [], []
if gets is not None and layer in gets:
layer_gets = gets[layer]
if sets is not None and layer in sets:
layer_sets = sets[layer]
for set in layer_sets:
output = torch.cat([output[:,:set["start"]], set["intervention"], output[:,set["end"]:]], dim = 1)
for get in layer_gets:
self.activation[f'{get["layer"]}-{get["start"]}-{get["end"]}'] = output[:,get["start"]: get["end"] ]
return output
return hook
def _gets_sets(self,gets=None, sets = None):
handlers = []
for layer in range(len(self.layers)):
hook = self.make_hook(gets,sets, layer)
both_handler = self.layers[layer].register_forward_hook(hook)
handlers.append(both_handler)
return handlers
def retrieve_activations(self, input, get, sets):
input = input.type(torch.FloatTensor).to(self.device)
self.activation = dict()
get_val = {get["layer"]: [get]} if get is not None else None
set_val = {sets["layer"]: [sets]} if sets is not None else None
handlers = self._gets_sets(get_val, set_val)
logits = self.model(input)
for handler in handlers:
handler.remove()
return self.activation[f'{get["layer"]}-{get["start"]}-{get["end"]}']
class CrossEntropyLossIIT(nn.Module):
def __init__(self):
super().__init__()
self.loss = nn.CrossEntropyLoss(reduction="mean")
def forward(self, preds, labels):
return self.loss(preds[0], labels[: , 0]) + self.loss(preds[1], labels[:,1])
class TorchDeepNeuralClassifierIIT(TorchDeepNeuralClassifier):
def __init__(self, id_to_coords=None, **base_kwargs):
super().__init__(**base_kwargs)
self.loss = CrossEntropyLossIIT()
self.id_to_coords = id_to_coords
self.shuffle_train = False
def build_graph(self):
model = super().build_graph()
IITmodel = IITModel(model, self.layers, self.id_to_coords, self.device)
return IITmodel
def batched_indices(self, max_len):
batch_indices = [x for x in range((max_len // self.batch_size))]
output = []
while len(batch_indices) != 0:
batch_index = random.sample(batch_indices, 1)[0]
batch_indices.remove(batch_index)
output.append([batch_index*self.batch_size + x for x in range(self.batch_size)])
return output
def build_dataset(self, base, sources, base_y, IIT_y, coord_ids):
base = torch.FloatTensor(np.array(base))
sources = [torch.FloatTensor(np.array(source)) for source in sources]
self.input_dim = base.shape[1]
coord_ids = torch.FloatTensor(np.array(coord_ids))
base_y = np.array(base_y)
self.classes_ = sorted(set(base_y))
self.n_classes_ = len(self.classes_)
class2index = dict(zip(self.classes_, range(self.n_classes_)))
base_y = [class2index[label] for label in base_y]
base_y = torch.tensor(base_y)
IIT_y = np.array(IIT_y)
IIT_y = [class2index[int(label)] for label in IIT_y]
IIT_y = torch.tensor(IIT_y)
bigX = torch.stack([base, coord_ids.unsqueeze(1).expand(-1, base.shape[1])] + sources, dim=1)
bigy = torch.stack((IIT_y, base_y), dim=1)
dataset = torch.utils.data.TensorDataset(bigX, bigy)
return dataset
def prep_input(self, base, sources, coord_ids):
bigX = torch.stack([base, coord_ids.unsqueeze(1).expand(-1, base.shape[1])] + sources, dim=1)
return bigX
def iit_predict(self, base, sources, coord_ids):
IIT_test = self.prep_input(base, sources, coord_ids)
IIT_preds, base_preds = self.model(IIT_test)
IIT_preds = np.array(IIT_preds.argmax(axis=1).cpu())
base_preds = np.array(base_preds.argmax(axis=1).cpu())
return IIT_preds, base_preds
if __name__ == '__main__':
import iit
from sklearn.metrics import classification_report
import utils
utils.fix_random_seeds()
V1 = 0
data_size = 10000
embedding_dim = 4
id_to_coords = {
V1: [{"layer": 1, "start": 0, "end": embedding_dim}]
}
iit_equality_dataset = iit.get_IIT_equality_dataset(
"V1", embedding_dim, data_size)
X_base_train, X_sources_train, y_base_train, y_IIT_train, interventions = iit_equality_dataset
model = TorchDeepNeuralClassifierIIT(
hidden_dim=embedding_dim*4,
hidden_activation=torch.nn.ReLU(),
num_layers=3,
id_to_coords=id_to_coords)
model.fit(
X_base_train,
X_sources_train,
y_base_train,
y_IIT_train,
interventions)
X_base_test, X_sources_test, y_base_test, y_IIT_test, interventions = iit.get_IIT_equality_dataset(
"V1", embedding_dim, 100)
IIT_preds, base_preds = model.iit_predict(
X_base_test, X_sources_test, interventions)
print("\nStandard evaluation")
print(classification_report(y_base_test, base_preds))
print("V1 counterfactual evaluation")
print(classification_report(y_IIT_test, IIT_preds))