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 | class TfModel(Model):
"""This is a wrapper for TensorFlow / Keras models"""
def __init__(self, input_model:tf.keras.Model) -> None:
"""Wraps a TensorFlow model to the superclass"""
if not isinstance(input_model, tf.keras.Model):
raise TypeError("input_model must be an instance of tf.keras.Model")
super().__init__(input_model)
if not self.model.built:
if hasattr(self.model, "input_shape") and self.model.input_shape:
self.model.build(self.model.input_shape)
self.dtype = self.model.input_dtype
def __call__(self, X:npt.ArrayLike) -> npt.NDArray[Any]:
"Model inference"
X_tensor = tf.convert_to_tensor(X, dtype=self.dtype)
output = self.model(X_tensor, training=False)
return output.numpy()
def activation_at_layer(self, X:npt.NDArray, layer_identifier: int | str | list[int | str]) -> list[npt.NDArray]:
"""
Forward passes input X up to the specified layer.
layer_identifier can be an integer index or layer name string.
"""
layer_identifier = self._get_layer_idx(layer_identifier)
X_tensor = tf.convert_to_tensor(X, dtype=self.dtype)
if not isinstance(layer_identifier, list):
layer_identifier = [layer_identifier]
if isinstance(self.model, keras.Sequential):
activation_map = {}
activations = []
curr = X_tensor
for i, layer in enumerate(self.model.layers):
curr = layer(curr)
activation_map[i] = curr
for layer_idx in layer_identifier:
activations.append(activation_map[layer_idx].numpy())
else:
intermediate_model = tf.keras.Model(
inputs=self.model.input,
outputs=layer_identifier
)
activations = intermediate_model(X_tensor, training=False)
if isinstance(activations, list):
activations = [a.numpy() for a in activations]
return activations
def compute_gradients(
self,
X: npt.NDArray | tf.Tensor,
objective: int | str | Tuple[int, int],
objective_fn: Callable[[tf.Tensor], tf.Tensor],
wrt: int | str | Tuple[int, int] |None = None,
):
r"""Computes the gradient \(\frac{\partial \text{Objective}}{\partial X}\).
`objective` is Objective, `wrt` is X.
The gradient is calculated after applying the objective function to the objective layer w.r.t the wrt layer.
Tuple forms of `objective` or `wrt` specifies specific neurons, e.g. (0, 1) means the second neuron of the first layer.
Note: When no `wrt` is provided, it will default to `'input'`.
"""
if isinstance(X, tf.Tensor):
X_tensor = tf.cast(X, dtype=self.dtype)
else:
X_tensor = tf.convert_to_tensor(X, dtype=self.dtype)
is_wrt_input = wrt is None or wrt == "input"
obj_idx = self._get_layer_idx(objective)
wrt_idx = None if is_wrt_input else self._get_layer_idx(wrt)
if isinstance(self.model, keras.Sequential):
with tf.GradientTape() as tape:
curr = X_tensor
if is_wrt_input:
tape.watch(X_tensor)
wrt_act = X_tensor
else:
wrt_act = None
for i, layer in enumerate(self.model.layers):
curr = layer(curr)
if not is_wrt_input and i == wrt_idx:
wrt_act = curr
tape.watch(wrt_act)
if i == obj_idx:
obj_act = curr
break
if isinstance(objective, Tuple):
loss = objective_fn(tf.squeeze(obj_act)[objective[1]])
else:
loss = objective_fn(obj_act)
target_tensor = X_tensor if is_wrt_input else wrt_act
grad = tape.gradient(loss, target_tensor)
else:
obj_output = self.model.layers[obj_idx].output
if is_wrt_input:
intermediate_model = tf.keras.Model(
inputs=self.model.input,
outputs=obj_output
)
with tf.GradientTape() as tape:
tape.watch(X_tensor)
obj_act = intermediate_model(X_tensor, training=False)
if isinstance(objective, Tuple):
loss = objective_fn(tf.squeeze(obj_act)[objective[1]])
else:
loss = objective_fn(obj_act)
grad = tape.gradient(loss, X_tensor)
else:
wrt_output = self.model.layers[wrt_idx].output
intermediate_model = tf.keras.Model(
inputs=self.model.input,
outputs=[wrt_output, obj_output]
)
with tf.GradientTape() as tape:
wrt_act, obj_act = intermediate_model(X_tensor, training=False)
tape.watch(wrt_act)
if isinstance(objective, Tuple):
loss = objective_fn(tf.squeeze(obj_act)[objective[1]])
else:
loss = objective_fn(obj_act)
grad = tape.gradient(loss, wrt_act)
if grad is None:
raise RuntimeError(
f"Gradients could not be computed for objective layer '{objective}' "
f"with respect to '{wrt}'."
)
if isinstance(wrt, Tuple):
grad = tf.squeeze(grad)[wrt[1]]
return grad.numpy()
def _get_layer_idx(self, identifier: int | str | list[int | str] | Tuple[int, int]) -> int:
if isinstance(identifier, int):
if identifier < 0 or identifier >= len(self.model.layers):
raise ValueError(f"Layer index '{identifier}' out of bounds.")
return identifier
if isinstance(identifier, Tuple):
return identifier[0]
if isinstance(identifier, list):
layer_map = {}
layer_indices = []
for idx, layer in enumerate(self.model.layers):
layer_map[layer] = idx
for layer in identifier:
if isinstance(layer, int):
layer_indices.append(layer)
elif isinstance(layer, str):
layer_indices.append(layer_map[layer])
else:
raise TypeError("Layer identifiers can only be int or str.")
return layer_indices
for idx, layer in enumerate(self.model.layers):
if layer.name == identifier:
return idx
raise ValueError(f"Layer '{identifier}' not found in TensorFlow model.")
|