Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -85,78 +85,23 @@ def sequence_to_kmer_vector(sequence: str, k: int = 4) -> np.ndarray:
|
|
85 |
# 3. SHAP-VALUE (ABLATION) CALCULATION
|
86 |
###############################################################################
|
87 |
|
88 |
-
def calculate_shap_values(model, x_tensor
|
89 |
-
"""
|
90 |
-
Calculate feature attributions using Integrated Gradients with multiple baselines.
|
91 |
-
|
92 |
-
Args:
|
93 |
-
model: A PyTorch model.
|
94 |
-
x_tensor: Input tensor of shape (1, num_features).
|
95 |
-
baselines: A list of baseline tensors, each of shape (1, num_features).
|
96 |
-
If None, defaults to n_baselines copies of the zero vector.
|
97 |
-
steps: Number of interpolation steps between the baseline and the input.
|
98 |
-
n_baselines: Number of baselines to use if baselines is None.
|
99 |
-
|
100 |
-
Returns:
|
101 |
-
avg_attributions: A numpy array of shape (num_features,) with averaged feature attributions.
|
102 |
-
avg_full_prob: The model's predicted probability for the target class ('human')
|
103 |
-
computed on the full input, averaged over baselines.
|
104 |
-
"""
|
105 |
model.eval()
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
112 |
-
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
full_prob = torch.softmax(full_output, dim=1)[0, 1].item()
|
122 |
-
full_probs.append(full_prob)
|
123 |
-
|
124 |
-
# Create interpolated inputs from baseline to x_tensor.
|
125 |
-
scaled_inputs = [
|
126 |
-
baseline + (float(i) / steps) * (x_tensor - baseline)
|
127 |
-
for i in range(steps + 1)
|
128 |
-
]
|
129 |
-
scaled_inputs = torch.cat(scaled_inputs, dim=0) # Shape: (steps+1, num_features)
|
130 |
-
scaled_inputs.requires_grad = True
|
131 |
-
|
132 |
-
# Forward pass: compute outputs and target class probabilities for all interpolated inputs.
|
133 |
-
outputs = model(scaled_inputs)
|
134 |
-
probs = torch.softmax(outputs, dim=1)[:, 1] # Probabilities for 'human' class
|
135 |
-
|
136 |
-
# Backward pass: compute gradients of the probabilities with respect to inputs.
|
137 |
-
grads = torch.autograd.grad(
|
138 |
-
outputs=probs,
|
139 |
-
inputs=scaled_inputs,
|
140 |
-
grad_outputs=torch.ones_like(probs),
|
141 |
-
create_graph=False,
|
142 |
-
retain_graph=False
|
143 |
-
)[0] # Shape: (steps+1, num_features)
|
144 |
-
|
145 |
-
# Approximate the integral using the trapezoidal rule.
|
146 |
-
avg_grads = (grads[:-1] + grads[1:]) / 2.0 # Average gradients between successive steps.
|
147 |
-
integrated_grad = avg_grads.mean(dim=0, keepdim=True) # Mean over all steps.
|
148 |
-
|
149 |
-
# Multiply by the input difference to get attributions.
|
150 |
-
attributions = (x_tensor - baseline) * integrated_grad # Shape: (1, num_features)
|
151 |
-
all_attributions.append(attributions)
|
152 |
-
|
153 |
-
# Average attributions over all baselines.
|
154 |
-
avg_attributions = torch.stack(all_attributions, dim=0).mean(dim=0)
|
155 |
-
avg_full_prob = np.mean(full_probs)
|
156 |
-
|
157 |
-
return avg_attributions.squeeze().cpu().detach().numpy(), avg_full_prob
|
158 |
-
|
159 |
-
|
160 |
|
161 |
|
162 |
###############################################################################
|
|
|
85 |
# 3. SHAP-VALUE (ABLATION) CALCULATION
|
86 |
###############################################################################
|
87 |
|
88 |
+
def calculate_shap_values(model, x_tensor):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
model.eval()
|
90 |
+
with torch.no_grad():
|
91 |
+
baseline_output = model(x_tensor)
|
92 |
+
baseline_probs = torch.softmax(baseline_output, dim=1)
|
93 |
+
baseline_prob = baseline_probs[0, 1].item() # Prob of 'human'
|
94 |
+
shap_values = []
|
95 |
+
x_zeroed = x_tensor.clone()
|
96 |
+
for i in range(x_tensor.shape[1]):
|
97 |
+
original_val = x_zeroed[0, i].item()
|
98 |
+
x_zeroed[0, i] = 0.0
|
99 |
+
output = model(x_zeroed)
|
100 |
+
probs = torch.softmax(output, dim=1)
|
101 |
+
prob = probs[0, 1].item()
|
102 |
+
shap_values.append(baseline_prob - prob)
|
103 |
+
x_zeroed[0, i] = original_val
|
104 |
+
return np.array(shap_values), baseline_prob
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
105 |
|
106 |
|
107 |
###############################################################################
|