Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -87,17 +87,18 @@ def sequence_to_kmer_vector(sequence: str, k: int = 4) -> np.ndarray:
|
|
87 |
def calculate_shap_values(model, x_tensor):
|
88 |
model.eval()
|
89 |
device = next(model.parameters()).device
|
90 |
-
|
91 |
try:
|
92 |
-
# Create background as a torch tensor
|
93 |
background = torch.zeros((300, x_tensor.shape[1]), device=device)
|
94 |
explainer = shap.DeepExplainer(model, background)
|
95 |
shap_values_all = explainer.shap_values(x_tensor)
|
96 |
-
# For binary classification,
|
97 |
shap_values = shap_values_all[1][0]
|
98 |
except Exception as e:
|
99 |
print(f"DeepExplainer failed, falling back to KernelExplainer: {str(e)}")
|
100 |
-
|
|
|
101 |
def model_predict(x):
|
102 |
if not isinstance(x, np.ndarray):
|
103 |
x = np.array(x)
|
@@ -106,24 +107,26 @@ def calculate_shap_values(model, x_tensor):
|
|
106 |
with torch.no_grad():
|
107 |
tensor_x = torch.tensor(x, dtype=torch.float, device=device)
|
108 |
output = model(tensor_x)
|
109 |
-
probs = torch.softmax(output, dim=1)[:, 1]
|
110 |
return probs.cpu().numpy()
|
111 |
-
|
112 |
-
#
|
113 |
-
background = np.zeros((300, x_tensor.shape[1]))
|
114 |
-
explainer = shap.KernelExplainer(model_predict, background)
|
115 |
x_numpy = x_tensor.cpu().numpy()
|
|
|
|
|
|
|
|
|
116 |
shap_values = explainer.shap_values(x_numpy, nsamples=1000)
|
117 |
-
# If
|
118 |
if isinstance(shap_values, list):
|
119 |
shap_values = shap_values[0]
|
120 |
-
|
121 |
-
# Get human probability from model
|
122 |
with torch.no_grad():
|
123 |
output = model(x_tensor)
|
124 |
probs = torch.softmax(output, dim=1)
|
125 |
prob_human = probs[0, 1].item()
|
126 |
-
|
127 |
return np.array(shap_values), prob_human
|
128 |
|
129 |
|
|
|
87 |
def calculate_shap_values(model, x_tensor):
|
88 |
model.eval()
|
89 |
device = next(model.parameters()).device
|
90 |
+
|
91 |
try:
|
92 |
+
# Create background as a torch tensor (using zeros may be acceptable for DeepExplainer)
|
93 |
background = torch.zeros((300, x_tensor.shape[1]), device=device)
|
94 |
explainer = shap.DeepExplainer(model, background)
|
95 |
shap_values_all = explainer.shap_values(x_tensor)
|
96 |
+
# For binary classification, get SHAP for class 1 and first sample
|
97 |
shap_values = shap_values_all[1][0]
|
98 |
except Exception as e:
|
99 |
print(f"DeepExplainer failed, falling back to KernelExplainer: {str(e)}")
|
100 |
+
|
101 |
+
# Define a wrapper that ensures proper input shape and conversion to tensor
|
102 |
def model_predict(x):
|
103 |
if not isinstance(x, np.ndarray):
|
104 |
x = np.array(x)
|
|
|
107 |
with torch.no_grad():
|
108 |
tensor_x = torch.tensor(x, dtype=torch.float, device=device)
|
109 |
output = model(tensor_x)
|
110 |
+
probs = torch.softmax(output, dim=1)[:, 1] # human probability
|
111 |
return probs.cpu().numpy()
|
112 |
+
|
113 |
+
# Instead of using zeros as background, use the input sample repeated 300 times.
|
|
|
|
|
114 |
x_numpy = x_tensor.cpu().numpy()
|
115 |
+
background = np.repeat(x_numpy, 300, axis=0)
|
116 |
+
|
117 |
+
explainer = shap.KernelExplainer(model_predict, background)
|
118 |
+
# Increase nsamples for a more robust estimate.
|
119 |
shap_values = explainer.shap_values(x_numpy, nsamples=1000)
|
120 |
+
# If a list is returned, select the first element.
|
121 |
if isinstance(shap_values, list):
|
122 |
shap_values = shap_values[0]
|
123 |
+
|
124 |
+
# Get the human probability from the model output.
|
125 |
with torch.no_grad():
|
126 |
output = model(x_tensor)
|
127 |
probs = torch.softmax(output, dim=1)
|
128 |
prob_human = probs[0, 1].item()
|
129 |
+
|
130 |
return np.array(shap_values), prob_human
|
131 |
|
132 |
|