Error using shap with SimpleRNN sequential model
Asked Answered
B

1

6

In the code below, I import a saved sparse numpy matrix, created with python, densify it, add a masking, batchnorm and dense ouptput layer to a many to one SimpleRNN. The keras sequential model works fine, however, I am unable to use shap. This is run in Jupyter lab from Winpython 3830 on a Windows 10 desktop. The X matrix has a shape of (4754, 500, 64): 4754 examples with 500 timesteps and 64 variables. I've created a function to simulate the data so the code can be tested. The simulated data returns the same error.

from sklearn.model_selection import train_test_split
import tensorflow as tf
from tensorflow.keras.models import Sequential
import tensorflow.keras.backend as Kb
from tensorflow.keras import layers
from tensorflow.keras.layers import BatchNormalization
from tensorflow import keras as K
import numpy as np
import shap
import random

def create_x():
    dims = [10,500,64]
    data = []
    y = []
    for i in range(dims[0]):
        data.append([])

        for j in range(dims[1]):
            data[i].append([])
            for k in range(dims[2]):
                isnp = random.random()
                if isnp > .2:
                    data[i][j].append(np.nan)
                else:
                    data[i][j].append(random.random())
        if isnp > .5:
            y.append(0)
        else:
            y.append(1)
    return np.asarray(data), np.asarray(y)

def first_valid(arr, axis, invalid_val=0):
    #return the 2nd index of 3 for  the first non np.nan on the 3rd axis
    mask = np.invert(np.isnan(arr))
    return np.where(mask.any(axis=axis), mask.argmax(axis=axis), invalid_val)

def densify_np(X):
    X_copy = np.empty_like (X)
    X_copy[:] = X
    #loop over the first index
    for i in range(len(X_copy)):

        old_row = []
        #get the 2nd index of the first valid value for each 3rd index
        indices = first_valid(X_copy[i,:,:],axis=0, invalid_val=0)
        for j in range(len(indices)):
            if np.isnan(X_copy[i,indices[j],j]):
                old_row.append(0)
            else:
                old_row.append(X_copy[i,indices[j],j])
        X_copy[i,0,:]= old_row
        for k in range(1,len(X_copy[i,:])):
            for l in range(len(X_copy[i,k,:])):
                if np.isnan(X_copy[i,k,l]):
                    X_copy[i,k,l] = X_copy[i,k-1,l]
           
    return(X_copy)
#this is what I do in the actual code
#X = np.load('C:/WinPython/WPy64-3830/data/X.npy')
#Y = np.load('C:/WinPython/WPy64-3830/scripts/Y.npy')

#simulated junk data
X, Y = create_x()

#create a dense matrix from the sparse one.
X = densify_np(X)

seed = 7
np.random.seed(seed)
array_size = 64
X_train, X_test, y_train, y_test = train_test_split(X, Y, test_size=0.33, random_state=seed)
batch = 64
model = Sequential()


model.add(layers.Input(shape=(500,array_size)))
model.add(layers.Masking(mask_value=0.,input_shape=(500, array_size)))
model.add(BatchNormalization())
model.add(layers.SimpleRNN(1, activation=None, dropout = 0, recurrent_dropout=.2))
model.add(layers.Dense(1, activation = 'sigmoid'))
opt = K.optimizers.Adam(learning_rate=.001)

model.compile(loss='binary_crossentropy', optimizer=opt)
model.fit(X_train, y_train.astype(int), validation_data=(X_test,y_test.astype(int)), epochs=25, batch_size=batch)

explainer = shap.DeepExplainer(model, X_test)
shap_values = explainer.shap_values(X_train)

Running the last line to create the shap_values yields the error below.

StagingError                              Traceback (most recent call last)
<ipython-input-6-f789203da9c8> in <module>
      1 import shap
      2 explainer = shap.DeepExplainer(model, X_test)
----> 3 shap_values = explainer.shap_values(X_train)
      4 print('done')

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\__init__.py in shap_values(self, X, ranked_outputs, output_rank_order, check_additivity)
    117         were chosen as "top".
    118         """
--> 119         return self.explainer.shap_values(X, ranked_outputs, output_rank_order, check_additivity=check_additivity)

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\deep_tf.py in shap_values(self, X, ranked_outputs, output_rank_order, check_additivity)
    302                 # run attribution computation graph
    303                 feature_ind = model_output_ranks[j,i]
--> 304                 sample_phis = self.run(self.phi_symbolic(feature_ind), self.model_inputs, joint_input)
    305 
    306                 # assign the attributions to the right part of the output arrays

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\deep_tf.py in run(self, out, model_inputs, X)
    359 
    360                 return final_out
--> 361             return self.execute_with_overridden_gradients(anon)
    362 
    363     def custom_grad(self, op, *grads):

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\deep_tf.py in execute_with_overridden_gradients(self, f)
    395         # define the computation graph for the attribution values using a custom gradient-like computation
    396         try:
--> 397             out = f()
    398         finally:
    399             # reinstate the backpropagatable check

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\deep_tf.py in anon()
    355                     v = tf.constant(data, dtype=self.model_inputs[i].dtype)
    356                     inputs.append(v)
--> 357                 final_out = out(inputs)
    358                 tf_execute.record_gradient = tf_backprop._record_gradient
    359 

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\def_function.py in __call__(self, *args, **kwds)
    778       else:
    779         compiler = "nonXla"
--> 780         result = self._call(*args, **kwds)
    781 
    782       new_tracing_count = self._get_tracing_count()

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\def_function.py in _call(self, *args, **kwds)
    821       # This is the first call of __call__, so we have to initialize.
    822       initializers = []
--> 823       self._initialize(args, kwds, add_initializers_to=initializers)
    824     finally:
    825       # At this point we know that the initialization is complete (or less

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\def_function.py in _initialize(self, args, kwds, add_initializers_to)
    694     self._graph_deleter = FunctionDeleter(self._lifted_initializer_graph)
    695     self._concrete_stateful_fn = (
--> 696         self._stateful_fn._get_concrete_function_internal_garbage_collected(  # pylint: disable=protected-access
    697             *args, **kwds))
    698 

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\function.py in _get_concrete_function_internal_garbage_collected(self, *args, **kwargs)
   2853       args, kwargs = None, None
   2854     with self._lock:
-> 2855       graph_function, _, _ = self._maybe_define_function(args, kwargs)
   2856     return graph_function
   2857 

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\function.py in _maybe_define_function(self, args, kwargs)
   3211 
   3212       self._function_cache.missed.add(call_context_key)
-> 3213       graph_function = self._create_graph_function(args, kwargs)
   3214       self._function_cache.primary[cache_key] = graph_function
   3215       return graph_function, args, kwargs

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\function.py in _create_graph_function(self, args, kwargs, override_flat_arg_shapes)
   3063     arg_names = base_arg_names + missing_arg_names
   3064     graph_function = ConcreteFunction(
-> 3065         func_graph_module.func_graph_from_py_func(
   3066             self._name,
   3067             self._python_function,

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\framework\func_graph.py in func_graph_from_py_func(name, python_func, args, kwargs, signature, func_graph, autograph, autograph_options, add_control_dependencies, arg_names, op_return_value, collections, capture_by_value, override_flat_arg_shapes)
    984         _, original_func = tf_decorator.unwrap(python_func)
    985 
--> 986       func_outputs = python_func(*func_args, **func_kwargs)
    987 
    988       # invariant: `func_outputs` contains only Tensors, CompositeTensors,

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\def_function.py in wrapped_fn(*args, **kwds)
    598         # __wrapped__ allows AutoGraph to swap in a converted function. We give
    599         # the function a weak reference to itself to avoid a reference cycle.
--> 600         return weak_wrapped_fn().__wrapped__(*args, **kwds)
    601     weak_wrapped_fn = weakref.ref(wrapped_fn)
    602 

C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\framework\func_graph.py in wrapper(*args, **kwargs)
    971           except Exception as e:  # pylint:disable=broad-except
    972             if hasattr(e, "ag_error_metadata"):
--> 973               raise e.ag_error_metadata.to_exception(e)
    974             else:
    975               raise

StagingError: in user code:

    C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\shap\explainers\deep\deep_tf.py:244 grad_graph  *
        x_grad = tape.gradient(out, shap_rAnD)
    C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\backprop.py:1067 gradient  **
        flat_grad = imperative_grad.imperative_grad(
    C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\imperative_grad.py:71 imperative_grad
        return pywrap_tfe.TFE_Py_TapeGradient(
    C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\eager\backprop.py:151 _gradient_function
        grad_fn = ops._gradient_registry.lookup(op_name)  # pylint: disable=protected-access
    C:\WinPython\WPy64-3830\python-3.8.3.amd64\lib\site-packages\tensorflow\python\framework\registry.py:96 lookup
        raise LookupError(

    LookupError: gradient registry has no entry for: shap_TensorListStack
Besprinkle answered 17/10, 2020 at 10:44 Comment(4)
Seems to be a frequent issue, see e.g. #1110 or #1490.Virtuosic
I've seen those issues. My code generates the same error with the masking and batchnorm removed.Besprinkle
what is your shap, tensorflow and keras version?Psychro
shap 0.35.0, tensoflow 2.3.0. Keras 2.3.1Besprinkle
P
4

The owner of the shap repo said:

The fundamental issue here is that DeepExplainer does not yet support TF 2.0.

That was on 11 Dec 2019. Is this still the case? Try it with Tensorflow 1.15 and see if that works.

Another issue on the shap repo about this (2 Jun 2020) says:

Alright, thank you. I did not see the post by Lundberg. I will stick to the workaround of using TF 1.15 until a new version of SHAP is released.

Psychro answered 24/10, 2020 at 14:36 Comment(0)

© 2022 - 2024 — McMap. All rights reserved.