Error occurs when injecting faults with Resnet50 using ImageNet dataset

[sabuj7177]

Hi, I was trying to inject faults with TensorFI2 in different pre-trained models in Keras with the ImageNet dataset. I was able to successfully inject faults with VGG16, VGG19, and MobileNet. However, when injecting faults in Resnet50, Resnet101, Resnet152, NasnetMobile, etc, I got the following error.

Traceback (most recent call last):
  File "resnet-imagenet.py", line 42, in <module>
    res = tfi.inject(model=model, x_test=image, confFile=conf)
  File "/home/sabuj/TensorFI2/src/tensorfi2.py", line 42, in __init__
    fiFunc(model, fiConf, **kwargs)
  File "/home/sabuj/TensorFI2/src/tensorfi2.py", line 223, in layer_outputs
    pred = get_pred([fiLayerOutputs])
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/keras/backend.py", line 3792, in __call__
    outputs = self._graph_fn(*converted_inputs)
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 1605, in __call__
    return self._call_impl(args, kwargs)
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 1645, in _call_impl
    return self._call_flat(args, self.captured_inputs, cancellation_manager)
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 1746, in _call_flat
    ctx, args, cancellation_manager=cancellation_manager))
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/eager/function.py", line 598, in call
    ctx=ctx)
  File "/home/sabuj/anaconda3/envs/tfi2/lib/python3.7/site-packages/tensorflow/python/eager/execute.py", line 60, in quick_execute
    inputs, attrs, num_outputs)
tensorflow.python.framework.errors_impl.InvalidArgumentError:  You must feed a value for placeholder tensor 'input_1' with dtype float and shape [?,224,224,3]
     [[node input_1 (defined at /home/sabuj/TensorFI2/src/tensorfi2.py:222) ]] [Op:__inference_keras_scratch_graph_10471]

The problem is, this error is not generating every time. When the layer number for fault injection is low, there is no error. For example, when I injected faults in the first 6 layers for Resnet50, there was no error(I changed random layer number to fixed layer number in tensorfi2 implementation for finding the reason). When injecting faults in the later layers, most of the time, it throws errors.

I ran the experiments/layer_outputs/resnet-imagenet.py file for my testing. However, I simplified it for ease of my testing. I am adding the code snippet here.

import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'

import tensorflow as tf

import numpy as np
import random
from tensorflow.keras.preprocessing.image import load_img, img_to_array
from tensorflow.keras.applications.resnet import preprocess_input, decode_predictions

import time, sys, math

from src import tensorfi2 as tfi

model = tf.keras.applications.ResNet50(
    include_top=True, weights='imagenet', input_tensor=None,
    input_shape=None, pooling=None, classes=1000)

model.compile(optimizer='sgd', loss='categorical_crossentropy')

#model.save_weights('h5/vgg16-trained.h5')

path = 'dog.jpg'
image = load_img(path, target_size=(224, 224))
image = img_to_array(image)
image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
image = preprocess_input(image)
out = model.predict(image)
label = decode_predictions(out)
label = label[0][0]
print(label)

conf = 'confFiles/sample.yaml'

numFaults = 10
for i in range(numFaults):
    # model.load_weights('h5/resnet-trained.h5')
    image = load_img(path, target_size=(224, 224))
    image = img_to_array(image)
    image = image.reshape((1, image.shape[0], image.shape[1], image.shape[2]))
    image = preprocess_input(image)
    res = tfi.inject(model=model, x_test=image, confFile=conf)

Here 'dog.jpg' is a random file, that I used for testing. The confFiles/sample.yml is added below.

Target: layer_outputs
Mode: single
Type: bitflips
Amount: 1
Bit: N

I tried this both in windows and ubuntu 20.04. I used both conda and pip environments for my experiment. I found the same issue in all the cases. I am adding my env file for reference.

# packages in environment at /home/sabuj/anaconda3/envs/tfi2:
#
# Name                    Version                   Build  Channel
_libgcc_mutex             0.1                        main  
_openmp_mutex             4.5                       1_gnu  
_tflow_select             2.1.0                       gpu  
absl-py                   0.15.0             pyhd3eb1b0_0  
aiohttp                   3.8.1            py37h7f8727e_0  
aiosignal                 1.2.0              pyhd3eb1b0_0  
astor                     0.8.1            py37h06a4308_0  
astunparse                1.6.3                      py_0  
async-timeout             4.0.1              pyhd3eb1b0_0  
asynctest                 0.13.0                     py_0  
attrs                     21.4.0             pyhd3eb1b0_0  
blas                      1.0                         mkl  
blinker                   1.4              py37h06a4308_0  
brotlipy                  0.7.0           py37h27cfd23_1003  
c-ares                    1.18.1               h7f8727e_0  
ca-certificates           2021.10.26           h06a4308_2  
cachetools                4.2.2              pyhd3eb1b0_0  
certifi                   2021.10.8        py37h06a4308_2  
cffi                      1.15.0           py37hd667e15_1  
charset-normalizer        2.0.4              pyhd3eb1b0_0  
click                     8.0.3              pyhd3eb1b0_0  
cryptography              3.4.8            py37hd23ed53_0  
cudatoolkit               10.1.243             h6bb024c_0  
cudnn                     7.6.5                cuda10.1_0  
cupti                     10.1.168                      0  
dataclasses               0.8                pyh6d0b6a4_7  
frozenlist                1.2.0            py37h7f8727e_0  
gast                      0.3.3                      py_0  
google-auth               1.33.0             pyhd3eb1b0_0  
google-auth-oauthlib      0.4.4              pyhd3eb1b0_0  
google-pasta              0.2.0              pyhd3eb1b0_0  
grpcio                    1.42.0           py37hce63b2e_0  
h5py                      2.10.0           py37hd6299e0_1  
hdf5                      1.10.6               hb1b8bf9_0  
idna                      3.3                pyhd3eb1b0_0  
importlib-metadata        4.8.2            py37h06a4308_0  
intel-openmp              2021.4.0          h06a4308_3561  
keras-preprocessing       1.1.2              pyhd3eb1b0_0  
ld_impl_linux-64          2.35.1               h7274673_9  
libffi                    3.3                  he6710b0_2  
libgcc-ng                 9.3.0               h5101ec6_17  
libgfortran-ng            7.5.0               ha8ba4b0_17  
libgfortran4              7.5.0               ha8ba4b0_17  
libgomp                   9.3.0               h5101ec6_17  
libprotobuf               3.19.1               h4ff587b_0  
libstdcxx-ng              9.3.0               hd4cf53a_17  
markdown                  3.3.4            py37h06a4308_0  
mkl                       2021.4.0           h06a4308_640  
mkl-service               2.4.0            py37h7f8727e_0  
mkl_fft                   1.3.1            py37hd3c417c_0  
mkl_random                1.2.2            py37h51133e4_0  
multidict                 5.2.0            py37h7f8727e_2  
ncurses                   6.3                  h7f8727e_2  
numpy                     1.21.2           py37h20f2e39_0  
numpy-base                1.21.2           py37h79a1101_0  
oauthlib                  3.1.1              pyhd3eb1b0_0  
openssl                   1.1.1m               h7f8727e_0  
opt_einsum                3.3.0              pyhd3eb1b0_1  
pillow                    9.0.1                    pypi_0    pypi
pip                       21.2.2           py37h06a4308_0  
protobuf                  3.19.1           py37h295c915_0  
pyasn1                    0.4.8              pyhd3eb1b0_0  
pyasn1-modules            0.2.8                      py_0  
pycparser                 2.21               pyhd3eb1b0_0  
pyjwt                     2.1.0            py37h06a4308_0  
pyopenssl                 21.0.0             pyhd3eb1b0_1  
pysocks                   1.7.1                    py37_1  
python                    3.7.9                h7579374_0  
pyyaml                    5.3.1                    pypi_0    pypi
readline                  8.1.2                h7f8727e_1  
requests                  2.27.1             pyhd3eb1b0_0  
requests-oauthlib         1.3.0                      py_0  
rsa                       4.7.2              pyhd3eb1b0_1  
scipy                     1.7.3            py37hc147768_0  
setuptools                58.0.4           py37h06a4308_0  
six                       1.16.0             pyhd3eb1b0_0  
sqlite                    3.37.0               hc218d9a_0  
tensorboard               2.4.0              pyhc547734_0  
tensorboard-plugin-wit    1.6.0                      py_0  
tensorflow                2.2.0           gpu_py37h1a511ff_0  
tensorflow-base           2.2.0           gpu_py37h8a81be8_0  
tensorflow-estimator      2.6.0              pyh7b7c402_0  
tensorflow-gpu            2.2.0                h0d30ee6_0  
termcolor                 1.1.0            py37h06a4308_1  
tk                        8.6.11               h1ccaba5_0  
typing-extensions         3.10.0.2             hd3eb1b0_0  
typing_extensions         3.10.0.2           pyh06a4308_0  
urllib3                   1.26.8             pyhd3eb1b0_0  
werkzeug                  2.0.2              pyhd3eb1b0_0  
wheel                     0.37.1             pyhd3eb1b0_0  
wrapt                     1.13.3           py37h7f8727e_2  
xz                        5.2.5                h7b6447c_0  
yarl                      1.6.3            py37h27cfd23_0  
zipp                      3.7.0              pyhd3eb1b0_0  
zlib                      1.2.11               h7f8727e_4

Any solution of this error will be very helpful for me. Thank you.

[nniranjhana]

Thanks @sabuj7177 for the detailed observations and logs. I think this is related to issue #2. I am not working on this presently, but if you are able to fix this, feel free to send a commit and I can take a look.