TVMError: relay.frontend.from_keras() fails with models compiled in tensorflow 2.3

[iswariyam]

The following error occurs when feeding a tensorflow 2.3 compiled model to relay.frontend.from_keras()

AttributeError: 'Node' object has no attribute 'node_indices'

The following code can be used to reproduce the error.

import numpy as np
import tensorflow as tf
import tensorflow.keras as keras
from matplotlib import pyplot as plt
from PIL import Image
from tensorflow.keras.applications.resnet50 import preprocess_input

import tvm
import tvm.relay as relay
from tvm import te
from tvm.contrib.download import download_testdata

keras_resnet50 = keras.applications.resnet50.ResNet50(
    include_top=True, input_shape=(224, 224, 3), classes=1000
)

img_url = "https://github.com/dmlc/mxnet.js/blob/master/data/cat.png?raw=true"
img_path = download_testdata(img_url, "cat.png", module="data")
img = Image.open(img_path).resize((224, 224))

# input preprocess
data = np.array(img)[np.newaxis, :].astype("float32")
data = preprocess_input(data).transpose([0, 3, 1, 2])
print("input_1", data.shape)

shape_dict = {"input_1": data.shape}
mod, params = relay.frontend.from_keras(keras_resnet50, shape_dict)

Reason

The issue occurs due to changes in the attributes of the Node class in tensorflow 2.3 as compared to tensorflow 2.2

Environment:

• Python : 3.7.7
• Tensorflow : 2.3
• TVM : 0.7.dev1
• OS : Ubuntu 16.04