Missing one conv2d layer in inceptionv3?

[LucaUrbinati44]

Summary

I think that there is a missing layer in the implementation of Inceptionv3.

According to the paper C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” arXiv:1512.00567 [cs], Dec. 2015, Accessed: Dec. 30, 2020. [Online]. Available: http://arxiv.org/abs/1512.00567, in Table 1 there should be a conv2d layer that takes 35x35x192 as input and provides 35x35x288 as output.

To have a graphical view, the layer I'm point to is the sixth red conv2d block in this picture showing the architecture of Inceptionv3: https://hackmd.io/@bouteille/SkD5Xd4DL#I-Summary.

The missing layer should be place between line 175 and 176 of the file: https://github.com/keras-team/keras-applications/blob/master/keras_applications/inception_v3.py, and it should be: x = conv2d_bn(x, 288, 3, 3, padding='same')

Moreover, I don't understand why the concatenation named "mixed0" outputs 35x35x256 instead of 35x35x288. In this case the line to edit should be line 191 from this: branch_pool = conv2d_bn(branch_pool, 32, 1, 1) to this: branch_pool = conv2d_bn(branch_pool, 64, 1, 1)

Let me know if I had an hallucination or it is really missing. Thank you!!!

Environment

• Python version: 3.6.9
• Keras version: 2.4.3
• Keras-applications version: 1.0.8
• Keras backend with version: 10.1

Logs or source codes for reproduction

import tensorflow as tf

from tensorflow.keras.applications import InceptionV3

from tensorflow.keras.models import Sequential
from tensorflow.keras import layers

IMG_SIZE = 299

def build_model():
    inputs = layers.Input(shape=(IMG_SIZE, IMG_SIZE, 3))
    model = InceptionV3(include_top=True, weights=None, input_tensor=inputs)
    optimizer = tf.keras.optimizers.Adam(learning_rate=1e-2)
    model.compile(
        optimizer=optimizer, loss="categorical_crossentropy", metrics=["accuracy"]
    )
    return model

model = build_model()
model.summary()