class AlexNet(Model):
"""Original AlexNet weights ported to TF.
AlexNet is the breakthrough vision model from Krizhevsky, et al (2012):
https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
This implementation is a caffe re-implementation:
http://www.cs.toronto.edu/~guerzhoy/tf_alexnet/
It was converted to TensorFlow by this GitHub project:
https://github.com/huanzhang12/tensorflow-alexnet-model
It appears the parameters are the actual original parameters.
"""
# The authors of code to convert AlexNet to TF host weights at
# http://jaina.cs.ucdavis.edu/datasets/adv/imagenet/alexnet_frozen.pb
# but it seems more polite and reliable to host our own.
model_path = 'gs://modelzoo/vision/other_models/AlexNet.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [227, 227, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255-IMAGENET_MEAN_BGR)
input_name = 'Placeholder'
# TODO - Sanity check this graph and layers
layers = _layers_from_list_of_dicts([
{'tags': ['conv'], 'name': 'concat_2', 'depth': 256},
{'tags': ['conv'], 'name': 'conv5_1', 'depth': 256},
{'tags': ['dense'], 'name': 'Relu', 'depth': 4096},
{'tags': ['dense'], 'name': 'Relu_1', 'depth': 4096},
{'tags': ['dense'], 'name': 'Softmax', 'depth': 1000},
])
class InceptionV1_caffe(Model):
"""InceptionV1 (or 'GoogLeNet') as reimplemented in caffe.
This model is a reimplementation of GoogLeNet:
https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
reimplemented in caffe by BVLC / Sergio Guadarrama:
https://github.com/BVLC/caffe/tree/master/models/bvlc_googlenet
and then ported to TensorFlow using caffe-tensorflow.
"""
model_path = 'gs://modelzoo/vision/caffe_models/InceptionV1.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255-IMAGENET_MEAN_BGR)
input_name = 'data'
layers = _layers_from_list_of_dicts([
{'tags': ['conv'], 'name': 'conv1_7x7_s2/conv1_7x7_s2', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3_reduce/conv2_3x3_reduce', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3/conv2_3x3', 'depth': 192},
{'tags': ['conv'], 'name': 'inception_3a_output', 'depth': 256},
{'tags': ['conv'], 'name': 'inception_3b_output', 'depth': 480},
{'tags': ['conv'], 'name': 'inception_4a_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4b_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4c_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4d_output', 'depth': 528},
{'tags': ['conv'], 'name': 'inception_4e_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5a_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5b_output', 'depth': 1024},
{'tags': ['dense'], 'name': 'prob', 'depth': 1000},
])
class ResnetV2_101_slim(Model):
"""ResnetV2_101 as implemented by the TensorFlow slim framework.
ResnetV2_101 was introduced by He, et al (2016):
https://arxiv.org/pdf/1603.05027.pdf
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
corresponding to the name "resnet_v2_101".
"""
model_path = 'gs://modelzoo/vision/slim_models/ResnetV2_101.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{'tags': ['conv'], 'name': 'resnet_v2_101/block1/unit_1/bottleneck_v2/preact/Relu', 'depth': 64},
{'tags': ['conv'], 'name': 'resnet_v2_101/block1/unit_1/bottleneck_v2/add', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v2_101/block1/unit_2/bottleneck_v2/add', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v2_101/block1/unit_3/bottleneck_v2/add', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v2_101/block2/unit_1/bottleneck_v2/preact/Relu', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v2_101/block2/unit_1/bottleneck_v2/add', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v2_101/block2/unit_2/bottleneck_v2/add', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v2_101/block2/unit_3/bottleneck_v2/add', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v2_101/block2/unit_4/bottleneck_v2/add', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_1/bottleneck_v2/preact/Relu', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_1/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_2/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_3/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_4/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_5/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_6/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_7/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_8/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_9/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_10/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_11/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_12/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_13/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_14/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_15/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_16/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_17/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_18/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_19/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_20/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_21/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_22/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block3/unit_23/bottleneck_v2/add', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block4/unit_1/bottleneck_v2/preact/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v2_101/block4/unit_1/bottleneck_v2/add', 'depth': 2048},
{'tags': ['conv'], 'name': 'resnet_v2_101/block4/unit_2/bottleneck_v2/add', 'depth': 2048},
{'tags': ['conv'], 'name': 'resnet_v2_101/block4/unit_3/bottleneck_v2/add', 'depth': 2048},
{'tags': ['conv'], 'name': 'resnet_v2_101/postnorm/Relu', 'depth': 2048},
{'tags': ['dense'], 'name': 'resnet_v2_101/predictions/Softmax', 'depth': 1001},
])
class InceptionV1_caffe_Places365(Model):
"""InceptionV1 (or 'GoogLeNet') trained on Places365.
This model is a caffe reimplementation of GoogLeNet:
https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
trained on the MIT Places365 dataset, retrieved here:
https://github.com/CSAILVision/places365
and then ported to TensorFlow using caffe-tensorflow.
"""
model_path = 'gs://modelzoo/vision/caffe_models/InceptionV1_places365.pb'
labels_path = 'gs://modelzoo/labels/Places365.txt'
dataset = 'Places365'
image_shape = [224, 224, 3]
# What is the correct input range???
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255-IMAGENET_MEAN_BGR)
input_name = 'data'
layers = _layers_from_list_of_dicts([
{'tags': ['conv'], 'name': 'conv1_7x7_s2/conv1_7x7_s2', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3_reduce/conv2_3x3_reduce', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3/conv2_3x3', 'depth': 192},
{'tags': ['conv'], 'name': 'inception_3a_output', 'depth': 256},
{'tags': ['conv'], 'name': 'inception_3b_output', 'depth': 480},
{'tags': ['conv'], 'name': 'inception_4a_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4b_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4c_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4d_output', 'depth': 528},
{'tags': ['conv'], 'name': 'inception_4e_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5a_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5b_output', 'depth': 1024},
{'tags': ['dense'], 'name': 'prob', 'depth': 365},
])
class ResnetV1_101_slim(Model):
"""ResnetV1101 as implemented by the TensorFlow slim framework.
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
"""
model_path = 'gs://modelzoo/vision/slim_models/ResnetV1_101.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
image_value_range = (-117, 255-117) # Inferred by testing, may not be exactly right
input_name = 'input'
# In ResNetV1, each add (joining the residual branch) is followed by a Relu
# this seems to be the natural "layer" position
layers = _layers_from_list_of_dicts([
{'tags': ['conv'], 'name': 'resnet_v1_101/conv1/Relu', 'depth': 64},
{'tags': ['conv'], 'name': 'resnet_v1_101/block1/unit_1/bottleneck_v1/Relu', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v1_101/block1/unit_2/bottleneck_v1/Relu', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v1_101/block1/unit_3/bottleneck_v1/Relu', 'depth': 256},
{'tags': ['conv'], 'name': 'resnet_v1_101/block2/unit_1/bottleneck_v1/Relu', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v1_101/block2/unit_2/bottleneck_v1/Relu', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v1_101/block2/unit_3/bottleneck_v1/Relu', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v1_101/block2/unit_4/bottleneck_v1/Relu', 'depth': 512},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_1/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_2/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_3/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_4/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_5/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_6/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_7/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_8/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_9/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_10/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_11/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_12/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_13/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_14/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_15/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_16/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_17/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_18/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_19/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_20/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_21/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_22/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block3/unit_23/bottleneck_v1/Relu', 'depth': 1024},
{'tags': ['conv'], 'name': 'resnet_v1_101/block4/unit_1/bottleneck_v1/Relu', 'depth': 2048},
{'tags': ['conv'], 'name': 'resnet_v1_101/block4/unit_2/bottleneck_v1/Relu', 'depth': 2048},
{'tags': ['conv'], 'name': 'resnet_v1_101/block4/unit_3/bottleneck_v1/Relu', 'depth': 2048},
{'tags': ['dense'], 'name': 'resnet_v1_101/predictions/Softmax', 'depth': 1000},
])
class AlexNet_caffe_Places365(Model):
"""AlexNet re-implementation trained on Places365.
This model is a reimplementation of AlexNet
https://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks.pdf
trained on the MIT Places365 dataset, retrieved here:
https://github.com/CSAILVision/places365
and then ported to TensorFlow using caffe-tensorflow.
"""
model_path = 'gs://modelzoo/vision/caffe_models/AlexNet_places365.pb'
labels_path = 'gs://modelzoo/labels/Places365.txt'
dataset = 'Places365'
image_shape = [227, 227, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255 - IMAGENET_MEAN_BGR)
input_name = 'input'
# TODO - Sanity check this graph and layers
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'conv5/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv5/conv5',
'depth': 256
},
{
'tags': ['dense'],
'name': 'fc6/fc6',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'fc7/fc7',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'prob',
'depth': 365
},
])
class CaffeNet_caffe(Model):
"""CaffeNet (AlexNet variant included in Caffe)
CaffeNet is a slight variant on AlexNet, described here:
https://github.com/BVLC/caffe/tree/master/models/bvlc_reference_caffenet
"""
model_path = 'gs://modelzoo/vision/caffe_models/CaffeNet.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [227, 227, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255 - IMAGENET_MEAN_BGR)
input_name = 'data'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'conv5/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv5/conv5',
'depth': 256
},
{
'tags': ['dense'],
'name': 'fc6/fc6',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'fc7/fc7',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'prob',
'depth': 1000
},
])
ResnetV2_50_slim.layers = _layers_from_list_of_dicts(ResnetV2_50_slim(), [
{
'tags': ['conv'],
'name': 'resnet_v2_50/block1/unit_1/bottleneck_v2/preact/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block1/unit_1/bottleneck_v2/add',
'depth': 256
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block1/unit_2/bottleneck_v2/add',
'depth': 256
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block1/unit_3/bottleneck_v2/add',
'depth': 256
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block2/unit_1/bottleneck_v2/preact/Relu',
'depth': 256
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block2/unit_1/bottleneck_v2/add',
'depth': 512
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block2/unit_2/bottleneck_v2/add',
'depth': 512
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block2/unit_3/bottleneck_v2/add',
'depth': 512
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block2/unit_4/bottleneck_v2/add',
'depth': 512
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_1/bottleneck_v2/preact/Relu',
'depth': 512
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_1/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_2/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_3/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_4/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_5/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block3/unit_6/bottleneck_v2/add',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block4/unit_1/bottleneck_v2/preact/Relu',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block4/unit_1/bottleneck_v2/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block4/unit_2/bottleneck_v2/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/block4/unit_3/bottleneck_v2/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'resnet_v2_50/postnorm/Relu',
'depth': 2048
},
{
'tags': ['dense'],
'name': 'resnet_v2_50/predictions/Softmax',
'depth': 1001
},
])
CovidNetB.layers = _layers_from_list_of_dicts(CovidNetB(), [
{
'tags': ['conv'],
'name': 'conv5_block3_out/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'conv5_block2_out/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'conv5_block1_out/add',
'depth': 2048
},
{
'tags': ['conv'],
'name': 'conv4_block6_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv4_block5_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv4_block4_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv4_block3_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv4_block2_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv4_block1_out/add',
'depth': 824
},
{
'tags': ['conv'],
'name': 'conv3_block4_out/add',
'depth': 416
},
{
'tags': ['conv'],
'name': 'conv3_block3_out/add',
'depth': 416
},
{
'tags': ['conv'],
'name': 'conv3_block2_out/add',
'depth': 416
},
{
'tags': ['conv'],
'name': 'conv3_block1_out/add',
'depth': 416
},
{
'tags': ['conv'],
'name': 'conv2_block3_out/add',
'depth': 200
},
{
'tags': ['conv'],
'name': 'conv2_block2_out/add',
'depth': 200
},
{
'tags': ['conv'],
'name': 'conv2_block1_out/add',
'depth': 200
},
{
'tags': ['dense'],
'name': 'norm_dense_1/Softmax',
'depth': 3
},
])
InceptionV1_slim.layers = _layers_from_list_of_dicts(InceptionV1_slim, [
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_1a_7x7/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_2b_1x1/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_2c_3x3/Relu',
'depth': 192
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_3b/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_3c/concat',
'depth': 480
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4b/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4c/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4d/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4e/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4f/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_5b/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_5c/concat',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'InceptionV1/Logits/Predictions/Softmax',
'depth': 1001
},
])
class InceptionV1_slim(Model):
"""InceptionV1 as implemented by the TensorFlow slim framework.
InceptionV1 was introduced by Szegedy, et al (2014):
https://arxiv.org/pdf/1409.4842v1.pdf
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
corresponding to the name "inception_v1".
"""
model_path = 'gs://modelzoo/vision/slim_models/InceptionV1.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_1a_7x7/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_2b_1x1/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Conv2d_2c_3x3/Relu',
'depth': 192
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_3b/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_3c/concat',
'depth': 480
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4b/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4c/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4d/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4e/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_4f/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_5b/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'InceptionV1/InceptionV1/Mixed_5c/concat',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'InceptionV1/Logits/Predictions/Softmax',
'depth': 1001
},
])
class MobilenetV1_025_slim(Model):
"""MobilenetV1025 as implemented by the TensorFlow slim framework.
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
"""
model_path = 'gs://modelzoo/vision/slim_models/MobilenetV1025.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt' #TODO
dataset = 'ImageNet'
image_shape = [224, 224, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_0/Relu6',
'depth': 8
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_1_pointwise/Relu6',
'depth': 16
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_2_pointwise/Relu6',
'depth': 32
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_3_pointwise/Relu6',
'depth': 32
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_4_pointwise/Relu6',
'depth': 64
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_5_pointwise/Relu6',
'depth': 64
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_6_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_7_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_8_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_9_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_10_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_11_pointwise/Relu6',
'depth': 128
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_12_pointwise/Relu6',
'depth': 256
},
{
'tags': ['conv'],
'name': 'MobilenetV1/MobilenetV1/Conv2d_13_pointwise/Relu6',
'depth': 256
},
# {'tags': 'avgpool', 'name': 'MobilenetV1/Logits/AvgPool_1a/AvgPool', 'depth': 256},
# {'tags': ['conv'], 'name': 'MobilenetV1/Logits/Conv2d_1c_1x1/Conv2D', 'depth': 1001},
{
'tags': ['dense'],
'name': 'MobilenetV1/Predictions/Softmax',
'depth': 1001
},
])
class InceptionV1(Model):
"""InceptionV1 (or 'GoogLeNet')
This is a (re?)implementation of InceptionV1
https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
The weights were trained at Google and released in an early TensorFlow
tutorial. It is possible the parameters are the original weights
(trained in TensorFlow's predecessor), but we haven't been able to
confirm this.
As far as we can tell, it is exactly the same as the model described in
the original paper, where as the slim and caffe implementations have
minor implementation differences (such as eliding the heads).
"""
model_path = 'gs://modelzoo/vision/other_models/InceptionV1.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_alternate.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
image_value_range = (-117, 255 - 117)
input_name = 'input:0'
def post_import(self, scope):
_populate_inception_bottlenecks(scope)
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'conv2d0',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv2d1',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv2d2',
'depth': 192
},
{
'tags': ['conv'],
'name': 'mixed3a',
'depth': 256
},
{
'tags': ['conv'],
'name': 'mixed3b',
'depth': 480
},
{
'tags': ['conv'],
'name': 'mixed4a',
'depth': 508
},
{
'tags': ['conv'],
'name': 'mixed4b',
'depth': 512
},
{
'tags': ['conv'],
'name': 'mixed4c',
'depth': 512
},
{
'tags': ['conv'],
'name': 'mixed4d',
'depth': 528
},
{
'tags': ['conv'],
'name': 'mixed4e',
'depth': 832
},
{
'tags': ['conv'],
'name': 'mixed5a',
'depth': 832
},
{
'tags': ['conv'],
'name': 'mixed5b',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'head0_bottleneck',
'depth': 128
},
{
'tags': ['dense'],
'name': 'nn0',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'softmax0',
'depth': 1008
},
{
'tags': ['conv'],
'name': 'head1_bottleneck',
'depth': 128
},
{
'tags': ['dense'],
'name': 'nn1',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'softmax1',
'depth': 1008
},
{
'tags': ['dense'],
'name': 'softmax2',
'depth': 1008
},
])
class VGG16_caffe(Model):
"""VGG16 model used in ImageNet ILSVRC-2014, ported from caffe.
VGG16 was introduced by Simonyan & Zisserman (2014):
https://arxiv.org/pdf/1409.1556.pdf
http://www.robots.ox.ac.uk/~vgg/research/very_deep/
as the Oxford Visual Geometry Group's submission for the ImageNet ILSVRC-2014
contest. We download their caffe trained model from
https://gist.github.com/ksimonyan/211839e770f7b538e2d8#file-readme-md
and convert it with caffe-tensorflow.
"""
model_path = 'gs://modelzoo/vision/caffe_models/VGG16.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255 - IMAGENET_MEAN_BGR)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'conv1_1/conv1_1',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv1_2/conv1_2',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv2_1/conv2_1',
'depth': 128
},
{
'tags': ['conv'],
'name': 'conv2_2/conv2_2',
'depth': 128
},
{
'tags': ['conv'],
'name': 'conv3_1/conv3_1',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv3_2/conv3_2',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv3_3/conv3_3',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv4_1/conv4_1',
'depth': 512
},
{
'tags': ['conv'],
'name': 'conv4_2/conv4_2',
'depth': 512
},
{
'tags': ['conv'],
'name': 'conv4_3/conv4_3',
'depth': 512
},
{
'tags': ['conv'],
'name': 'conv5_1/conv5_1',
'depth': 512
},
{
'tags': ['conv'],
'name': 'conv5_2/conv5_2',
'depth': 512
},
{
'tags': ['conv'],
'name': 'conv5_3/conv5_3',
'depth': 512
},
{
'tags': ['dense'],
'name': 'fc6/fc6',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'fc7/fc7',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'prob',
'depth': 1000
},
])
class MobilenetV2_14_slim(Model):
"""MobilenetV2 1.4 as implemented by the TensorFlow slim framework.
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
"""
model_path = 'gs://modelzoo/vision/slim_models/MobilenetV2_14.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_2/add',
'depth': 32
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_4/add',
'depth': 48
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_5/add',
'depth': 48
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_7/add',
'depth': 88
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_8/add',
'depth': 88
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_9/add',
'depth': 88
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_11/add',
'depth': 136
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_12/add',
'depth': 136
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_14/add',
'depth': 224
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_15/add',
'depth': 224
},
{
'tags': ['dense'],
'name': 'MobilenetV2/Predictions/Softmax',
'depth': 1001
},
])
MobilenetV2_10_slim.layers = _layers_from_list_of_dicts(
MobilenetV2_10_slim(), [
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_2/add',
'depth': 24
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_4/add',
'depth': 32
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_5/add',
'depth': 32
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_7/add',
'depth': 64
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_8/add',
'depth': 64
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_9/add',
'depth': 64
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_11/add',
'depth': 96
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_12/add',
'depth': 96
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_14/add',
'depth': 160
},
{
'tags': ['conv'],
'name': 'MobilenetV2/expanded_conv_15/add',
'depth': 160
},
{
'tags': ['dense'],
'name': 'MobilenetV2/Predictions/Softmax',
'depth': 1001
},
])
image_shape = [224, 224, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
MobilenetV1_slim.layers = _layers_from_list_of_dicts(MobilenetV1_slim, [
{'name': 'MobilenetV1/MobilenetV1/Conv2d_0/Relu6', 'depth': 32, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_1_pointwise/Relu6', 'depth': 64, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_2_pointwise/Relu6', 'depth': 128, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_3_pointwise/Relu6', 'depth': 128, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_4_pointwise/Relu6', 'depth': 256, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_5_pointwise/Relu6', 'depth': 256, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_6_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_7_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_8_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_9_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_10_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_11_pointwise/Relu6', 'depth': 512, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_12_pointwise/Relu6', 'depth': 1024, 'tags': ['conv']},
{'name': 'MobilenetV1/MobilenetV1/Conv2d_13_pointwise/Relu6', 'depth': 1024, 'tags': ['conv']},
# {'name': 'MobilenetV1/Logits/AvgPool_1a/AvgPool', 'depth': 1024, 'type': 'avgpool'},
# {'name': 'MobilenetV1/Logits/Conv2d_1c_1x1/Conv2D', 'depth': 1001, 'tags': ['conv']},
{'name': 'MobilenetV1/Predictions/Softmax', 'depth': 1001, 'tags': ['dense']},
])
class MobilenetV1_050_slim(Model):
"""MobilenetV1050 as implemented by the TensorFlow slim framework.
This function provides the pre-trained reimplementation from TF slim:
class InceptionV4_slim(Model):
"""InceptionV4 as implemented by the TensorFlow slim framework.
InceptionV4 was introduced by Szegedy, et al (2016):
https://arxiv.org/pdf/1602.07261.pdf
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
corresponding to the name "inception_v4".
"""
model_path = 'gs://modelzoo/vision/slim_models/InceptionV4.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt'
dataset = 'ImageNet'
image_shape = [299, 299, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Conv2d_1a_3x3/Relu',
'depth': 32
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Conv2d_2a_3x3/Relu',
'depth': 32
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Conv2d_2b_3x3/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_3a/concat',
'depth': 160
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_4a/concat',
'depth': 192
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_5a/concat',
'depth': 384
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_5b/concat',
'depth': 384
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_5c/concat',
'depth': 384
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_5d/concat',
'depth': 384
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_5e/concat',
'depth': 384
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6a/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6b/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6c/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6d/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6e/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6f/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6g/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_6h/concat',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_7a/concat',
'depth': 1536
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_7b/concat',
'depth': 1536
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_7c/concat',
'depth': 1536
},
{
'tags': ['conv'],
'name': 'InceptionV4/InceptionV4/Mixed_7d/concat',
'depth': 1536
},
{
'tags': ['dense'],
'name': 'InceptionV4/Logits/Predictions',
'depth': 1001
},
])
# TODO - Sanity check this graph and layers
AlexNet.layers = _layers_from_list_of_dicts(AlexNet, [
{
'tags': ['conv'],
'name': 'concat_2',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv5_1',
'depth': 256
},
{
'tags': ['dense'],
'name': 'Relu',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'Relu_1',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'Softmax',
'depth': 1000
},
])
I3D.layers = _layers_from_list_of_dicts(
I3D,
[
{
'tags': ['conv'],
'name': 'inceptioni3d/Conv3d_1a_7x7/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Conv3d_2b_1x1/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Conv3d_2c_3x3/Relu',
'depth': 192
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_3b/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_3c/concat',
'depth': 480
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_4b/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_4c/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_4d/concat',
'depth': 512
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_4e/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_4f/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_5b/concat',
'depth': 832
},
{
'tags': ['conv'],
'name': 'inceptioni3d/Mixed_5c/concat',
'depth': 1024
},
{
'tags': ['conv'],
# 'name' : 'inceptioni3d/Logits/Conv3d_0c_1x1/conv_3d/add', # 'inceptioni3d/Logits/Conv3d_0c_1x1',
'name': 'inceptioni3d/Logits/SpatialSqueeze',
'depth': 400
}
])
model_path = 'gs://modelzoo/vision/caffe_models/InceptionV1.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard.txt'
dataset = 'ImageNet'
image_shape = [224, 224, 3]
is_BGR = True
image_value_range = (-IMAGENET_MEAN_BGR, 255-IMAGENET_MEAN_BGR)
input_name = 'data'
InceptionV1_caffe.layers = _layers_from_list_of_dicts(InceptionV1_caffe, [
{'tags': ['conv'], 'name': 'conv1_7x7_s2/conv1_7x7_s2', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3_reduce/conv2_3x3_reduce', 'depth': 64},
{'tags': ['conv'], 'name': 'conv2_3x3/conv2_3x3', 'depth': 192},
{'tags': ['conv'], 'name': 'inception_3a_output', 'depth': 256},
{'tags': ['conv'], 'name': 'inception_3b_output', 'depth': 480},
{'tags': ['conv'], 'name': 'inception_4a_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4b_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4c_output', 'depth': 512},
{'tags': ['conv'], 'name': 'inception_4d_output', 'depth': 528},
{'tags': ['conv'], 'name': 'inception_4e_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5a_output', 'depth': 832},
{'tags': ['conv'], 'name': 'inception_5b_output', 'depth': 1024},
{'tags': ['dense'], 'name': 'prob', 'depth': 1000},
])
class InceptionV1_caffe_Places205(Model):
"""InceptionV1 (or 'GoogLeNet') trained on Places205.
This model is a caffe reimplementation of GoogLeNet:
https://www.cs.unc.edu/~wliu/papers/GoogLeNet.pdf
trained on the MIT Places205 dataset, retrieved here:
AlexNet.layers = _layers_from_list_of_dicts(AlexNet, [
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D',
'depth': 96
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_1',
'depth': 128
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_2',
'depth': 128
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_3',
'depth': 384
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_4',
'depth': 192
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_5',
'depth': 192
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_6',
'depth': 128
},
{
'tags': ['pre_relu', 'conv'],
'name': 'Conv2D_7',
'depth': 128
},
{
'tags': ['dense'],
'name': 'Relu',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'Relu_1',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'Softmax',
'depth': 1000
},
])
InceptionV1.layers = _layers_from_list_of_dicts(InceptionV1, [
{
'tags': ['conv'],
'name': 'conv2d0',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv2d1',
'depth': 64
},
{
'tags': ['conv'],
'name': 'conv2d2',
'depth': 192
},
{
'tags': ['conv'],
'name': 'mixed3a',
'depth': 256
},
{
'tags': ['conv'],
'name': 'mixed3b',
'depth': 480
},
{
'tags': ['conv'],
'name': 'mixed4a',
'depth': 508
},
{
'tags': ['conv'],
'name': 'mixed4b',
'depth': 512
},
{
'tags': ['conv'],
'name': 'mixed4c',
'depth': 512
},
{
'tags': ['conv'],
'name': 'mixed4d',
'depth': 528
},
{
'tags': ['conv'],
'name': 'mixed4e',
'depth': 832
},
{
'tags': ['conv'],
'name': 'mixed5a',
'depth': 832
},
{
'tags': ['conv'],
'name': 'mixed5b',
'depth': 1024
},
{
'tags': ['conv'],
'name': 'head0_bottleneck',
'depth': 128
},
{
'tags': ['dense'],
'name': 'nn0',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'softmax0',
'depth': 1008
},
{
'tags': ['conv'],
'name': 'head1_bottleneck',
'depth': 128
},
{
'tags': ['dense'],
'name': 'nn1',
'depth': 1024
},
{
'tags': ['dense'],
'name': 'softmax1',
'depth': 1008
},
{
'tags': ['dense'],
'name': 'softmax2',
'depth': 1008
},
])
CaffeNet_caffe.layers = _layers_from_list_of_dicts(CaffeNet_caffe, [
{
'tags': ['conv'],
'name': 'conv5/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv5/conv5',
'depth': 256
},
{
'tags': ['dense'],
'name': 'fc6/fc6',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'fc7/fc7',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'prob',
'depth': 1000
},
])
# TODO - Sanity check this graph and layers
AlexNet_caffe_Places365.layers = _layers_from_list_of_dicts(
AlexNet_caffe_Places365(), [
{
'tags': ['conv'],
'name': 'conv5/concat',
'depth': 256
},
{
'tags': ['conv'],
'name': 'conv5/conv5',
'depth': 256
},
{
'tags': ['dense'],
'name': 'fc6/fc6',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'fc7/fc7',
'depth': 4096
},
{
'tags': ['dense'],
'name': 'prob',
'depth': 365
},
])
NasnetMobile_slim.layers = _layers_from_list_of_dicts(
NasnetMobile_slim(),
[
{
'tags': ['conv'],
'name': 'conv0/Conv2D',
'depth': 32
},
{
'tags': ['conv'],
'name': 'cell_stem_0/cell_output/concat',
'depth': 44
},
{
'tags': ['conv'],
'name': 'cell_stem_1/cell_output/concat',
'depth': 88
},
{
'tags': ['conv'],
'name': 'cell_0/cell_output/concat',
'depth': 264
},
{
'tags': ['conv'],
'name': 'cell_1/cell_output/concat',
'depth': 264
},
{
'tags': ['conv'],
'name': 'cell_2/cell_output/concat',
'depth': 264
},
{
'tags': ['conv'],
'name': 'cell_3/cell_output/concat',
'depth': 264
},
{
'tags': ['conv'],
'name': 'reduction_cell_0/cell_output/concat',
'depth': 352
},
{
'tags': ['conv'],
'name': 'cell_4/cell_output/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'cell_5/cell_output/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'cell_6/cell_output/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'cell_7/cell_output/concat',
'depth': 528
},
{
'tags': ['conv'],
'name': 'reduction_cell_1/cell_output/concat',
'depth': 704
},
{
'tags': ['conv'],
'name': 'cell_8/cell_output/concat',
'depth': 1056
},
{
'tags': ['conv'],
'name': 'cell_9/cell_output/concat',
'depth': 1056
},
{
'tags': ['conv'],
'name': 'cell_10/cell_output/concat',
'depth': 1056
},
{
'tags': ['conv'],
'name': 'cell_11/cell_output/concat',
'depth': 1056
},
# {'tags': ['dense'], 'name': 'final_layer/FC/BiasAdd', 'depth': 1001},
{
'tags': ['dense'],
'name': 'final_layer/predictions',
'depth': 1001
},
])
class PnasnetLarge_slim(Model):
"""PnasnetLarge as implemented by the TensorFlow slim framework.
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
"""
model_path = 'gs://modelzoo/vision/slim_models/PnasnetLarge.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt' #TODO
dataset = 'ImageNet'
image_shape = [331, 331, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
layers = _layers_from_list_of_dicts([{
'tags': ['conv'],
'name': 'conv0/Conv2D',
'depth': 96
}, {
'tags': ['conv'],
'name': 'cell_stem_0/cell_output/concat',
'depth': 270
}, {
'tags': ['conv'],
'name': 'cell_stem_1/cell_output/concat',
'depth': 540
}, {
'tags': ['conv'],
'name': 'cell_0/cell_output/concat',
'depth': 1080
}, {
'tags': ['conv'],
'name': 'cell_1/cell_output/concat',
'depth': 1080
}, {
'tags': ['conv'],
'name': 'cell_2/cell_output/concat',
'depth': 1080
}, {
'tags': ['conv'],
'name': 'cell_3/cell_output/concat',
'depth': 1080
}, {
'tags': ['conv'],
'name': 'cell_4/cell_output/concat',
'depth': 2160
}, {
'tags': ['conv'],
'name': 'cell_5/cell_output/concat',
'depth': 2160
}, {
'tags': ['conv'],
'name': 'cell_6/cell_output/concat',
'depth': 2160
}, {
'tags': ['conv'],
'name': 'cell_7/cell_output/concat',
'depth': 2160
}, {
'tags': ['conv'],
'name': 'cell_8/cell_output/concat',
'depth': 4320
}, {
'tags': ['conv'],
'name': 'cell_9/cell_output/concat',
'depth': 4320
}, {
'tags': ['conv'],
'name': 'cell_10/cell_output/concat',
'depth': 4320
}, {
'tags': ['conv'],
'name': 'cell_11/cell_output/concat',
'depth': 4320
}, {
'tags': ['dense'],
'name': 'final_layer/predictions',
'depth': 1001
}])
class InceptionResnetV2_slim(Model):
"""InceptionResnetV2 as implemented by the TensorFlow slim framework.
InceptionResnetV2 was introduced in this paper by Szegedy, et al (2016):
https://arxiv.org/pdf/1602.07261.pdf
This function provides the pre-trained reimplementation from TF slim:
https://github.com/tensorflow/models/tree/master/research/slim
corresponding to the name "inception_resnet_v2".
"""
model_path = 'gs://modelzoo/vision/slim_models/InceptionResnetV2.pb'
labels_path = 'gs://modelzoo/labels/ImageNet_standard_with_dummy.txt'
dataset = 'ImageNet'
image_shape = [299, 299, 3]
# inpute range taken from:
# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/inception_preprocessing.py#L280
image_value_range = (-1, 1)
input_name = 'input'
# TODO: understand this graph, see if we can delete some add or relu nodes from layers
layers = _layers_from_list_of_dicts([
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_1a_3x3/Relu',
'depth': 32
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_2a_3x3/Relu',
'depth': 32
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_2b_3x3/Relu',
'depth': 64
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_3b_1x1/Relu',
'depth': 80
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_4a_3x3/Relu',
'depth': 192
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Mixed_5b/concat',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_1/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_1/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_2/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_2/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_3/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_3/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_4/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_4/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_5/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_5/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_6/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_6/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_7/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_7/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_8/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_8/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Repeat/block35_9/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_9/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_10/add',
'depth': 320
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat/block35_10/Relu',
'depth': 320
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Mixed_6a/concat',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_1/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_1/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_2/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_2/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_3/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_3/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_4/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_4/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_5/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_5/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_6/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_6/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_7/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_7/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_8/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_8/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_9/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_9/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_10/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_10/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_11/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_11/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_12/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_12/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_13/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_13/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_14/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_14/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_15/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_15/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_16/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_16/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_17/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_17/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_18/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_18/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_19/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_19/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_20/add',
'depth': 1088
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_1/block17_20/Relu',
'depth': 1088
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Mixed_7a/concat',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_1/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_1/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_2/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_2/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_3/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_3/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_4/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_4/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_5/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_5/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_6/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_6/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_7/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_7/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_8/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_8/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_9/add',
'depth': 2080
},
{
'tags': ['conv'],
'name':
'InceptionResnetV2/InceptionResnetV2/Repeat_2/block8_9/Relu',
'depth': 2080
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Block8/add',
'depth': 2080
},
{
'tags': ['conv'],
'name': 'InceptionResnetV2/InceptionResnetV2/Conv2d_7b_1x1/Relu',
'depth': 1536
},
{
'tags': ['dense'],
'name': 'InceptionResnetV2/Logits/Predictions',
'depth': 1001
},
])
