TensorNets

High level network definitions with pre-trained weights in TensorFlow (tested with >= 1.2.0).

Guiding principles

• Applicability. Many people already have their own ML workflows, and want to put a new model on their workflows. TensorNets can be easily plugged together because it is designed as simple functional interfaces without custom classes.
• Manageability. Models are written in tf.contrib.layers, which is lightweight like PyTorch and Keras, and allows for ease of accessibility to every weight and end-point. Also, it is easy to deploy and expand a collection of pre-processing and pre-trained weights.
• Readability. With recent TensorFlow APIs, more factoring and less indenting can be possible. For example, all the inception variants are implemented as about 500 lines of code in TensorNets while 2000+ lines in official TensorFlow models.

A quick example

Each network (see full list) is not a custom class but a function that takes and returns tf.Tensor as its input and output. Here is an example of ResNet50:

import tensorflow as tf
import tensornets as nets

inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
model = nets.ResNet50(inputs)

assert isinstance(model, tf.Tensor)

You can load an example image by using utils.load_img returning a np.ndarray as the NHWC format:

from tensornets import utils
img = utils.load_img('cat.png', target_size=256, crop_size=224)
assert img.shape == (1, 224, 224, 3)

Once your network is created, you can run with regular TensorFlow APIs 😊 because all the networks in TensorNets always return tf.Tensor. Using pre-trained weights and pre-processing are as easy as pretrained() and preprocess() to reproduce the original results:

with tf.Session() as sess:
    img = model.preprocess(img)  # equivalent to img = nets.preprocess(model, img)
    sess.run(model.pretrained())  # equivalent to nets.pretrained(model)
    preds = sess.run(model, {inputs: img})

You can see the most probable classes:

print(utils.decode_predictions(preds, top=2)[0])
[(u'n02124075', u'Egyptian_cat', 0.28067636), (u'n02127052', u'lynx', 0.16826575)]

TensorNets enables us to deploy well-known architectures and benchmark those results faster ⚡️. For more information, you can check out the lists of utilities, examples, and architectures.

Object detection example

Each object detection model can be coupled with any network in TensorNets (see performances) and takes two arguments: a placeholder and a function acting as a stem layer. Here is an example of YOLOv2 for PASCAL VOC:

import tensorflow as tf
import tensornets as nets

inputs = tf.placeholder(tf.float32, [None, 416, 416, 3])
model = nets.YOLOv2(inputs, nets.Darknet19)

img = nets.utils.load_img('cat.png')

with tf.Session() as sess:
    sess.run(model.pretrained())
    preds = sess.run(model, {inputs: model.preprocess(img)})
    boxes = model.get_boxes(preds, img.shape[1:3])

Like other models, a detection model also returns tf.Tensor as its output. You can see the bounding box predictions (x1, y1, x2, y2, score) by using model.get_boxes(model_output, original_img_shape) and visualize the results:

from tensornets.datasets import voc
print("%s: %s" % (voc.classnames[7], boxes[7][0]))  # 7 is cat

import numpy as np
import matplotlib.pyplot as plt
box = boxes[7][0]
plt.imshow(img[0].astype(np.uint8))
plt.gca().add_patch(plt.Rectangle(
    (box[0], box[1]), box[2] - box[0], box[3] - box[1],
    fill=False, edgecolor='r', linewidth=2))
plt.show()

More detection examples such as FasterRCNN on VOC2007 are here 😎.

Utilities

An example output of utils.print_summary(model):

Scope: resnet50
Total layers: 54
Total weights: 320
Total parameters: 25,636,712

An example output of utils.print_weights(model):

Scope: resnet50
conv1/conv/weights:0 (7, 7, 3, 64)
conv1/conv/biases:0 (64,)
conv1/bn/beta:0 (64,)
conv1/bn/gamma:0 (64,)
conv1/bn/moving_mean:0 (64,)
conv1/bn/moving_variance:0 (64,)
conv2/block1/0/conv/weights:0 (1, 1, 64, 256)
conv2/block1/0/conv/biases:0 (256,)
conv2/block1/0/bn/beta:0 (256,)
conv2/block1/0/bn/gamma:0 (256,)
...

• utils.get_weights(model) returns a list of all the tf.Tensor weights as shown in the above

An example output of utils.print_outputs(model):

Scope: resnet50
conv1/pad:0 (?, 230, 230, 3)
conv1/conv/BiasAdd:0 (?, 112, 112, 64)
conv1/bn/batchnorm/add_1:0 (?, 112, 112, 64)
conv1/relu:0 (?, 112, 112, 64)
pool1/pad:0 (?, 114, 114, 64)
pool1/MaxPool:0 (?, 56, 56, 64)
conv2/block1/0/conv/BiasAdd:0 (?, 56, 56, 256)
conv2/block1/0/bn/batchnorm/add_1:0 (?, 56, 56, 256)
conv2/block1/1/conv/BiasAdd:0 (?, 56, 56, 64)
conv2/block1/1/bn/batchnorm/add_1:0 (?, 56, 56, 64)
conv2/block1/1/relu:0 (?, 56, 56, 64)
...

• utils.get_outputs(model) returns a list of all the tf.Tensor end-points as shown in the above

Examples

• Comparison of different networks:

inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
models = [
    nets.MobileNet75(inputs),
    nets.MobileNet100(inputs),
    nets.SqueezeNet(inputs),
]

img = utils.load_img('cat.png', target_size=256, crop_size=224)
imgs = nets.preprocess(models, img)

with tf.Session() as sess:
    nets.pretrained(models)
    for (model, img) in zip(models, imgs):
        preds = sess.run(model, {inputs: img})
        print(utils.decode_predictions(preds, top=2)[0])

• Transfer learning:

inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
outputs = tf.placeholder(tf.float32, [None, 50])
model = nets.DenseNet169(inputs, is_training=True, classes=50)

loss = tf.losses.softmax_cross_entropy(outputs, model)
train = tf.train.AdamOptimizer(learning_rate=1e-5).minimize(loss)

with tf.Session() as sess:
    nets.pretrained(model)
    # for (x, y) in your NumPy data (the NHWC and one-hot format):
        sess.run(train, {inputs: x, outputs: y})

• Using multi-GPU:

inputs = tf.placeholder(tf.float32, [None, 224, 224, 3])
models = []

with tf.device('gpu:0'):
    models.append(nets.ResNeXt50(inputs))

with tf.device('gpu:1'):
    models.append(nets.DenseNet201(inputs))

from tensornets.preprocess import fb_preprocess
img = utils.load_img('cat.png', target_size=256, crop_size=224)
img = fb_preprocess(img)

with tf.Session() as sess:
    nets.pretrained(models)
    preds = sess.run(models, {inputs: img})
    for pred in preds:
        print(utils.decode_predictions(pred, top=2)[0])

Performances

Image classification

• The top-k errors were obtained with TensorNets on ImageNet validation set and may slightly differ from the original ones. The crop size is 224x224 for all but 331x331 for NASNetAlarge, 299x299 for Inception3,4,ResNet2, and ResNet50-152v2.
  • Top-1: single center crop, top-1 error
  • Top-5: single center crop, top-5 error
  • 10-5: ten crops (1 center + 4 corners and those mirrored ones), top-5 error
  • Size: rounded the number of parameters
• The computation times were measured on NVIDIA Tesla P100 (3584 cores, 16 GB global memory) with cuDNN 6.0 and CUDA 8.0.
  • Speed: milliseconds for inferences of 100 images

	Top-1	Top-5	10-5	Size	Speed	References
ResNet50	25.126	7.982	6.842	26M	195.4	[paper] [tf-slim] [torch-fb] [caffe] [keras]
ResNet101	23.580	7.214	6.092	45M	311.7	[paper] [tf-slim] [torch-fb] [caffe]
ResNet152	23.396	6.882	5.908	60M	439.1	[paper] [tf-slim] [torch-fb] [caffe]
ResNet50v2	24.526	7.252	6.012	26M	209.7	[paper] [tf-slim] [torch-fb]
ResNet101v2	23.116	6.488	5.230	45M	326.2	[paper] [tf-slim] [torch-fb]
ResNet152v2	22.236	6.080	4.960	60M	455.2	[paper] [tf-slim] [torch-fb]
ResNet200v2	21.714	5.848	4.830	65M	618.3	[paper] [tf-slim] [torch-fb]
ResNeXt50c32	22.260	6.190	5.410	25M	267.4	[paper] [torch-fb]
ResNeXt101c32	21.270	5.706	4.842	44M	427.9	[paper] [torch-fb]
ResNeXt101c64	20.506	5.408	4.564	84M	877.8	[paper] [torch-fb]
WideResNet50	21.982	6.066	5.116	69M	358.1	[paper] [torch]
Inception1	33.160	12.324	10.246	7.0M	165.1	[paper] [tf-slim] [caffe-zoo]
Inception2	26.296	8.270	6.882	11M	134.3	[paper] [tf-slim]
Inception3	22.102	6.280	5.038	24M	314.6	[paper] [tf-slim] [keras]
Inception4	19.880	5.022	4.206	43M	582.1	[paper] [tf-slim]
InceptionResNet2	19.744	4.748	3.962	56M	656.8	[paper] [tf-slim]
NASNetAlarge	17.502	3.996	3.412	94M	2081	[paper] [tf-slim]
NASNetAmobile	25.634	8.146	6.758	7.7M	165.8	[paper] [tf-slim]
VGG16	28.732	9.950	8.834	138M	348.4	[paper] [keras]
VGG19	28.744	10.012	8.774	144M	399.8	[paper] [keras]
DenseNet121	25.480	8.022	6.842	8.1M	202.9	[paper] [torch]
DenseNet169	23.926	6.892	6.140	14M	219.1	[paper] [torch]
DenseNet201	22.936	6.542	5.724	20M	272.0	[paper] [torch]
MobileNet25	48.418	24.208	21.196	0.48M	29.27	[paper] [tf-slim]
MobileNet50	35.708	14.376	12.180	1.3M	42.32	[paper] [tf-slim]
MobileNet75	31.588	11.758	9.878	2.6M	57.23	[paper] [tf-slim]
MobileNet100	29.576	10.496	8.774	4.3M	70.69	[paper] [tf-slim]
SqueezeNet	45.566	21.960	18.578	1.2M	71.43	[paper] [caffe]

Object detection

• The object detection models can be coupled with any network but mAPs could be measured only for the models with pre-trained weights. Note that:
  • YOLOv2VOC is equivalent to YOLOv2(inputs, Darknet19),
  • TinyYOLOv2VOC: TinyYOLOv2(inputs, TinyDarknet19),
  • FasterRCNN_ZF_VOC: FasterRCNN(inputs, ZF),
  • FasterRCNN_VGG16_VOC: FasterRCNN(inputs, VGG16, stem_out='conv5/3').
• The mAPs were obtained with TensorNets on PASCAL VOC2007 test set and may slightly differ from the original ones.
• The test input sizes were the numbers reported as the best in the papers:
  • YOLOv2: 416x416
  • FasterRCNN: min_shorter_side=600, max_longer_side=1000
• The sizes stand for rounded the number of parameters.
• The computation times were measured on NVIDIA Tesla P100 (3584 cores, 16 GB global memory) with cuDNN 6.0 and CUDA 8.0.
  • Speed: milliseconds only for network inferences of a 416x416 single image
  • FPS: 1000 / speed

	mAP	Size	Speed	FPS	References
YOLOv2VOC	0.7320	51M	14.75	67.80	[paper] [darknet] [darkflow]
TinyYOLOv2VOC	0.5303	16M	6.534	153.0	[paper] [darknet] [darkflow]
FasterRCNN_ZF_VOC	0.4466	59M	241.4	3.325	[paper] [caffe] [roi-pooling]
FasterRCNN_VGG16_VOC	0.6872	137M	300.7	4.143	[paper] [caffe] [roi-pooling]