Add PNASNetlarge with pretrained weights

README.md

@@ -272,6 +272,7 @@ with tf.Session() as sess:
 | [InceptionResNet2](tensornets/inceptions.py#L264) | 19.744      | 4.748       | 3.962       | 55.9M  | 54.3M  | 656.8 | [[paper]](https://arxiv.org/abs/1602.07261) [[tf-slim]](https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py) |
 | [NASNetAlarge](tensornets/nasnets.py#L100)        | 17.502      | 3.996       | 3.412       | 93.5M  | 89.5M  | 2081  | [[paper]](https://arxiv.org/abs/1707.07012) [[tf-slim]](https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet) |
 | [NASNetAmobile](tensornets/nasnets.py#L108)       | 25.634      | 8.146       | 6.758       | 7.7M   | 6.7M   | 165.8 | [[paper]](https://arxiv.org/abs/1707.07012) [[tf-slim]](https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet) |
+| [PNASNetlarge](tensornets/nasnets.py#L100)        | 17.366      | 3.950       | 3.358       | 86.2M  | 81.9M  | 2081  | [[paper]](https://arxiv.org/abs/1712.00559) [[tf-slim]](https://github.com/tensorflow/models/tree/master/research/slim/nets/nasnet) |
 | [VGG16](tensornets/vggs.py#L71)                   | 28.732      | 9.950       | 8.834       | 138.4M | 14.7M  | 348.4 | [[paper]](https://arxiv.org/abs/1409.1556) [[keras]](https://github.com/keras-team/keras/blob/master/keras/applications/vgg16.py) |
 | [VGG19](tensornets/vggs.py#L78)                   | 28.744      | 10.012      | 8.774       | 143.7M | 20.0M  | 399.8 | [[paper]](https://arxiv.org/abs/1409.1556) [[keras]](https://github.com/keras-team/keras/blob/master/keras/applications/vgg19.py) |
 | [DenseNet121](tensornets/densenets.py#L63)        | 25.480      | 8.022       | 6.842       | 8.1M   | 7.0M   | 202.9 | [[paper]](https://arxiv.org/abs/1608.06993) [[torch]](https://github.com/liuzhuang13/DenseNet/blob/master/models/densenet.lua) |
@@ -316,14 +317,15 @@ with tf.Session() as sess:
 
 ## News 📰
 
+- PNASNetlarge is released, [12 May 2018]().
 - The six variants of MobileNetv2 are released, [5 May 2018](https://github.com/taehoonlee/tensornets/commit/fb429b6637f943875249dff50f4bc6220d9d50bf).
 - YOLOv3 for COCO and VOC are released, [4 April 2018](https://github.com/taehoonlee/tensornets/commit/d8b2d8a54dc4b775a174035da63561028deb6624).
 - Generic object detection models for YOLOv2 and FasterRCNN are released, [26 March 2018](https://github.com/taehoonlee/tensornets/commit/67915e659d2097a96c82ba7740b9e43a8c69858d).
 
 ## Future work 🔥
 
 - Add training codes.
-- Add image classification models (PolyNet, PNASNet).
+- Add image classification models (PolyNet).
 - Add object detection models (MaskRCNN, SSD).
 - Add image segmentation models (FCN, UNet).
 - Add image datasets (COCO, OpenImages).

tensornets/__init__.py

@@ -23,6 +23,7 @@
 
 from .nasnets import NASNetAlarge
 from .nasnets import NASNetAmobile
+from .nasnets import PNASNetlarge
 
 from .vggs import VGG16
 from .vggs import VGG19

tensornets/middles.py

@@ -35,9 +35,12 @@ def names_squeezenet():
 
 
 def names_nasnets(k):
-    names = ["normal%d/concat:0" % (i + 1) for i in range(k)]
-    names.insert(k // 3, "reduction%d/concat:0" % (k // 3))
-    names.insert(k // 3 * 2 + 1, "reduction%d/concat:0" % (k // 3 * 2))
+    names = []
+    for i in range(3):
+        base = sum(k[:i])
+        names += ["normal%d/concat:0" % (base + j + 1) for j in range(k[i])]
+        if i < 2:
+            names += ["reduction%d/concat:0" % (base + k[i])]
     return names
 
 
@@ -210,15 +213,21 @@ def direct(model_name):
     'nasnetAlarge': (
         list(range(145, 371, 45)) + [416] + list(range(466, 692, 45)) + [748] +
         list(range(798, 1024, 45)),
-        names_nasnets(18),
+        names_nasnets([6, 6, 6]),
         -8
     ),
     'nasnetAmobile': (
         list(range(145, 281, 45)) + [326] + list(range(376, 512, 45)) + [568] +
         list(range(618, 754, 45)),
-        names_nasnets(12),
+        names_nasnets([4, 4, 4]),
         -6
     ),
+    'pnasnetlarge': (
+        list(range(169, 323, 51)) + [376] + list(range(432, 535, 51)) + [588] +
+        list(range(644, 747, 51)),
+        names_nasnets([4, 3, 3]),
+        -5
+    ),
     'vgg16': (
         list(range(11, 16, 2)) + list(range(18, 23, 2)) +
         list(range(25, 30, 2)),

tensornets/nasnets.py

@@ -1,17 +1,22 @@
 """Collection of NASNet variants
 
-The reference paper:
+The reference papers:
 
+1. Original (a.k.a. NASNet)
  - Learning Transferable Architectures for Scalable Image Recognition,
    arXiv 2017
  - Barret Zoph, Vijay Vasudevan, Jonathon Shlens, Quoc V. Le
  - https://arxiv.org/abs/1707.07012
+2. PNASNet
+ - Progressive Neural Architecture Search, arXiv 2017
+ - Chenxi Liu et al.
+ - https://arxiv.org/abs/1712.00559
 
 The reference implementation:
 
 1. TF Slim
  - https://github.com/tensorflow/models/blob/master/research/slim/nets/
-   nasnet/nasnet.py
+   nasnet/{nasnet,pnasnet}.py
 """
 from __future__ import absolute_import
 from __future__ import division
@@ -66,16 +71,15 @@ def nasnet(x, stem_filters, normals, filters, skip_reduction, use_aux,
     x, p = reductionA(x, None, filters * scaling ** (-2), scope='stem1')
     x, p = reductionA(x, p, filters * scaling ** (-1), scope='stem2')
 
-    for i in range(normals):
-        x, p = normalA(x, p, filters, scope="normal%d" % (i + 1))
+    for i in range(1, normals + 1):
+        x, p = normalA(x, p, filters, scope="normal%d" % i)
 
     x, p0 = reductionA(x, p, filters * scaling,
                        scope="reduction%d" % normals)
     p = p0 if not skip_reduction else p
 
-    for i in range(normals):
-        x, p = normalA(x, p, filters * scaling,
-                       scope="normal%d" % (i + normals + 1))
+    for i in range(normals + 1, normals * 2 + 1):
+        x, p = normalA(x, p, filters * scaling, scope="normal%d" % i)
 
     if use_aux is True:
         a = aux(x, classes, scope='aux')
@@ -84,9 +88,8 @@ def nasnet(x, stem_filters, normals, filters, skip_reduction, use_aux,
                        scope="reduction%d" % (normals * 2))
     p = p0 if not skip_reduction else p
 
-    for i in range(normals):
-        x, p = normalA(x, p, filters * scaling ** 2,
-                       scope="normal%d" % (i + normals * 2 + 1))
+    for i in range(normals * 2 + 1, normals * 3 + 1):
+        x, p = normalA(x, p, filters * scaling ** 2, scope="normal%d" % i)
 
     x = relu(x, name='relu')
     if stem: return x
@@ -113,6 +116,44 @@ def nasnetAmobile(x, is_training=False, classes=1000,
                   is_training, classes, stem, scope, reuse)
 
 
+def pnasnet(x, stem_filters, blocks, filters,
+            scaling, is_training, classes, stem, scope=None, reuse=None):
+    x = conv(x, stem_filters, 3, stride=2, padding='VALID', scope='conv0')
+
+    x, p = normalP(x, None, filters * scaling ** (-2), stride=2, scope='stem1')
+    x, p = normalP(x, p, filters * scaling ** (-1), stride=2, scope='stem2')
+
+    for i in range(1, blocks + 1):
+        x, p = normalP(x, p, filters, scope="normal%d" % i)
+
+    x, p = normalP(x, p, filters * scaling, stride=2,
+                   scope="reduction%d" % blocks)
+
+    for i in range(blocks + 1, blocks * 2):
+        x, p = normalP(x, p, filters * scaling, scope="normal%d" % i)
+
+    x, p = normalP(x, p, filters * scaling ** 2, stride=2,
+                   scope="reduction%d" % (blocks * 2 - 1))
+
+    for i in range(blocks * 2, blocks * 3 - 1):
+        x, p = normalP(x, p, filters * scaling ** 2, scope="normal%d" % i)
+
+    x = relu(x, name='relu')
+    if stem: return x
+    x = reduce_mean(x, [1, 2], name='avgpool')
+    x = dropout(x, keep_prob=0.5, scope='dropout')
+    x = fc(x, classes, scope='logits')
+    x = softmax(x, name='probs')
+    return x
+
+
+@var_scope('pnasnetlarge')
+@set_args(__args__)
+def pnasnetlarge(x, is_training=False, classes=1000,
+                 stem=False, scope=None, reuse=None):
+    return pnasnet(x, 96, 4, 216, 2, is_training, classes, stem, scope, reuse)
+
+
 @var_scope('adjust')
 def adjust(p, x, filters, scope=None):
     if p is None:
@@ -173,6 +214,37 @@ def reductionA(x, p, filters, scope=None):
     return concat([x2, x3, x5, x4], axis=3, name='concat'), x
 
 
+@var_scope('pool')
+def pool(x, filters, stride, scope=None):
+    y = max_pool2d(x, 3, stride=stride)
+    if int(x.shape[-1]) != filters:
+        y = conv(y, filters, 1, scope='1x1')
+    return y
+
+
+@var_scope('normalP')
+def normalP(x, p, filters, stride=1, scope=None):
+    p = adjust(p, x, filters)
+
+    h = relu(x)
+    h = conv(h, filters, 1, scope='1x1')
+
+    x1 = add(sconv(p, filters, 5, stride, scope='left1'),
+             pool(p, filters, stride, scope='right1'), name='add1')
+    x2 = add(sconv(h, filters, 7, stride, scope='left2'),
+             pool(h, filters, stride, scope='right2'), name='add2')
+    x3 = add(sconv(h, filters, 5, stride, scope='left3'),
+             sconv(h, filters, 3, stride, scope='right3'), name='add3')
+    x4 = add(sconv(x3, filters, 3, scope='left4'),
+             pool(h, filters, stride, scope='right4'), name='add4')
+    x5 = add(
+        sconv(p, filters, 3, stride, scope='left5'),
+        conv(relu(h), filters, 1, stride, scope='right5') if stride > 1 else h,
+        name='add5')
+
+    return concat([x1, x2, x3, x4, x5], axis=3, name='concat'), x
+
+
 @var_scope('aux')
 def aux(x, classes, scope=None):
     x = relu(x, name='relu1')
@@ -190,3 +262,4 @@ def aux(x, classes, scope=None):
 # Simple alias.
 NASNetAlarge = nasnetAlarge
 NASNetAmobile = nasnetAmobile
+PNASNetlarge = pnasnetlarge

tensornets/preprocess.py

@@ -154,6 +154,7 @@ def faster_rcnn_preprocess(x):
     'wideresnet50': wrn_preprocess,
     'nasnetAlarge': tfslim_preprocess,
     'nasnetAmobile': tfslim_preprocess,
+    'pnasnetlarge': tfslim_preprocess,
     'vgg16': keras_resnet_preprocess,
     'vgg19': keras_resnet_preprocess,
     'densenet': fb_preprocess,

tensornets/pretrained.py

@@ -629,6 +629,17 @@ def load_nasnetAmobile(scopes, return_fn=_assign):
     return return_fn(scopes, values)
 
 
+def load_pnasnetlarge(scopes, return_fn=_assign):
+    """Converted from the [TF Slim][2]."""
+    filename = 'pnasnet_large.npz'
+    weights_path = get_file(
+        filename, __model_url__ + 'nasnet/' + filename,
+        cache_subdir='models',
+        file_hash='a1afc7679b950b643865aa7ea716c901')
+    values = parse_weights(weights_path)
+    return return_fn(scopes, values)
+
+
 def load_vgg16(scopes, return_fn=_assign):
     """Copied from [Keras][3]."""
     filename = 'vgg16_weights_tf_dim_ordering_tf_kernels.h5'
@@ -761,6 +772,7 @@ def load_ref_faster_rcnn_vgg16_voc(scopes, return_fn=_assign):
     'wideresnet50': load_wideresnet50,
     'nasnetAlarge': load_nasnetAlarge,
     'nasnetAmobile': load_nasnetAmobile,
+    'pnasnetlarge': load_pnasnetlarge,
     'vgg16': load_vgg16,
     'vgg19': load_vgg19,
     'densenet121': load_densenet121,

tests/basics_test.py

@@ -50,6 +50,11 @@
         marks=pytest.mark.xfail(
             LooseVersion(tf.__version__) < LooseVersion('1.3.0'),
             reason='NASNetAmobile requires TensorFlow >= 1.3.0')),
+    pytest.param(
+        nets.PNASNetlarge, (331, 331, 3),
+        marks=pytest.mark.xfail(
+            LooseVersion(tf.__version__) < LooseVersion('1.3.0'),
+            reason='PNASNetlarge requires TensorFlow >= 1.3.0')),
     random.choice([
         (nets.VGG16, (224, 224, 3)),
         (nets.VGG19, (224, 224, 3)),
@@ -87,6 +92,7 @@
     'InceptionResNet2',
     'NASNetAlarge',
     'NASNetAmobile',
+    'PNASNetlarge',
     'VGG',
     'DenseNet',
     'MobileNet',

translations/tfslim.py

@@ -0,0 +1,63 @@
+import numpy as np
+import tensorflow as tf
+import tensornets as nets
+import tensorflow_hub as hub
+
+models_list = [
+    (nets.MobileNet35v2, (224, 224, 3), 'mobilenet_v2_035_224'),
+    (nets.MobileNet50v2, (224, 224, 3), 'mobilenet_v2_050_224'),
+    (nets.MobileNet75v2, (224, 224, 3), 'mobilenet_v2_075_224'),
+    (nets.MobileNet100v2, (224, 224, 3), 'mobilenet_v2_100_224'),
+    (nets.MobileNet130v2, (224, 224, 3), 'mobilenet_v2_130_224'),
+    (nets.MobileNet140v2, (224, 224, 3), 'mobilenet_v2_140_224'),
+    (nets.PNASNetlarge, (331, 331, 3), 'pnasnet_large'),
+]
+
+url = 'https://tfhub.dev/google/imagenet'
+
+
+for (net, shape, model_name) in models_list:
+
+    with tf.Graph().as_default():
+
+        inputs = tf.placeholder(tf.float32, [None] + list(shape))
+        model = net(inputs, scope='a')
+
+        tfhub = hub.Module("%s/%s/classification/1" % (url, model_name))
+        features = tfhub(inputs, signature="image_classification",
+                         as_dict=True)
+        model_tfhub = tf.nn.softmax(features['default'])
+
+        img = nets.utils.load_img('cat.png',
+                                  target_size=int(shape[0] * 8 / 7),
+                                  crop_size=shape[0])
+
+        with tf.Session() as sess:
+
+            # Retrieve values
+            sess.run(tf.global_variables_initializer())
+            weights = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
+                                        scope='module')
+            values = sess.run(weights)
+            for i in range(-2, 0):
+                values[i] = np.delete(np.squeeze(values[i]), 0, axis=-1)
+
+            # Adjust the order of the values to cover TF < 1.4.0
+            names = [w.name[2:] for w in model.get_weights()]
+            for i in range(len(names) - 1):
+                if 'gamma:0' in names[i] and 'beta:0' in names[i + 1]:
+                    names[i], names[i + 1] = names[i + 1], names[i]
+                    values[i], values[i + 1] = values[i + 1], values[i]
+
+            # Save the values as the TensorNets format
+            np.savez(model_name, names=names, values=values)
+
+            # Load and set the values
+            weights = model.get_weights()
+            values = nets.utils.parse_weights(model_name + '.npz')
+            sess.run([w.assign(v) for (w, v) in zip(weights, values)])
+
+            # Run equivalence tests
+            preds = sess.run(model, {inputs: model.preprocess(img)})
+            preds_tfhub = sess.run(model_tfhub, {inputs: img / 255.})
+            np.testing.assert_allclose(preds, preds_tfhub[:, 1:], atol=1e-4)