PifPaf_tracking_heads.py

densenet of tensorlayer by yjk

densenet.py

@@ -1,153 +1,153 @@
-import os
-os.environ['TL_BACKEND'] = 'tensorflow'
-import time
-import multiprocessing
-import tensorflow as tf
-from tensorlayer.models import TrainOneStep
-from tensorlayer.layers import Module
-import tensorlayer as tl
-from torchsummary import summary
-from tensorlayer import logging
-from tensorlayer.files import (assign_weights, maybe_download_and_extract)
-from tensorlayer.layers import (BatchNorm, Conv2d, Dense, Elementwise, AdaptiveMeanPool2d, MaxPool2d , MeanPool2d,Concat,Dropout)
-from tensorlayer.layers import Module, SequentialLayer
-
-
-class _DenseLayer(Module):
-    def __init__(self, in_channels, growth_rate, bn_size):
-        super(_DenseLayer, self).__init__()
-        W_init = tl.initializers.truncated_normal(stddev=5e-2)
-        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
-        b_init2 = tl.initializers.constant(value=0.1)
-        self.layer_list = []
-        self.layer_list.append(Conv2d(bn_size * growth_rate,(1,1),in_channels=in_channels,W_init=W_init))
-        self.layer_list.append(BatchNorm(num_features=bn_size * growth_rate,act='relu'))
-        self.layer_list.append(Conv2d(growth_rate, (3, 3), in_channels=bn_size * growth_rate,W_init=W_init))
-        self.layer_list.append(BatchNorm(num_features=growth_rate, act='relu'))
-        self.dense_layer = SequentialLayer(self.layer_list)
-        self.concat = Concat(1)
-
-    # 重载forward函数
-    def forward(self, x):
-        new_features = self.dense_layer(x)
-        return self.concat([x, new_features])
-
-
-class _DenseBlock(Module):
-    def __init__(self, num_layers, in_channels, bn_size, growth_rate):
-        super(_DenseBlock, self).__init__()
-        W_init = tl.initializers.truncated_normal(stddev=5e-2)
-        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
-        b_init2 = tl.initializers.constant(value=0.1)
-        self.layer_list = []
-        for i in range(num_layers):
-            self.layer_list.append(_DenseLayer(in_channels + growth_rate * i,growth_rate, bn_size))
-        self.dense_block = SequentialLayer(self.layer_list)
-
-    # 重载forward函数
-    def forward(self, x):
-        return self.dense_block(x)
-
-
-class _Transition(Module):
-    def __init__(self, in_channels, out_channels):
-        super(_Transition, self).__init__()
-        W_init = tl.initializers.truncated_normal(stddev=5e-2)
-        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
-        b_init2 = tl.initializers.constant(value=0.1)
-        self.layer_list = []
-        self.layer_list.append(Conv2d(out_channels,(1,1),in_channels=in_channels,W_init=W_init))
-        self.layer_list.append(BatchNorm(num_features=out_channels,act='relu'))
-        self.layer_list.append(MeanPool2d((2,2),strides=(2,2)))
-        self.transition_layer = SequentialLayer(self.layer_list)
-
-    # 重载forward函数
-    def forward(self, x):
-        return self.transition_layer(x)
-
-class DenseNet_BC(Module):
-    def __init__(self, growth_rate=12, block_config=(6, 12, 24, 16),
-                 bn_size=4, theta=0.5, num_classes=10):
-        super(DenseNet_BC, self).__init__()
-        W_init = tl.initializers.truncated_normal(stddev=5e-2)
-        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
-        b_init2 = tl.initializers.constant(value=0.1)
-        # 初始的卷积为filter:2倍的growth_rate
-        num_init_feature = 2 * growth_rate
-        self.layer_list = []
-        # 表示cifar-10
-        if num_classes == 10:
-            self.layer_list.append(Conv2d(num_init_feature,(3,3),strides=(1,1),in_channels=3,W_init=W_init))
-
-        else:
-            self.layer_list.append(Conv2d(num_init_feature,(7,7),strides=(2,2),padding="valid",in_channels=3,W_init=W_init))
-            self.layer_list.append(BatchNorm(num_features=num_init_feature,act='relu'))
-            self.layer_list.append(MaxPool2d((3, 3), strides=(2, 2)))
-
-
-        num_feature = num_init_feature
-        for i, num_layers in enumerate(block_config):
-            self.layer_list.append( _DenseBlock(num_layers, num_feature,bn_size, growth_rate))
-            num_feature = num_feature + growth_rate * num_layers
-            if i != len(block_config) - 1:
-                self.layer_list.append(_Transition(num_feature,int(num_feature * theta)))
-                num_feature = int(num_feature * theta)
-
-        self.layer_list.append(BatchNorm(num_features=num_feature,act='relu'))
-        self.layer_list.append(AdaptiveMeanPool2d((1,1)))
-
-        self.features = SequentialLayer(self.layer_list)
-        self.classifier = Dense(num_feature, num_classes,W_init=W_init2,b_init=b_init2)
-
-
-    def forward(self, x):
-        features = self.features(x)
-        out = features.view(features.size(0), -1)
-        out = self.classifier(out)
-        return out
-
-
-# DenseNet_BC for ImageNet
-def DenseNet121():
-    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 24, 16), num_classes=1000)
-
-
-def DenseNet169():
-    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 32, 32), num_classes=1000)
-
-
-def DenseNet201():
-    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 48, 32), num_classes=1000)
-
-
-def DenseNet161():
-    return DenseNet_BC(growth_rate=48, block_config=(6, 12, 36, 24), num_classes=1000, )
-
-
-# DenseNet_BC for cifar
-def densenet_BC_100():
-    return DenseNet_BC(growth_rate=12, block_config=(16, 16, 16))
-
-def builddensenet(name = "densenet-100"):
-    if name == "densenet-100":
-        return densenet_BC_100()
-    elif name == "densenet-121":
-        return DenseNet121()
-    else:
-        print("not found the net")
-        exit(0)
-
-def test():
-    net = densenet_BC_100()
-    print(summary(net, input_size=(3, 32, 32)))
-
-    #x = torch.randn(2, 3, 32, 32)
-    # y = net(x)
-    # print(y.size())
-
-
-if __name__ == '__main__':
-    test()
-
-
-
+import os
+os.environ['TL_BACKEND'] = 'tensorflow'
+import time
+import multiprocessing
+import tensorflow as tf
+from tensorlayer.models import TrainOneStep
+from tensorlayer.layers import Module
+import tensorlayer as tl
+from torchsummary import summary
+from tensorlayer import logging
+from tensorlayer.files import (assign_weights, maybe_download_and_extract)
+from tensorlayer.layers import (BatchNorm, Conv2d, Dense, Elementwise, AdaptiveMeanPool2d, MaxPool2d , MeanPool2d,Concat,Dropout)
+from tensorlayer.layers import Module, SequentialLayer
+
+
+class _DenseLayer(Module):
+    def __init__(self, in_channels, growth_rate, bn_size):
+        super(_DenseLayer, self).__init__()
+        W_init = tl.initializers.truncated_normal(stddev=5e-2)
+        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
+        b_init2 = tl.initializers.constant(value=0.1)
+        self.layer_list = []
+        self.layer_list.append(Conv2d(bn_size * growth_rate,(1,1),in_channels=in_channels,W_init=W_init))
+        self.layer_list.append(BatchNorm(num_features=bn_size * growth_rate,act='relu'))
+        self.layer_list.append(Conv2d(growth_rate, (3, 3), in_channels=bn_size * growth_rate,W_init=W_init))
+        self.layer_list.append(BatchNorm(num_features=growth_rate, act='relu'))
+        self.dense_layer = SequentialLayer(self.layer_list)
+        self.concat = Concat(1)
+
+    # 重载forward函数
+    def forward(self, x):
+        new_features = self.dense_layer(x)
+        return self.concat([x, new_features])
+
+
+class _DenseBlock(Module):
+    def __init__(self, num_layers, in_channels, bn_size, growth_rate):
+        super(_DenseBlock, self).__init__()
+        W_init = tl.initializers.truncated_normal(stddev=5e-2)
+        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
+        b_init2 = tl.initializers.constant(value=0.1)
+        self.layer_list = []
+        for i in range(num_layers):
+            self.layer_list.append(_DenseLayer(in_channels + growth_rate * i,growth_rate, bn_size))
+        self.dense_block = SequentialLayer(self.layer_list)
+
+    # 重载forward函数
+    def forward(self, x):
+        return self.dense_block(x)
+
+
+class _Transition(Module):
+    def __init__(self, in_channels, out_channels):
+        super(_Transition, self).__init__()
+        W_init = tl.initializers.truncated_normal(stddev=5e-2)
+        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
+        b_init2 = tl.initializers.constant(value=0.1)
+        self.layer_list = []
+        self.layer_list.append(Conv2d(out_channels,(1,1),in_channels=in_channels,W_init=W_init))
+        self.layer_list.append(BatchNorm(num_features=out_channels,act='relu'))
+        self.layer_list.append(MeanPool2d((2,2),strides=(2,2)))
+        self.transition_layer = SequentialLayer(self.layer_list)
+
+    # 重载forward函数
+    def forward(self, x):
+        return self.transition_layer(x)
+
+class DenseNet_BC(Module):
+    def __init__(self, growth_rate=12, block_config=(6, 12, 24, 16),
+                 bn_size=4, theta=0.5, num_classes=10):
+        super(DenseNet_BC, self).__init__()
+        W_init = tl.initializers.truncated_normal(stddev=5e-2)
+        W_init2 = tl.initializers.truncated_normal(stddev=0.04)
+        b_init2 = tl.initializers.constant(value=0.1)
+        # 初始的卷积为filter:2倍的growth_rate
+        num_init_feature = 2 * growth_rate
+        self.layer_list = []
+        # 表示cifar-10
+        if num_classes == 10:
+            self.layer_list.append(Conv2d(num_init_feature,(3,3),strides=(1,1),in_channels=3,W_init=W_init))
+
+        else:
+            self.layer_list.append(Conv2d(num_init_feature,(7,7),strides=(2,2),padding="valid",in_channels=3,W_init=W_init))
+            self.layer_list.append(BatchNorm(num_features=num_init_feature,act='relu'))
+            self.layer_list.append(MaxPool2d((3, 3), strides=(2, 2)))
+
+
+        num_feature = num_init_feature
+        for i, num_layers in enumerate(block_config):
+            self.layer_list.append( _DenseBlock(num_layers, num_feature,bn_size, growth_rate))
+            num_feature = num_feature + growth_rate * num_layers
+            if i != len(block_config) - 1:
+                self.layer_list.append(_Transition(num_feature,int(num_feature * theta)))
+                num_feature = int(num_feature * theta)
+
+        self.layer_list.append(BatchNorm(num_features=num_feature,act='relu'))
+        self.layer_list.append(AdaptiveMeanPool2d((1,1)))
+
+        self.features = SequentialLayer(self.layer_list)
+        self.classifier = Dense(num_feature, num_classes,W_init=W_init2,b_init=b_init2)
+
+
+    def forward(self, x):
+        features = self.features(x)
+        out = features.view(features.size(0), -1)
+        out = self.classifier(out)
+        return out
+
+
+# DenseNet_BC for ImageNet
+def DenseNet121():
+    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 24, 16), num_classes=1000)
+
+
+def DenseNet169():
+    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 32, 32), num_classes=1000)
+
+
+def DenseNet201():
+    return DenseNet_BC(growth_rate=32, block_config=(6, 12, 48, 32), num_classes=1000)
+
+
+def DenseNet161():
+    return DenseNet_BC(growth_rate=48, block_config=(6, 12, 36, 24), num_classes=1000, )
+
+
+# DenseNet_BC for cifar
+def densenet_BC_100():
+    return DenseNet_BC(growth_rate=12, block_config=(16, 16, 16))
+
+def builddensenet(name = "densenet-100"):
+    if name == "densenet-100":
+        return densenet_BC_100()
+    elif name == "densenet-121":
+        return DenseNet121()
+    else:
+        print("not found the net")
+        exit(0)
+
+def test():
+    net = densenet_BC_100()
+    print(summary(net, input_size=(3, 32, 32)))
+
+    #x = torch.randn(2, 3, 32, 32)
+    # y = net(x)
+    # print(y.size())
+
+
+if __name__ == '__main__':
+    test()
+
+
+

headmeta.py

@@ -0,0 +1,192 @@
+"""Head meta objects contain meta information about head networks.
+
+This includes the name, the name of the individual fields, the composition, etc.
+"""
+
+from dataclasses import dataclass, field
+from typing import Any, ClassVar, List, Tuple
+
+import numpy as np
+
+
+@dataclass
+class Base:
+    name: str
+    dataset: str
+
+    head_index: int = field(default=None, init=False)
+    base_stride: int = field(default=None, init=False)
+    upsample_stride: int = field(default=1, init=False)
+
+    @property
+    def stride(self) -> int:
+        if self.base_stride is None:
+            return None
+        return self.base_stride // self.upsample_stride
+
+    @property
+    def n_fields(self) -> int:
+        raise NotImplementedError
+
+
+@dataclass
+class Cif(Base):
+    """Head meta data for a Composite Intensity Field (CIF)."""
+
+    keypoints: List[str]
+    sigmas: List[float]
+    pose: Any = None
+    draw_skeleton: List[Tuple[int, int]] = None
+    score_weights: List[float] = None
+
+    n_confidences: ClassVar[int] = 1
+    n_vectors: ClassVar[int] = 1
+    n_scales: ClassVar[int] = 1
+
+    vector_offsets = [True]
+    decoder_min_scale = 0.0
+    decoder_seed_mask: List[int] = None
+
+    training_weights: List[float] = None
+
+    @property
+    def n_fields(self):
+        return len(self.keypoints)
+
+
+@dataclass
+class Caf(Base):
+    """Head meta data for a Composite Association Field (CAF)."""
+
+    keypoints: List[str]
+    sigmas: List[float]
+    skeleton: List[Tuple[int, int]]
+    pose: Any = None
+    sparse_skeleton: List[Tuple[int, int]] = None
+    dense_to_sparse_radius: float = 2.0
+    only_in_field_of_view: bool = False
+
+    n_confidences: ClassVar[int] = 1
+    n_vectors: ClassVar[int] = 2
+    n_scales: ClassVar[int] = 2
+
+    vector_offsets = [True, True]
+    decoder_min_distance = 0.0
+    decoder_max_distance = float('inf')
+    decoder_confidence_scales: List[float] = None
+
+    training_weights: List[float] = None
+
+    @property
+    def n_fields(self):
+        return len(self.skeleton)
+
+    @staticmethod
+    def concatenate(metas):
+        # TODO: by keypoint name, update skeleton indices if meta.keypoints
+        # is not the same for all metas.
+        concatenated = Caf(
+            name='_'.join(m.name for m in metas),
+            dataset=metas[0].dataset,
+            keypoints=metas[0].keypoints,
+            sigmas=metas[0].sigmas,
+            pose=metas[0].pose,
+            skeleton=[s for meta in metas for s in meta.skeleton],
+            sparse_skeleton=metas[0].sparse_skeleton,
+            only_in_field_of_view=metas[0].only_in_field_of_view,
+            decoder_confidence_scales=[
+                s
+                for meta in metas
+                for s in (meta.decoder_confidence_scales
+                          if meta.decoder_confidence_scales
+                          else [1.0 for _ in meta.skeleton])
+            ]
+        )
+        concatenated.head_index = metas[0].head_index
+        concatenated.base_stride = metas[0].base_stride
+        concatenated.upsample_stride = metas[0].upsample_stride
+        return concatenated
+
+
+@dataclass
+class CifDet(Base):
+    """Head meta data for a Composite Intensity Field (CIF) for Detection."""
+
+    categories: List[str]
+
+    n_confidences: ClassVar[int] = 1
+    n_vectors: ClassVar[int] = 2
+    n_scales: ClassVar[int] = 0
+
+    vector_offsets = [True, False]
+    decoder_min_scale = 0.0
+
+    training_weights: List[float] = None
+
+    @property
+    def n_fields(self):
+        return len(self.categories)
+
+
+@dataclass
+class TSingleImageCif(Cif):
+    """Single-Image CIF head in tracking models."""
+
+
+@dataclass
+class TSingleImageCaf(Caf):
+    """Single-Image CAF head in tracking models."""
+
+
+@dataclass
+class Tcaf(Base):
+    """Tracking Composite Association Field."""
+
+    keypoints_single_frame: List[str]
+    sigmas_single_frame: List[float]
+    pose_single_frame: Any
+    draw_skeleton_single_frame: List[Tuple[int, int]] = None
+    keypoints: List[str] = None
+    sigmas: List[float] = None
+    pose: Any = None
+    draw_skeleton: List[Tuple[int, int]] = None
+    only_in_field_of_view: bool = False
+
+    n_confidences: ClassVar[int] = 1
+    n_vectors: ClassVar[int] = 2
+    n_scales: ClassVar[int] = 2
+
+    training_weights: List[float] = None
+
+    vector_offsets = [True, True]
+
+    def __post_init__(self):
+        if self.keypoints is None:
+            self.keypoints = np.concatenate((
+                self.keypoints_single_frame,
+                self.keypoints_single_frame,
+            ), axis=0)
+        if self.sigmas is None:
+            self.sigmas = np.concatenate((
+                self.sigmas_single_frame,
+                self.sigmas_single_frame,
+            ), axis=0)
+        if self.pose is None:
+            self.pose = np.concatenate((
+                self.pose_single_frame,
+                self.pose_single_frame,
+            ), axis=0)
+        if self.draw_skeleton is None:
+            self.draw_skeleton = np.concatenate((
+                self.draw_skeleton_single_frame,
+                self.draw_skeleton_single_frame,
+            ), axis=0)
+
+    @property
+    def skeleton(self):
+        return [(i + 1, i + 1 + len(self.keypoints_single_frame))
+                for i, _ in enumerate(self.keypoints_single_frame)]
+
+    @property
+    def n_fields(self):
+        return len(self.keypoints_single_frame)

heads.py

@@ -0,0 +1,412 @@
+"""Head networks."""
+import os
+import argparse
+import functools
+import math
+import tensorlayer as tl
+from tensorlayer import logging
+from tensorlayer.layers import  Conv2d,Dropout
+import torch
+os.environ['TL_BACKEND'] = 'pytorch'
+from .import headmeta
+from tensorlayer.layers import Module
+import numpy as np
+LOG = logging.getLogger(__name__)
+
+@functools.lru_cache(maxsize=16)
+def index_field_torch(shape, *, device=None, unsqueeze=(0, 0)):
+    assert len(shape) == 2
+    xy = np.empty((2, shape[0], shape[1]), device=device)
+    xy = tl.convert_to_tensor(xy)
+    xy[0] = np.arange(shape[1], device=device)
+    xy[0] = tl.convert_to_tensor(xy[0])
+    xy[1] = np.arange(shape[0], device=device).unsqueeze(1)
+    xy[1] = tl.convert_to_tensor(xy[1])
+
+    for dim in unsqueeze:
+        xy = torch.unsqueeze(xy, dim)
+        xy = xy.cpu().numpy()
+        xy = tl.convert_to_tensor(xy)
+
+    return xy
+
+
+class PifHFlip(Module):
+    def __init__(self, keypoints, hflip):
+        super().__init__()
+
+        flip_indices = torch.LongTensor([
+            keypoints.index(hflip[kp_name]) if kp_name in hflip else kp_i
+            for kp_i, kp_name in enumerate(keypoints)
+        ])
+        LOG.debug('hflip indices: %s', flip_indices)
+        self.register_buffer('flip_indices', flip_indices)
+
+    def forward(self, *args):
+        out = []
+        for field in args:
+            field = torch.index_select(field, 1, self.flip_indices)
+            field = torch.flip(field, dims=[len(field.shape) - 1])
+            out.append(field)
+
+        # flip the x-coordinate of the vector component
+        out[1][:, :, 0, :, :] *= -1.0
+
+        return out
+
+
+class PafHFlip(Module):
+    def __init__(self, keypoints, skeleton, hflip):
+        super().__init__()
+        skeleton_names = [
+            (keypoints[j1 - 1], keypoints[j2 - 1])
+            for j1, j2 in skeleton
+        ]
+        flipped_skeleton_names = [
+            (hflip[j1] if j1 in hflip else j1, hflip[j2] if j2 in hflip else j2)
+            for j1, j2 in skeleton_names
+        ]
+        LOG.debug('skeleton = %s, flipped_skeleton = %s',
+                  skeleton_names, flipped_skeleton_names)
+
+        flip_indices = list(range(len(skeleton)))
+        reverse_direction = []
+        for paf_i, (n1, n2) in enumerate(skeleton_names):
+            if (n1, n2) in flipped_skeleton_names:
+                flip_indices[paf_i] = flipped_skeleton_names.index((n1, n2))
+            if (n2, n1) in flipped_skeleton_names:
+                flip_indices[paf_i] = flipped_skeleton_names.index((n2, n1))
+                reverse_direction.append(paf_i)
+        LOG.debug('hflip indices: %s, reverse: %s', flip_indices, reverse_direction)
+
+        self.register_buffer('flip_indices', torch.LongTensor(flip_indices))
+        self.register_buffer('reverse_direction', torch.LongTensor(reverse_direction))
+
+    def forward(self, *args):
+        out = []
+        for field in args:
+            field = torch.index_select(field, 1, self.flip_indices)
+            field = torch.flip(field, dims=[len(field.shape) - 1])
+            out.append(field)
+
+        # flip the x-coordinate of the vector components
+        out[1][:, :, 0, :, :] *= -1.0
+        out[2][:, :, 0, :, :] *= -1.0
+
+        # reverse direction
+        for paf_i in self.reverse_direction:
+            cc = torch.clone(out[1][:, paf_i])
+            out[1][:, paf_i] = out[2][:, paf_i]
+            out[2][:, paf_i] = cc
+
+        return out
+
+
+class HeadNetwork(Module):
+    """Base class for head networks.
+
+    :param meta: head meta instance to configure this head network
+    :param in_features: number of input features which should be equal to the
+        base network's output features
+    """
+    def __init__(self, meta: headmeta.Base, in_features: int):
+        super().__init__()
+        self.meta = meta
+        self.in_features = in_features
+
+    @classmethod
+    def cli(cls, parser: argparse.ArgumentParser):
+        """Command line interface (CLI) to extend argument parser."""
+
+    @classmethod
+    def configure(cls, args: argparse.Namespace):
+        """Take the parsed argument parser output and configure class variables."""
+
+    def forward(self, x):
+        raise NotImplementedError
+
+
+class CompositeField3(HeadNetwork):
+    dropout_p = 0.0
+    inplace_ops = True
+
+    def __init__(self,
+                 meta: headmeta.Base,
+                 in_features, *,
+                 kernel_size=1, padding=0, dilation=1):
+        super().__init__(meta, in_features)
+
+        LOG.debug('%s config: fields = %d, confidences = %d, vectors = %d, scales = %d '
+                  'kernel = %d, padding = %d, dilation = %d',
+                  meta.name, meta.n_fields, meta.n_confidences, meta.n_vectors, meta.n_scales,
+                  kernel_size, padding, dilation)
+
+        self.dropout = Dropout(p=self.dropout_p)
+
+
+        # convolution
+        out_features = meta.n_fields * (meta.n_confidences + meta.n_vectors * 3 + meta.n_scales)
+        self.conv = Conv2d(out_features * (meta.upsample_stride ** 2),
+                           (kernel_size,kernel_size), padding=(padding,padding), dilation=(dilation,dilation),in_channels=in_features)
+
+        # upsample
+        assert meta.upsample_stride >= 1
+        self.upsample_op = None
+        if meta.upsample_stride > 1:
+            self.upsample_op = torch.nn.PixelShuffle(meta.upsample_stride)
+
+    @classmethod
+    def cli(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group('CompositeField3')
+        group.add_argument('--cf3-dropout', default=cls.dropout_p, type=float,
+                           help='[experimental] zeroing probability of feature in head input')
+        assert cls.inplace_ops
+        group.add_argument('--cf3-no-inplace-ops', dest='cf3_inplace_ops',
+                           default=True, action='store_false',
+                           help='alternative graph without inplace ops')
+
+    @classmethod
+    def configure(cls, args: argparse.Namespace):
+        cls.dropout_p = args.cf3_dropout
+        cls.inplace_ops = args.cf3_inplace_ops
+
+    @property
+    def sparse_task_parameters(self):
+        return [self.conv.weight]
+
+    def forward(self, x):  # pylint: disable=arguments-differ
+        x = self.dropout(x)
+        x = self.conv(x)
+        # upscale
+        if self.upsample_op is not None:
+            x = self.upsample_op(x)
+            low_cut = (self.meta.upsample_stride - 1) // 2
+            high_cut = math.ceil((self.meta.upsample_stride - 1) / 2.0)
+            if self.training:
+                # negative axes not supported by ONNX TensorRT
+                x = x[:, :, low_cut:-high_cut, low_cut:-high_cut]
+            else:
+                # the int() forces the tracer to use static shape
+                x = x[:, :, low_cut:int(x.shape[2]) - high_cut, low_cut:int(x.shape[3]) - high_cut]
+
+        # Extract some shape parameters once.
+        # Convert to int so that shape is constant in ONNX export.
+        x_size = x.size()
+        batch_size = x_size[0]
+        feature_height = int(x_size[2])
+        feature_width = int(x_size[3])
+
+        x = x.view(
+            batch_size,
+            self.meta.n_fields,
+            self.meta.n_confidences + self.meta.n_vectors * 3 + self.meta.n_scales,
+            feature_height,
+            feature_width
+        )
+
+        if not self.training and self.inplace_ops:
+            # classification
+            classes_x = x[:, :, 0:self.meta.n_confidences]
+            tl.sigmoid(classes_x)
+
+            # regressions x: add index
+            if self.meta.n_vectors > 0:
+                index_field = index_field_torch((feature_height, feature_width), device=x.device)
+                first_reg_feature = self.meta.n_confidences
+                for i, do_offset in enumerate(self.meta.vector_offsets):
+                    if not do_offset:
+                        continue
+                    reg_x = x[:, :, first_reg_feature + i * 2:first_reg_feature + (i + 1) * 2]
+                    reg_x.add_(index_field)
+
+            # scale
+            first_scale_feature = self.meta.n_confidences + self.meta.n_vectors * 3
+            scales_x = x[:, :, first_scale_feature:first_scale_feature + self.meta.n_scales]
+            scales_x[:] = torch.nn.functional.softplus(scales_x)
+
+            # remove width in the middle and add one to the front (v4 style)
+            first_width_feature = self.meta.n_confidences + self.meta.n_vectors * 2
+            x = tl.concat([
+                x[:, :, first_width_feature:first_width_feature + 1],
+                x[:, :, :first_width_feature],
+                x[:, :, self.meta.n_confidences + self.meta.n_vectors * 3:],
+            ], dim=2)
+        elif not self.training and not self.inplace_ops:
+            # TODO: CoreMLv4 does not like strided slices.
+            # Strides are avoided when switching the first and second dim
+            # temporarily.
+            x = tl.transpose(x, 1, 2)
+
+            # classification
+            classes_x = x[:, 0:self.meta.n_confidences]
+            classes_x = tl.sigmoid(classes_x)
+
+            # regressions x
+            first_reg_feature = self.meta.n_confidences
+            regs_x = [
+                x[:, first_reg_feature + i * 2:first_reg_feature + (i + 1) * 2]
+                for i in range(self.meta.n_vectors)
+            ]
+            # regressions x: add index
+            index_field = index_field_torch(
+                (feature_height, feature_width), device=x.device, unsqueeze=(1, 0))
+            # TODO: coreml export does not work with the index_field creation in the graph.
+            index_field = tl.convert_to_tensor(index_field.numpy())
+            regs_x = [reg_x + index_field if do_offset else reg_x
+                      for reg_x, do_offset in zip(regs_x, self.meta.vector_offsets)]
+
+            # regressions logb
+            first_reglogb_feature = self.meta.n_confidences + self.meta.n_vectors * 2
+            single_reg_logb = x[:, first_reglogb_feature:first_reglogb_feature + 1]
+
+            # scale
+            first_scale_feature = self.meta.n_confidences + self.meta.n_vectors * 3
+            scales_x = x[:, first_scale_feature:first_scale_feature + self.meta.n_scales]
+            scales_x = torch.nn.functional.softplus(scales_x)
+
+            # concat with width in front (v4 style)
+            x = tl.concat([single_reg_logb, classes_x, *regs_x, scales_x], dim=1)
+
+            # TODO: CoreMLv4 problem (see above).
+            x = tl.transpose(x, 1, 2)
+
+        return x
+
+
+class CompositeField4(HeadNetwork):
+    dropout_p = 0.0
+    inplace_ops = True
+
+    def __init__(self,
+                 meta: headmeta.Base,
+                 in_features, *,
+                 kernel_size=1, padding=0, dilation=1):
+        super().__init__(meta, in_features)
+
+        LOG.debug('%s config: fields = %d, confidences = %d, vectors = %d, scales = %d '
+                  'kernel = %d, padding = %d, dilation = %d',
+                  meta.name, meta.n_fields, meta.n_confidences, meta.n_vectors, meta.n_scales,
+                  kernel_size, padding, dilation)
+
+        self.dropout = Dropout(p=self.dropout_p)
+
+        # convolution
+        self.n_components = 1 + meta.n_confidences + meta.n_vectors * 2 + meta.n_scales
+        self.conv = Conv2d(
+            meta.n_fields * self.n_components * (meta.upsample_stride ** 2),
+            (kernel_size,kernel_size), padding=(padding,padding), dilation=(dilation,dilation),in_channels= in_features
+        )
+
+        # upsample
+        assert meta.upsample_stride >= 1
+        self.upsample_op = None
+        if meta.upsample_stride > 1:
+            self.upsample_op = torch.nn.PixelShuffle(meta.upsample_stride)
+
+    @classmethod
+    def cli(cls, parser: argparse.ArgumentParser):
+        group = parser.add_argument_group('CompositeField4')
+        group.add_argument('--cf4-dropout', default=cls.dropout_p, type=float,
+                           help='[experimental] zeroing probability of feature in head input')
+        assert cls.inplace_ops
+        group.add_argument('--cf4-no-inplace-ops', dest='cf4_inplace_ops',
+                           default=True, action='store_false',
+                           help='alternative graph without inplace ops')
+
+    @classmethod
+    def configure(cls, args: argparse.Namespace):
+        cls.dropout_p = args.cf4_dropout
+        cls.inplace_ops = args.cf4_inplace_ops
+
+    @property
+    def sparse_task_parameters(self):
+        return [self.conv.weight]
+
+    def forward(self, x):  # pylint: disable=arguments-differ
+        x = self.dropout(x)
+        x = self.conv(x)
+        # upscale
+        if self.upsample_op is not None:
+            x = self.upsample_op(x)
+            low_cut = (self.meta.upsample_stride - 1) // 2
+            high_cut = math.ceil((self.meta.upsample_stride - 1) / 2.0)
+            if self.training:
+                # negative axes not supported by ONNX TensorRT
+                x = x[:, :, low_cut:-high_cut, low_cut:-high_cut]
+            else:
+                # the int() forces the tracer to use static shape
+                x = x[:, :, low_cut:int(x.shape[2]) - high_cut, low_cut:int(x.shape[3]) - high_cut]
+
+        # Extract some shape parameters once.
+        # Convert to int so that shape is constant in ONNX export.
+        x_size = x.size()
+        batch_size = x_size[0]
+        feature_height = int(x_size[2])
+        feature_width = int(x_size[3])
+
+        x = x.view(
+            batch_size,
+            self.meta.n_fields,
+            self.n_components,
+            feature_height,
+            feature_width
+        )
+
+        if not self.training and self.inplace_ops:
+            # classification
+            classes_x = x[:, :, 1:1 + self.meta.n_confidences]
+            tl.sigmoid(classes_x)
+
+            # regressions x: add index
+            if self.meta.n_vectors > 0:
+                index_field = index_field_torch((feature_height, feature_width), device=x.device)
+                first_reg_feature = 1 + self.meta.n_confidences
+                for i, do_offset in enumerate(self.meta.vector_offsets):
+                    if not do_offset:
+                        continue
+                    reg_x = x[:, :, first_reg_feature + i * 2:first_reg_feature + (i + 1) * 2]
+                    reg_x.add_(index_field)
+
+            # scale
+            first_scale_feature = 1 + self.meta.n_confidences + self.meta.n_vectors * 2
+            scales_x = x[:, :, first_scale_feature:first_scale_feature + self.meta.n_scales]
+            scales_x[:] = torch.nn.functional.softplus(scales_x)
+        elif not self.training and not self.inplace_ops:
+            # TODO: CoreMLv4 does not like strided slices.
+            # Strides are avoided when switching the first and second dim
+            # temporarily.
+            x = torch.transpose(x, 1, 2)
+
+            # width
+            width_x = x[:, 0:1]
+
+            # classification
+            classes_x = x[:, 1:1 + self.meta.n_confidences]
+            classes_x = torch.sigmoid(classes_x)
+
+            # regressions x
+            first_reg_feature = 1 + self.meta.n_confidences
+            regs_x = [
+                x[:, first_reg_feature + i * 2:first_reg_feature + (i + 1) * 2]
+                for i in range(self.meta.n_vectors)
+            ]
+            # regressions x: add index
+            index_field = index_field_torch(
+                (feature_height, feature_width), device=x.device, unsqueeze=(1, 0))
+            # TODO: coreml export does not work with the index_field creation in the graph.
+            index_field = tl.convert_to_tensor(index_field.numpy())
+            regs_x = [reg_x + index_field if do_offset else reg_x
+                      for reg_x, do_offset in zip(regs_x, self.meta.vector_offsets)]
+
+            # scale
+            first_scale_feature = 1 + self.meta.n_confidences + self.meta.n_vectors * 2
+            scales_x = x[:, first_scale_feature:first_scale_feature + self.meta.n_scales]
+            scales_x = torch.nn.functional.softplus(scales_x)
+
+            # concat
+            x = tl.concat([width_x, classes_x, *regs_x, scales_x], dim=1)
+
+            # TODO: CoreMLv4 problem (see above).
+            x = tl.transpose(x, 1, 2)
+
+        return x

tracking_heads.py

@@ -0,0 +1,105 @@
+import tensorlayer as tl
+from tensorlayer.layers import Conv2d
+from tensorlayer.layers import SequentialLayer
+
+
+from .heads import HeadNetwork, CompositeField4
+
+
+class TBaseSingleImage(HeadNetwork):
+    """Filter the feature map so that they can be used by single image loss.
+
+    Training: only apply loss to image 0 of an image pair of image 0 and 1.
+    Evaluation with forward tracking pose: only keep image 0.
+    Evaluation with full tracking pose: keep all but stack group along feature dim.
+    """
+    forward_tracking_pose = True
+    tracking_pose_length = 2
+
+    def __init__(self, meta, in_features):
+        super().__init__(meta, in_features)
+        self.head = CompositeField4(meta, in_features)
+
+    def forward(self, *args):
+        x = args[0]
+
+        if self.training:
+            x = x[::2]
+        elif self.forward_tracking_pose:
+            x = x[::self.tracking_pose_length]
+
+        x = self.head(x)
+
+        if not self.training and not self.forward_tracking_pose:
+            # full tracking pose eval
+            # TODO: stack batch dimension in feature dimension and adjust
+            # meta information (make it a property to dynamically return
+            # a different meta for evaluation)
+            raise NotImplementedError
+
+        return x
+
+
+class Tcaf(HeadNetwork):
+    """Filter the feature map so that they can be used by single image loss.
+
+    Training: only apply loss to image 0 of an image pair of image 0 and 1.
+    Evaluation with forward tracking pose: only keep image 0.
+    Evaluation with full tracking pose: keep all.
+    """
+    tracking_pose_length = 2
+    reduced_features = 512
+
+    _global_feature_reduction = None
+    _global_feature_compute = None
+
+    def __init__(self, meta, in_features):
+        super().__init__(meta, in_features)
+
+        if self._global_feature_reduction is None:
+            self.__class__._global_feature_reduction = SequentialLayer(
+                  [Conv2d(self.reduced_features,
+                                kernel_size=(1,1), bias=True,in_channels=in_features),
+                tl.ReLU(inplace=True)]
+            )
+        self.feature_reduction = self._global_feature_reduction
+
+        if self._global_feature_compute is None:
+            self.__class__._global_feature_compute = SequentialLayer(
+                [Conv2d(self.reduced_features * 2,kernel_size=(1,1) ,bias=True,in_channels=self.reduced_features * 2),
+                tl.ReLU(inplace=True)]
+            )
+        self.feature_compute = self._global_feature_compute
+
+        self.head = CompositeField4(meta, self.reduced_features * 2)
+
+    def forward(self, *args):
+        x = args[0]
+
+        # Batches that are not intended for tracking loss might have an
+        # odd number of images (or only 1 image).
+        # In that case, simply do not execute this head as the result should
+        # never be used.
+        if len(x) % 2 == 1:
+            return None
+
+        x = self.feature_reduction(x)
+
+        group_length = 2 if self.training else self.tracking_pose_length
+        primary = x[::group_length]
+        others = [x[i::group_length] for i in range(1, group_length)]
+
+        x = tl.stack([tl.concat([primary, o], dim=1) for o in others], dim=1)
+        x_shape = x.size()
+        x = tl.reshape(x, [x_shape[0] * x_shape[1]] + list(x_shape[2:]))
+
+        x = self.feature_compute(x)
+
+        x = self.head(x)
+
+        if self.tracking_pose_length != 2:
+            # TODO need to stack group from batch dim in feature dim and adjust
+            # meta info
+            raise NotImplementedError
+
+        return x