File size: 9,500 Bytes
f670afc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
# Copyright (C) 2021 NVIDIA CORPORATION & AFFILIATES.  All rights reserved.
#
# This work is made available under the Nvidia Source Code License-NC.
# To view a copy of this license, check out LICENSE.md
"""Utils for the pix2pixHD model."""
import numpy as np
import torch

from imaginaire.utils.data import get_paired_input_label_channel_number
from imaginaire.utils.distributed import dist_all_gather_tensor, is_master
from imaginaire.utils.distributed import master_only_print as print
from imaginaire.utils.trainer import (get_optimizer, get_optimizer_for_params,
                                      wrap_model_and_optimizer)
from sklearn.cluster import KMeans


def cluster_features(cfg, train_data_loader, net_E,
                     preprocess=None, small_ratio=0.0625, is_cityscapes=True):
    r"""Use clustering to compute the features.

    Args:
        cfg (obj): Global configuration file.
        train_data_loader (obj): Dataloader for iterate through the training
            set.
        net_E (nn.Module): Pytorch network.
        preprocess (function): Pre-processing function.
        small_ratio (float): We only consider instance that at least occupy
            $(small_ratio) amount of image space.
        is_cityscapes (bool): Is this is the cityscape dataset? In the
            Cityscapes dataset, the instance labels for car start with 26001,
            26002, ...

    Returns:
        ( num_labels x num_cluster_centers x feature_dims): cluster centers.
    """
    # Encode features.
    label_nc = get_paired_input_label_channel_number(cfg.data)
    feat_nc = cfg.gen.enc.num_feat_channels
    n_clusters = getattr(cfg.gen.enc, 'num_clusters', 10)
    # Compute features.
    features = {}
    for label in range(label_nc):
        features[label] = np.zeros((0, feat_nc + 1))
    for data in train_data_loader:
        if preprocess is not None:
            data = preprocess(data)
        feat = encode_features(net_E, feat_nc, label_nc,
                               data['images'], data['instance_maps'],
                               is_cityscapes)
        # We only collect the feature vectors for the master GPU.
        if is_master():
            for label in range(label_nc):
                features[label] = np.append(
                    features[label], feat[label], axis=0)
    # Clustering.
    # We only perform clustering for the master GPU.
    if is_master():
        for label in range(label_nc):
            feat = features[label]
            # We only consider segments that are greater than a pre-set
            # threshold.
            feat = feat[feat[:, -1] > small_ratio, :-1]
            if feat.shape[0]:
                n_clusters = min(feat.shape[0], n_clusters)
                kmeans = KMeans(n_clusters=n_clusters, random_state=0).fit(feat)
                n, d = kmeans.cluster_centers_.shape
                this_cluster = getattr(net_E, 'cluster_%d' % label)
                this_cluster[0:n, :] = torch.Tensor(
                    kmeans.cluster_centers_).float()


def encode_features(net_E, feat_nc, label_nc, image, inst,
                    is_cityscapes=True):
    r"""Compute feature embeddings for an image image.
    TODO(Ting-Chun): To make this funciton dataset independent.

    Args:
        net_E (nn.Module): The encoder network.
        feat_nc (int): Feature dimensions
        label_nc (int): Number of segmentation labels.
        image (tensor): Input image tensor.
        inst (tensor): Input instance map.
        is_cityscapes (bool): Is this is the cityscape dataset? In the
            Cityscapes dataset, the instance labels for car start with 26001,
            26002, ...
    Returns:
        (list of list of numpy vectors): We will have $(label_nc)
            list. For each list, it will record a list of feature vectors of
            dimension $(feat_nc+1) where the first $(feat_nc) dimensions is
            the representative feature of an instance and the last dimension
            is the proportion.
    """
    # h, w = inst.size()[2:]
    feat_map = net_E(image, inst)
    feature_map_gather = dist_all_gather_tensor(feat_map)
    inst_gathered = dist_all_gather_tensor(inst)
    # Initialize the cluster centers.
    # For each feature vector,
    #   0:feat_nc will be the feature vector.
    #   The feat_nc dimension record the percentage of the instance.
    feature = {}
    for i in range(label_nc):
        feature[i] = np.zeros((0, feat_nc + 1))
    if is_master():
        all_feat_map = torch.cat(feature_map_gather, 0)
        all_inst_map = torch.cat(inst_gathered, 0)
        # Scan through the batches.
        for n in range(all_feat_map.size()[0]):
            feat_map = all_feat_map[n:(n + 1), :, :, :]
            inst = all_inst_map[n:(n + 1), :, :, :]
            fh, fw = feat_map.size()[2:]
            inst_np = inst.cpu().numpy().astype(int)
            for i in np.unique(inst_np):
                if is_cityscapes:
                    label = i if i < 1000 else i // 1000
                else:
                    label = i
                idx = (inst == int(i)).nonzero()
                num = idx.size()[0]
                # We will just pick the middle pixel as its representative
                # feature.
                idx = idx[num // 2, :]
                val = np.zeros((1, feat_nc + 1))
                for k in range(feat_nc):
                    # We expect idx[0]=0 and idx[1]=0 as the number of sample
                    # per processing is 1 (idx[0]=0) and the channel number of
                    # the instance map is 1.
                    val[0, k] = feat_map[
                        idx[0], idx[1] + k, idx[2], idx[3]].item()
                val[0, feat_nc] = float(num) / (fh * fw)
                feature[label] = np.append(feature[label], val, axis=0)
        return feature
    else:
        return feature


def get_edges(t):
    r""" Compute edge maps for a given input instance map.

    Args:
        t (4D tensor): Input instance map.
    Returns:
        (4D tensor): Output edge map.
    """
    edge = torch.cuda.ByteTensor(t.size()).zero_()
    edge[:, :, :, 1:] = edge[:, :, :, 1:] | (
        t[:, :, :, 1:] != t[:, :, :, :-1]).byte()
    edge[:, :, :, :-1] = edge[:, :, :, :-1] | (
        t[:, :, :, 1:] != t[:, :, :, :-1]).byte()
    edge[:, :, 1:, :] = edge[:, :, 1:, :] | (
        t[:, :, 1:, :] != t[:, :, :-1, :]).byte()
    edge[:, :, :-1, :] = edge[:, :, :-1, :] | (
        t[:, :, 1:, :] != t[:, :, :-1, :]).byte()
    return edge.float()


def get_train_params(net, param_names_start_with=[], param_names_include=[]):
    r"""Get train parameters.

    Args:
        net (obj): Network object.
        param_names_start_with (list of strings): Params whose names
            start with any of the strings will be trained.
        param_names_include (list of strings): Params whose names include
            any of the strings will be trained.
    """
    params_to_train = []
    params_dict = net.state_dict()
    list_of_param_names_to_train = set()
    # Iterate through all params in the network and check if we need to
    # train it.
    for key, value in params_dict.items():
        do_train = False
        # If the param name starts with the target string (excluding
        # the 'module' part etc), we will train this param.
        key_s = key.replace('module.', '').replace('averaged_model.', '')
        for param_name in param_names_start_with:
            if key_s.startswith(param_name):
                do_train = True
                list_of_param_names_to_train.add(param_name)

        # Otherwise, if the param name includes the target string,
        # we will also train it.
        if not do_train:
            for param_name in param_names_include:
                if param_name in key_s:
                    do_train = True
                    full_param_name = \
                        key_s[:(key_s.find(param_name) + len(param_name))]
                    list_of_param_names_to_train.add(full_param_name)

        # If we decide to train the param, add it to the list to train.
        if do_train:
            module = net
            key_list = key.split('.')
            for k in key_list:
                module = getattr(module, k)
            params_to_train += [module]

    print('Training layers: ', sorted(list_of_param_names_to_train))
    return params_to_train


def get_optimizer_with_params(cfg, net_G, net_D, param_names_start_with=[],
                              param_names_include=[]):
    r"""Return the optimizer object.

    Args:
        cfg (obj): Global config.
        net_G (obj): Generator network.
        net_D (obj): Discriminator network.
        param_names_start_with (list of strings): Params whose names
            start with any of the strings will be trained.
        param_names_include (list of strings): Params whose names include
            any of the strings will be trained.
    """
    # If any of the param name lists is not empty, will only train
    # these params. Otherwise will train the entire network (all params).
    if param_names_start_with or param_names_include:
        params = get_train_params(net_G, param_names_start_with,
                                  param_names_include)
    else:
        params = net_G.parameters()

    opt_G = get_optimizer_for_params(cfg.gen_opt, params)
    opt_D = get_optimizer(cfg.dis_opt, net_D)
    return wrap_model_and_optimizer(cfg, net_G, net_D, opt_G, opt_D)