def resnet_50(input_image): arg_scope = resnet_v1.resnet_arg_scope() with slim.arg_scope(arg_scope): features, _ = resnet_v1.resnet_v1_50(input_image) # feature flatten features = tf.squeeze(features) return features
def top_feature_net(input, anchors, inds_inside, num_bases): stride=8 with tf.variable_scope("top_base") as sc: arg_scope = resnet_v1.resnet_arg_scope(weight_decay=0.0) with slim.arg_scope(arg_scope) : net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=8) #pdb.set_trace() block=end_points['top_base/resnet_v1_50/block4'] # block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='2') tf.summary.histogram('rpn_top_block', block) # tf.summary.histogram('rpn_top_block_weights', tf.get_collection('2/conv_weight')[0]) with tf.variable_scope('top') as scope: #up = upsample2d(block, factor = 2, has_bias=True, trainable=True, name='1') #up = block up = conv2d_bn_relu(block, num_kernels=128, kernel_size=(3,3), stride=[1,1,1,1], padding='SAME', name='2') scores = conv2d(up, num_kernels=2*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='score') probs = tf.nn.softmax( tf.reshape(scores,[-1,2]), name='prob') deltas = conv2d(up, num_kernels=4*num_bases, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='delta') #<todo> flip to train and test mode nms (e.g. different nms_pre_topn values): use tf.cond with tf.variable_scope('top-nms') as scope: #non-max batch_size, img_height, img_width, img_channel = input.get_shape().as_list() img_scale = 1 # pdb.set_trace() rois, roi_scores = tf_rpn_nms( probs, deltas, anchors, inds_inside, stride, img_width, img_height, img_scale, nms_thresh=0.7, min_size=stride, nms_pre_topn=300, nms_post_topn=50, name ='nms') #<todo> feature = upsample2d(block, factor = 4, ...) feature = block
def network_resnet_v1_50(): input_shape = [1, 224, 224, 3] input_ = tf.placeholder(dtype=tf.float32, name='input', shape=input_shape) net, _end_points = resnet_v1_50(input_, num_classes=1000, is_training=False) return net
def build_model(self, inp, mode, regularizer=None): net = inp['img'] training = (mode == tf.estimator.ModeKeys.TRAIN) with tf.variable_scope('encode'): with slim.arg_scope( resnet_v1.resnet_arg_scope( weight_decay=self.config_dict['ext'] ['encoder_l2_decay'])): net, _ = resnet_v1.resnet_v1_50(net, num_classes=None, is_training=training, global_pool=True) with tf.variable_scope('classify'): # net = tf.layers.max_pooling2d(net, net.shape.as_list()[1], 1) # net = tf.layers.conv2d(net, 1024, 1, kernel_regularizer=regularizer) net = tf.layers.conv2d(net, self.config_dict['label_cnt'], 1, kernel_regularizer=regularizer) logits = tf.squeeze(net, axis=(1, 2)) return logits
def res50(): image = tf.placeholder(tf.float32, [None, 224, 224, 3], 'image') with slim.arg_scope(resnet_arg_scope(is_training=False)): net_conv, end_point = resnet_v1.resnet_v1_50(image, global_pool=True, is_training=False) return net_conv, image
def run_model(total_gpu_num): """Train model.""" with epl.replicate(total_gpu_num): iterator = get_mock_iterator() images, labels = iterator.get_next() features = resnet_v1.resnet_v1_50(images, num_classes=None, is_training=True)[0] features = tf.squeeze(features, [1, 2]) with epl.split(total_gpu_num): logits = tf.layers.dense(features, class_num) loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits) global_step = tf.train.get_or_create_global_step() optimizer = tf.train.AdamOptimizer(learning_rate=0.9) train_op = optimizer.minimize(loss, global_step=global_step) hooks = [tf.train.StopAtStepHook(last_step=20)] with tf.train.MonitoredTrainingSession(hooks=hooks) as sess: while not sess.should_stop(): starttime = time.time() _, _, step = sess.run([loss, train_op, global_step]) endtime = time.time() tf.logging.info("[Iteration {} ], Time: {:.4} .".format( step, endtime - starttime)) tf.logging.info("[Finished]")
def getCNNFeatures(self, input_tensor, out_dim, fc_initializer): graph = tf.Graph() with graph.as_default(): with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, end_points = resnet_v1.resnet_v1_50(input_tensor, num_classes=None) model_path = os.path.join(self.checkpoints_dir, self.ckpt_name) init_fn = tf.contrib.framework.assign_from_checkpoint_fn( model_path, slim.get_model_variables('resnet_v1')) flattened = tf.reshape(end_points["resnet_v1_50/block4"], [-1, fc_dim]) print flattened.get_shape() with vs.variable_scope('fc_resnet'): W = vs.get_variable("W", [fc_dim, out_dim], initializer=fc_initializer) b = vs.get_variable("b", [out_dim], initializer=fc_initializer) output = tf.nn.relu(tf.matmul(flattened, W) + b) return init_fn, output #TEST: # cnn_f_extractor = CNN_FeatureExtractor() # inputt = tf.constant(np.arange(12288, dtype=np.float32), shape=[1, 64, 64, 3]) # inputfn, features = cnn_f_extractor.getCNNFeatures(inputt, 256, tf.contrib.layers.variance_scaling_initializer()) # print features.get_shape()
def tower_loss(scope): images, labels = read_and_decode() if net == 'vgg_16': with slim.arg_scope(vgg.vgg_arg_scope()): logits, end_points = vgg.vgg_16(images, num_classes=FLAGS.num_classes) elif net == 'vgg_19': with slim.arg_scope(vgg.vgg_arg_scope()): logits, end_points = vgg.vgg_19(images, num_classes=FLAGS.num_classes) elif net == 'resnet_v1_101': with slim.arg_scope(resnet_v1.resnet_arg_scope()): logits, end_points = resnet_v1.resnet_v1_101(images, num_classes=FLAGS.num_classes) logits = tf.reshape(logits, [FLAGS.batch_size, FLAGS.num_classes]) elif net == 'resnet_v1_50': with slim.arg_scope(resnet_v1.resnet_arg_scope()): logits, end_points = resnet_v1.resnet_v1_50(images, num_classes=FLAGS.num_classes) logits = tf.reshape(logits, [FLAGS.batch_size, FLAGS.num_classes]) elif net == 'resnet_v2_50': with slim.arg_scope(resnet_v2.resnet_arg_scope()): logits, end_points = resnet_v2.resnet_v2_50(images, num_classes=FLAGS.num_classes) logits = tf.reshape(logits, [FLAGS.batch_size, FLAGS.num_classes]) else: raise Exception('No network matched with net %s.' % net) assert logits.shape == (FLAGS.batch_size, FLAGS.num_classes) _ = cal_loss(logits, labels) losses = tf.get_collection('losses', scope) total_loss = tf.add_n(losses, name='total_loss') for l in losses + [total_loss]: loss_name = re.sub('%s_[0-9]*/' % TOWER_NAME, '', l.op.name) tf.summary.scalar(loss_name, l) return total_loss
def get_slim_resnet_v1_byname(net_name, inputs, num_classes=None, is_training=True, global_pool=True, output_stride=None, weight_decay=0.): if net_name == 'resnet_v1_50': with slim.arg_scope( resnet_v1.resnet_arg_scope(weight_decay=weight_decay)): logits, end_points = resnet_v1.resnet_v1_50( inputs=inputs, num_classes=num_classes, is_training=is_training, global_pool=global_pool, output_stride=output_stride, ) return logits, end_points if net_name == 'resnet_v1_101': with slim.arg_scope( resnet_v1.resnet_arg_scope(weight_decay=weight_decay)): logits, end_points = resnet_v1.resnet_v1_101( inputs=inputs, num_classes=num_classes, is_training=is_training, global_pool=global_pool, output_stride=output_stride, ) return logits, end_points
def build_graph(self): with slim.arg_scope(resnet_v1.resnet_arg_scope(weight_decay=1e-5)): logits, end_point = resnet_v1.resnet_v1_50( self.input, num_classes=self.num_classes, scope='resnet_v1_50') # logits [-1,1,1,dim] 全局池化 dim = logits.get_shape()[-1] assert dim == self.num_classes self.logits = tf.reshape(logits, [-1, dim])
def testing_network(self, image): _, endpoints = resnet_v1.resnet_v1_50(image, num_classes=702, is_training=False, global_pool=True, output_stride=None, reuse=tf.AUTO_REUSE, scope='resnet_v1_50') return endpoints
def network_test(self, inputs): net, end_points = resnet_v1.resnet_v1_50(inputs, num_classes=self.n_labels, is_training=False, global_pool=True, output_stride=None, reuse=tf.AUTO_REUSE, scope="resnet_v1_50") return net, end_points
def inference(self): x = tf.reshape(self.x, shape=[-1, self.input_shape[0], self.input_shape[1], self.input_shape[2]]) with slim.arg_scope(resnet_v1.resnet_arg_scope()): logits, end_points = resnet_v1.resnet_v1_50(x, num_classes=self.nclasses, is_training=self.is_training # , spatial_squeeze=True , global_pool=True ) # remove in the future if squeeze build in resnet_v1 function net = array_ops.squeeze(logits, [1,2], name='SpatialSqueeze') return net
def rgb_feature_net(input): arg_scope = resnet_v1.resnet_arg_scope(weight_decay=0.0) with slim.arg_scope(arg_scope): net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=8) block=end_points['resnet_v1_50/block4'] # block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='2') #<todo> feature = upsample2d(block, factor = 4, ...) tf.summary.histogram('rgb_top_block', block) feature = block return feature
def _vision(preprocessed_inputs, reuse=True): with tf.variable_scope("vision", reuse=reuse): with slim.arg_scope(resnet_v1.resnet_arg_scope()): resnet_output, _ = resnet_v1.resnet_v1_50( preprocessed_inputs, is_training=True) if not config["fine_tune_vision"]: resnet_output = tf.stop_gradient(resnet_output) resnet_output = tf.squeeze(resnet_output, axis=[1, 2]) resnet_output = tf.nn.dropout( resnet_output, keep_prob=self.vision_keep_prob_ph) vision_result = slim.fully_connected(resnet_output, num_hidden_hyper, activation_fn=None) return vision_result, resnet_output
def teacher(self, x, j): with slim.arg_scope(resnet_v1.resnet_arg_scope()): x = utils.nchw_to_nhwc(x) batch_out, batch_list = resnet_v1.resnet_v1_50(x, 1000, is_training=True) feature = batch_list['resnet_v1_50/block2/unit_4/bottleneck_v1/conv1'] self.init_fn_1 = slim.assign_from_checkpoint_fn( self.pre_dir + '/resnet_v1_50.ckpt', slim.get_model_variables('resnet_v1_50')) ''' del which has no gradient ''' # print(batch_list) x = utils.nhwc_to_nchw(feature) x, var = vnect(x, j) return x, var
def batch_pred(models_path, images_list, labels_nums, data_format): [batch_size, resize_height, resize_width, depths] = data_format input_images = tf.placeholder( dtype=tf.float32, shape=[None, resize_height, resize_width, depths], name='input') # model with slim.arg_scope(resnet_v1.resnet_arg_scope()): out, end_points = resnet_v1.resnet_v1_50(inputs=input_images, num_classes=labels_nums, is_training=False) out = tf.squeeze(out, [1, 2]) score = tf.nn.softmax(out, name='pre') class_id = tf.argmax(score, 1) gpu_options = tf.GPUOptions(allow_growth=False) with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess: sess.run(tf.global_variables_initializer()) saver = tf.train.Saver() saver.restore(sess, models_path) tot = len(images_list) for idx in range(0, tot, batch_size): images = list() idx_end = min(tot, idx + batch_size) print(idx) for i in range(idx, idx_end): image_path = images_list[i] image = open(image_path, 'rb').read() image = tf.image.decode_jpeg(image, channels=3) processed_image = preprocess_image(image, resize_height, resize_width) processed_image = sess.run(processed_image) # print("processed_image.shape", processed_image.shape) images.append(processed_image) images = np.array(images) start = time.time() sess.run([score, class_id], feed_dict={input_images: images}) end = time.time() print("time of batch {} is %f".format(batch_size) % (end - start)) sess.close()
def build(self): # Input self.input = tf.placeholder( dtype=tf.float32, shape=[None, self.img_size[0], self.img_size[1], self.img_size[2]]) self.input_mean = tfutils.mean_value(self.input, self.img_mean) if self.base_net == 'vgg16': with slim.arg_scope(vgg.vgg_arg_scope()): outputs, end_points = vgg.vgg_16(self.input_mean, self.num_classes) self.prob = tf.nn.softmax(outputs, -1) self.logits = outputs elif self.base_net == 'res50': with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, end_points = resnet_v1.resnet_v1_50( self.input_mean, self.num_classes, is_training=self.is_train) self.prob = tf.nn.softmax(net[:, 0, 0, :], -1) self.logits = net[:, 0, 0, :] elif self.base_net == 'res101': with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, end_points = resnet_v1.resnet_v1_101( self.input_mean, self.num_classes, is_training=self.is_train) self.prob = tf.nn.softmax(net[:, 0, 0, :], -1) self.logits = net[:, 0, 0, :] elif self.base_net == 'res152': with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, end_points = resnet_v1.resnet_v1_152( self.input_mean, self.num_classes, is_training=self.is_train) self.prob = tf.nn.softmax(net[:, 0, 0, :], -1) self.logits = net[:, 0, 0, :] else: raise ValueError( 'base network should be vgg16, res50, -101, -152...') self.gt = tf.placeholder(dtype=tf.int32, shape=[None]) # self.var_list = tf.trainable_variables() if self.is_train: self.loss()
def main(): tf.reset_default_graph() input_node = tf.placeholder(tf.float32, shape=(1, 224, 224, 3), name="input") print("input_node:", input_node) with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, _ = resnet_v1.resnet_v1_50(input_node, 1000, is_training=False) print("net:", net) saver = tf.train.Saver() with tf.Session() as sess: saver.restore(sess, model_path) tf.train.write_graph(sess.graph_def, './pb_model', 'model.pb') freeze_graph.freeze_graph('pb_model/model.pb', '', False, model_path, 'resnet_v1_50/logits/BiasAdd', 'save/restore_all', 'save/Const:0', 'pb_model/frozen_resnet_v1_50.pb', False, "") print("done")
def get_model(input_pls, is_training, bn=False, bn_decay=None, img_size=224, FLAGS=None): if FLAGS.act == "relu": activation_fn = tf.nn.relu elif FLAGS.act == "elu": activation_fn = tf.nn.elu input_imgs = input_pls['imgs'] input_pnts = input_pls['pnts'] input_gvfs = input_pls['gvfs'] input_onedge = input_pls['onedge'] input_trans_mat = input_pls['trans_mats'] input_obj_rot_mats = input_pls['obj_rot_mats'] batch_size = input_imgs.get_shape()[0].value # endpoints end_points = {} end_points['pnts'] = input_pnts if FLAGS.rot: end_points['gt_gvfs_xyz'] = tf.matmul(input_gvfs, input_obj_rot_mats) end_points['pnts_rot'] = tf.matmul(input_pnts, input_obj_rot_mats) else: end_points['gt_gvfs_xyz'] = input_gvfs #* 10 end_points['pnts_rot'] = input_pnts if FLAGS.edgeweight != 1.0: end_points['onedge'] = input_onedge input_pnts_rot = end_points['pnts_rot'] end_points['imgs'] = input_imgs # B*H*W*3|4 # Image extract features if input_imgs.shape[1] != img_size or input_imgs.shape[2] != img_size: if FLAGS.alpha: ref_img_rgb = tf.compat.v1.image.resize_bilinear( input_imgs[:, :, :, :3], [img_size, img_size]) ref_img_alpha = tf.image.resize_nearest_neighbor( tf.expand_dims(input_imgs[:, :, :, 3], axis=-1), [img_size, img_size]) ref_img = tf.concat([ref_img_rgb, ref_img_alpha], axis=-1) else: ref_img = tf.compat.v1.image.resize_bilinear( input_imgs, [img_size, img_size]) else: ref_img = input_imgs end_points['resized_ref_img'] = ref_img if FLAGS.encoder[:6] == "vgg_16": vgg.vgg_16.default_image_size = img_size with slim.arg_scope([slim.conv2d], weights_regularizer=slim.l2_regularizer(FLAGS.wd)): ref_feats_embedding, encdr_end_points = vgg.vgg_16( ref_img, num_classes=FLAGS.num_classes, is_training=False, scope='vgg_16', spatial_squeeze=False) elif FLAGS.encoder == "sim_res": ref_feats_embedding, encdr_end_points = res_sim_encoder.res_sim_encoder( ref_img, FLAGS.batch_size, is_training=is_training, activation_fn=activation_fn, bn=bn, bn_decay=bn_decay, wd=FLAGS.wd) elif FLAGS.encoder == "resnet_v1_50": resnet_v1.default_image_size = img_size with slim.arg_scope(resnet_v1.resnet_arg_scope()): ref_feats_embedding, encdr_end_points = resnet_v1.resnet_v1_50( ref_img, FLAGS.num_classes, is_training=is_training, scope='resnet_v1_50') scopelst = [ "resnet_v1_50/block1", "resnet_v1_50/block2", "resnet_v1_50/block3", 'resnet_v1_50/block4' ] elif FLAGS.encoder == "resnet_v1_101": resnet_v1.default_image_size = img_size with slim.arg_scope(resnet_v1.resnet_arg_scope()): ref_feats_embedding, encdr_end_points = resnet_v1.resnet_v1_101( ref_img, FLAGS.num_classes, is_training=is_training, scope='resnet_v1_101') scopelst = [ "resnet_v1_101/block1", "resnet_v1_101/block2", "resnet_v1_101/block3", 'resnet_v1_101/block4' ] elif FLAGS.encoder == "resnet_v2_50": resnet_v2.default_image_size = img_size with slim.arg_scope(resnet_v1.resnet_arg_scope()): ref_feats_embedding, encdr_end_points = resnet_v2.resnet_v2_50( ref_img, FLAGS.num_classes, is_training=is_training, scope='resnet_v2_50') scopelst = [ "resnet_v2_50/block1", "resnet_v2_50/block2", "resnet_v2_50/block3", 'resnet_v2_50/block4' ] elif FLAGS.encoder == "resnet_v2_101": resnet_v2.default_image_size = img_size with slim.arg_scope(resnet_v1.resnet_arg_scope()): ref_feats_embedding, encdr_end_points = resnet_v2.resnet_v2_101( ref_img, FLAGS.num_classes, is_training=is_training, scope='resnet_v2_101') scopelst = [ "resnet_v2_101/block1", "resnet_v2_101/block2", "resnet_v2_101/block3", 'resnet_v2_101/block4' ] end_points['img_embedding'] = ref_feats_embedding point_img_feat = None gvfs_feat = None sample_img_points = get_img_points(input_pnts, input_trans_mat) # B * N * 2 if FLAGS.img_feat_onestream: with tf.compat.v1.variable_scope("sdfimgfeat") as scope: if FLAGS.encoder[:3] == "vgg": conv1 = tf.compat.v1.image.resize_bilinear( encdr_end_points['vgg_16/conv1/conv1_2'], (FLAGS.img_h, FLAGS.img_w)) point_conv1 = tf.contrib.resampler.resampler( conv1, sample_img_points) conv2 = tf.compat.v1.image.resize_bilinear( encdr_end_points['vgg_16/conv2/conv2_2'], (FLAGS.img_h, FLAGS.img_w)) point_conv2 = tf.contrib.resampler.resampler( conv2, sample_img_points) conv3 = tf.compat.v1.image.resize_bilinear( encdr_end_points['vgg_16/conv3/conv3_3'], (FLAGS.img_h, FLAGS.img_w)) point_conv3 = tf.contrib.resampler.resampler( conv3, sample_img_points) if FLAGS.encoder[-7:] != "smaller": conv4 = tf.compat.v1.image.resize_bilinear( encdr_end_points['vgg_16/conv4/conv4_3'], (FLAGS.img_h, FLAGS.img_w)) point_conv4 = tf.contrib.resampler.resampler( conv4, sample_img_points) point_img_feat = tf.concat(axis=2, values=[ point_conv1, point_conv2, point_conv3, point_conv4 ]) # small else: print("smaller vgg") point_img_feat = tf.concat( axis=2, values=[point_conv1, point_conv2, point_conv3]) # small elif FLAGS.encoder[:3] == "res": # print(encdr_end_points.keys()) conv1 = tf.compat.v1.image.resize_bilinear( encdr_end_points[scopelst[0]], (FLAGS.img_h, FLAGS.img_w)) point_conv1 = tf.contrib.resampler.resampler( conv1, sample_img_points) conv2 = tf.compat.v1.image.resize_bilinear( encdr_end_points[scopelst[1]], (FLAGS.img_h, FLAGS.img_w)) point_conv2 = tf.contrib.resampler.resampler( conv2, sample_img_points) conv3 = tf.compat.v1.image.resize_bilinear( encdr_end_points[scopelst[2]], (FLAGS.img_h, FLAGS.img_w)) point_conv3 = tf.contrib.resampler.resampler( conv3, sample_img_points) # conv4 = tf.compat.v1.image.resize_bilinear(encdr_end_points[scopelst[3]], (FLAGS.img_h, FLAGS.img_w)) # point_conv4 = tf.contrib.resampler.resampler(conv4, sample_img_points) point_img_feat = tf.concat( axis=2, values=[point_conv1, point_conv2, point_conv3]) else: conv1 = tf.compat.v1.image.resize_bilinear( encdr_end_points[0], (FLAGS.img_h, FLAGS.img_w)) point_conv1 = tf.contrib.resampler.resampler( conv1, sample_img_points) conv2 = tf.compat.v1.image.resize_bilinear( encdr_end_points[1], (FLAGS.img_h, FLAGS.img_w)) point_conv2 = tf.contrib.resampler.resampler( conv2, sample_img_points) conv3 = tf.compat.v1.image.resize_bilinear( encdr_end_points[2], (FLAGS.img_h, FLAGS.img_w)) point_conv3 = tf.contrib.resampler.resampler( conv3, sample_img_points) # conv4 = tf.compat.v1.image.resize_bilinear(encdr_end_points[scopelst[3]], (FLAGS.img_h, FLAGS.img_w)) # point_conv4 = tf.contrib.resampler.resampler(conv4, sample_img_points) point_img_feat = tf.concat( axis=2, values=[point_conv1, point_conv2, point_conv3]) print("point_img_feat.shape", point_img_feat.get_shape()) point_img_feat = tf.expand_dims(point_img_feat, axis=2) if FLAGS.decoder == "att": gvfs_feat = gvfnet.get_gvf_att_imgfeat( input_pnts_rot, ref_feats_embedding, point_img_feat, is_training, batch_size, bn, bn_decay, wd=FLAGS.wd, activation_fn=activation_fn) elif FLAGS.decoder == "skip": gvfs_feat = gvfnet.get_gvf_basic_imgfeat_onestream_skip( input_pnts_rot, ref_feats_embedding, point_img_feat, is_training, batch_size, bn, bn_decay, wd=FLAGS.wd, activation_fn=activation_fn) else: gvfs_feat = gvfnet.get_gvf_basic_imgfeat_onestream( input_pnts_rot, ref_feats_embedding, point_img_feat, is_training, batch_size, bn, bn_decay, wd=FLAGS.wd, activation_fn=activation_fn) else: if not FLAGS.multi_view: with tf.compat.v1.variable_scope("sdfprediction") as scope: gvfs_feat = gvfnet.get_gvf_basic(input_pnts_rot, ref_feats_embedding, is_training, batch_size, bn, bn_decay, wd=FLAGS.wd, activation_fn=activation_fn) end_points['pred_gvfs_xyz'], end_points['pred_gvfs_dist'], end_points[ 'pred_gvfs_direction'] = None, None, None if FLAGS.XYZ: end_points['pred_gvfs_xyz'] = gvfnet.xyz_gvfhead( gvfs_feat, batch_size, wd=FLAGS.wd, activation_fn=activation_fn) end_points['pred_gvfs_dist'] = tf.sqrt( tf.reduce_sum(tf.square(end_points['pred_gvfs_xyz']), axis=2, keepdims=True)) end_points[ 'pred_gvfs_direction'] = end_points['pred_gvfs_xyz'] / tf.maximum( end_points['pred_gvfs_dist'], 1e-6) else: end_points['pred_gvfs_dist'], end_points[ 'pred_gvfs_direction'] = gvfnet.dist_direct_gvfhead( gvfs_feat, batch_size, wd=FLAGS.wd, activation_fn=activation_fn) end_points['pred_gvfs_xyz'] = end_points[ 'pred_gvfs_direction'] * end_points['pred_gvfs_dist'] end_points["sample_img_points"] = sample_img_points # end_points["ref_feats_embedding"] = ref_feats_embedding end_points["point_img_feat"] = point_img_feat return end_points
parser.add_argument('--weights', default="model.cktp", type=str) # define the model path parser.add_argument('--weight_dir', default='./Affwild_models/standard_ResNet/', type=str) # define the model path parser.add_argument('--input_file', default='video_T_01.csv', type=str) # define the input image path parser.add_argument('--save_file', default='video_T_01.mat', type=str) # define the path to save extracted features args = parser.parse_args() images_batch = tf.placeholder(tf.float32, [1, 96, 96, 3]) with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, end_point = resnet_v1.resnet_v1_50(inputs=images_batch, is_training=False, num_classes=None) net = tf.squeeze(net, [1, 2]) saver = tf.train.Saver() sess = tf.Session() weight_file = os.path.join(args.weight_dir, args.weights) saver.restore(sess, weight_file) files = pd.read_csv(args.input_file) files = files.values feature_list = [] for file_path in tqdm(files): file_path = file_path[0].strip() image = cv2.imread(file_path) inputs = cv2.resize(image, (96, 96)) inputs = cv2.cvtColor(inputs, cv2.COLOR_BGR2RGB).astype(np.float32) inputs -= 128.0
def model_fn(features, labels, mode, params): # Download the pretrained model bucket_name = params['bucket_name'] prefix_name = params['prefix_name'] s3 = boto3.resource('s3') try: s3.Bucket(bucket_name).download_file(prefix_name, 'resnet.ckpt') print("Pretrained model is downloaded.") except botocore.exceptions.ClientError as e: if e.response['Error']['Code'] == "404": print("The object does not exist.") else: raise # Input Layer input_layer = tf.reshape(features[INPUT_TENSOR_NAME], [-1, 32, 32, 3]) # Load Pretrained model from tensorflow.contrib.slim.python.slim.nets.resnet_v1 import resnet_v1_50 last_layer = resnet_v1_50(input_layer, num_classes=None, scope='resnet_v1_50') variables_to_restore = tf.contrib.slim.get_variables_to_restore() tf.train.init_from_checkpoint("./resnet.ckpt",{v.name.split(':')[0]: v for v in variables_to_restore if not 'biases' in v.name}) logits = tf.reshape(tf.layers.dense(inputs=last_layer[0], units=100), [-1, 100]) # Define operations if mode in (Modes.PREDICT, Modes.EVAL): predicted_indices = tf.argmax(input=logits, axis=1) probabilities = tf.nn.softmax(logits, name='softmax_tensor') if mode in (Modes.TRAIN, Modes.EVAL): global_step = tf.train.get_or_create_global_step() label_indices = tf.cast(labels, tf.int32) loss = tf.losses.softmax_cross_entropy( onehot_labels=tf.one_hot(label_indices, depth=100), logits=logits) tf.summary.scalar('OptimizeLoss', loss) if mode == Modes.PREDICT: predictions = { 'classes': predicted_indices, 'probabilities': probabilities } export_outputs = { SIGNATURE_NAME: tf.estimator.export.PredictOutput(predictions) } return tf.estimator.EstimatorSpec( mode, predictions=predictions, export_outputs=export_outputs) if mode == Modes.TRAIN: logging_hook = tf.train.LoggingTensorHook({"loss" : loss}, every_n_iter=10) optimizer = tf.train.AdamOptimizer(learning_rate=0.001) train_op = optimizer.minimize(loss, global_step=global_step) return tf.estimator.EstimatorSpec(mode, loss=loss, train_op=train_op, training_hooks = [logging_hook]) if mode == Modes.EVAL: eval_metric_ops = { 'accuracy': tf.metrics.accuracy(label_indices, predicted_indices) } return tf.estimator.EstimatorSpec( mode, loss=loss, eval_metric_ops=eval_metric_ops)
def evaluate(): g = tf.Graph() with g.as_default(): image_list, label_list = data_process.read_labeled_image_list( FLAGS.input_file) # split into sequences, note: in the cnn models case this is splitting into batches of length: seq_length ; # for the cnn-rnn models case, I do not check whether the images in a sequence are consecutive or the images are from the same video/the images are displaying the same person image_list, label_list = data_process.make_rnn_input_per_seq_length_size( image_list, label_list, FLAGS.seq_length) images = tf.convert_to_tensor(image_list) labels = tf.convert_to_tensor(label_list) # Makes an input queue input_queue = tf.train.slice_input_producer([images, labels, images], num_epochs=None, shuffle=False, seed=None, capacity=1000, shared_name=None, name=None) images_batch, labels_batch, image_locations_batch = data_process.decodeRGB( input_queue, FLAGS.seq_length, FLAGS.size) images_batch = tf.to_float(images_batch) images_batch -= 128.0 images_batch /= 128.0 # scale all pixel values in range: [-1,1] images_batch = tf.reshape(images_batch, [-1, 96, 96, 3]) labels_batch = tf.reshape(labels_batch, [-1, 2]) if FLAGS.network == 'vggface_4096': from vggface import vggface_4096x4096x2 as net network = net.VGGFace(FLAGS.batch_size * FLAGS.seq_length) network.setup(images_batch) prediction = network.get_output() elif FLAGS.network == 'vggface_2000': from vggface import vggface_4096x2000x2 as net network = net.VGGFace(FLAGS.batch_size * FLAGS.seq_length) network.setup(images_batch) prediction = network.get_output() elif FLAGS.network == 'affwildnet_resnet': from tensorflow.contrib.slim.python.slim.nets import resnet_v1 with slim.arg_scope(resnet_v1.resnet_arg_scope()): net, _ = resnet_v1.resnet_v1_50(inputs=images_batch, is_training=False, num_classes=None) with tf.variable_scope('rnn') as scope: cnn = tf.reshape( net, [FLAGS.batch_size, FLAGS.sequence_length, -1]) cell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.GRUCell(128) for _ in range(2)]) outputs, _ = tf.nn.dynamic_rnn(cell, cnn, dtype=tf.float32) outputs = tf.reshape( outputs, (FLAGS.batch_size * FLAGS.sequence_length, 128)) weights_initializer = tf.truncated_normal_initializer( stddev=0.01) weights = tf.get_variable('weights_output', shape=[128, 2], initializer=weights_initializer, trainable=True) biases = tf.get_variable('biases_output', shape=[2], initializer=tf.zeros_initializer, trainable=True) prediction = tf.nn.xw_plus_b(outputs, weights, biases) elif FLAGS.network == 'affwildnet_vggface': from affwildnet import vggface_gru as net network = net.VGGFace(FLAGS.batch_size, FLAGS.seq_length) network.setup(images_batch) prediction = network.get_output() num_batches = int(len(image_list) / FLAGS.batch_size) variables_to_restore = tf.global_variables() with tf.Session() as sess: init_fn = slim.assign_from_checkpoint_fn( FLAGS.pretrained_model_checkpoint_path, variables_to_restore, ignore_missing_vars=False) init_fn(sess) print('Loading model {}'.format( FLAGS.pretrained_model_checkpoint_path)) tf.train.start_queue_runners(sess=sess) coord = tf.train.Coordinator() evaluated_predictions = [] evaluated_labels = [] images = [] try: for _ in range(num_batches): pr, l, imm = sess.run( [prediction, labels_batch, image_locations_batch]) evaluated_predictions.append(pr) evaluated_labels.append(l) images.append(imm) if coord.should_stop(): break coord.request_stop() except Exception as e: coord.request_stop(e) predictions = np.reshape(evaluated_predictions, (-1, 2)) labels = np.reshape(evaluated_labels, (-1, 2)) images = np.reshape(images, (-1)) conc_arousal = concordance_cc2(predictions[:, 1], labels[:, 1]) conc_valence = concordance_cc2(predictions[:, 0], labels[:, 0]) for i in range(len(predictions)): print("Labels: ", labels[i], "Predictions: ", predictions[i], "Error: ", (abs(labels[i] - predictions[i]))) print( "------------------------------------------------------------------------------" ) print('Concordance on valence : {}'.format(conc_valence)) print('Concordance on arousal : {}'.format(conc_arousal)) print('Concordance on total : {}'.format( (conc_arousal + conc_valence) / 2)) mse_arousal = sum( (predictions[:, 1] - labels[:, 1])**2) / len(labels[:, 1]) print('MSE Arousal : {}'.format(mse_arousal)) mse_valence = sum( (predictions[:, 0] - labels[:, 0])**2) / len(labels[:, 0]) print('MSE Valence : {}'.format(mse_valence)) return conc_valence, conc_arousal, ( conc_arousal + conc_valence) / 2, mse_arousal, mse_valence
def get_featuremap(net_name, input, num_classes=None): ''' #tensorlayer input = tl.layers.InputLayer(input) if net_name == 'resnet_v1_50': with slim.arg_scope(resnet_v1.resnet_arg_scope(weight_decay=cfg.FEATURE_WEIGHT_DECAY)): featuremap = tl.layers.SlimNetsLayer(prev_layer=input, slim_layer=resnet_v1.resnet_v1_50, slim_args={ 'num_classes': num_classes, 'is_training': True, 'global_pool': False }, name='resnet_v1_50' ) sv = tf.train.Supervisor() with sv.managed_session() as sess: a = sess.run(featuremap.all_layers) print(a) feature_w_loss = tf.reduce_sum(slim.losses.get_regularization_losses()) return featuremap.outputs, feature_w_loss, featuremap.all_params if net_name == 'resnet_v1_101': with slim.arg_scope(resnet_v1.resnet_arg_scope()): featuremap = tl.layers.SlimNetsLayer(prev_layer=input, slim_layer=resnet_v1.resnet_v1_101, slim_args={ 'num_classes': num_classes, 'is_training': True, 'global_pool': False }, name='resnet_v1_101' ) feature_w_loss = tf.reduce_sum(slim.losses.get_regularization_losses()) return featuremap.outputs, feature_w_loss, featuremap.all_params if net_name == 'resnet_v1_152': with slim.arg_scope(resnet_v1.resnet_arg_scope()): featuremap = tl.layers.SlimNetsLayer(prev_layer=input, slim_layer=resnet_v1.resnet_v1_152, slim_args={ 'num_classes': num_classes, 'is_training': True, 'global_pool': False }, name='resnet_v1_152' ) feature_w_loss = tf.reduce_sum(slim.losses.get_regularization_losses()) return featuremap.outputs, feature_w_loss, featuremap.all_params if net_name == 'vgg16': with slim.arg_scope(vgg.vgg_arg_scope()): featuremap = tl.layers.SlimNetsLayer(prev_layer=input, slim_layer=vgg.vgg_16, slim_args={ 'num_classes': num_classes, 'is_training': True, 'spatial_squeeze': False }, name='vgg_16' ) feature_w_loss = tf.reduce_sum(slim.losses.get_regularization_losses()) return featuremap.outputs, feature_w_loss, featuremap.all_params ''' #slim if net_name == 'resnet_v1_50': with slim.arg_scope( resnet_v1.resnet_arg_scope( weight_decay=cfg.FEATURE_WEIGHT_DECAY)): featuremap, layer_dic = resnet_v1.resnet_v1_50( inputs=input, num_classes=num_classes, is_training=False, global_pool=False) if cfg.USE_FPN: feature_maps_dict = { 'C2': layer_dic[ 'resnet_v1_50/block1/unit_2/bottleneck_v1'], # [56, 56] 'C3': layer_dic[ 'resnet_v1_50/block2/unit_3/bottleneck_v1'], # [28, 28] 'C4': layer_dic[ 'resnet_v1_50/block3/unit_5/bottleneck_v1'], # [14, 14] 'C5': layer_dic['resnet_v1_50/block4'] # [7, 7] } return feature_maps_dict return layer_dic['resnet_v1_50/block3/unit_5/bottleneck_v1'] #return featuremap if net_name == 'resnet_v1_101': with slim.arg_scope( resnet_v1.resnet_arg_scope( weight_decay=cfg.FEATURE_WEIGHT_DECAY)): featuremap, layer_dic = resnet_v1.resnet_v1_101( inputs=input, num_classes=num_classes, is_training=True, global_pool=False) if cfg.USE_FPN: feature_maps_dict = { 'C2': layer_dic[ 'resnet_v1_101/block1/unit_2/bottleneck_v1'], # [56, 56] 'C3': layer_dic[ 'resnet_v1_101/block2/unit_3/bottleneck_v1'], # [28, 28] 'C4': layer_dic[ 'resnet_v1_101/block3/unit_22/bottleneck_v1'], # [14, 14] 'C5': layer_dic['resnet_v1_101/block4'] } return feature_maps_dict return featuremap if net_name == 'vgg_16': with slim.arg_scope( resnet_v1.resnet_arg_scope( weight_decay=cfg.FEATURE_WEIGHT_DECAY)): featuremap, layer_dic = vgg.vgg_16( inputs=input, num_classes=7, is_training=False, spatial_squeeze=False, ) return layer_dic['vgg_16/conv5/conv5_3']
def main(_): with tf.name_scope('input_placeholder'): mv_placeholder = tf.placeholder(tf.float32, shape=(None, FLAGS.num_segments, 224, 224, 3 ), name = 'mv_frame') flow_placeholder = tf.placeholder(tf.float32, shape=(None, FLAGS.num_segments, 224, 224, 3 ), name = 'flow_frame') i_placeholder = tf.placeholder(tf.float32, shape=(None, FLAGS.num_segments, 224, 224, 3 ), name = 'i_frame') r_placeholder = tf.placeholder(tf.float32, shape=(None, FLAGS.num_segments, 224, 224, 3 ), name = 'r_frame') with tf.name_scope('label_placeholder'): label_placeholder = tf.placeholder(tf.int32, shape=(None), name = 'labels') with tf.name_scope('accuracy'): combine_value_ = tf.placeholder(tf.float32, shape=(), name = 'accuracy') i_value_ = tf.placeholder(tf.float32, shape=(), name = 'accuracy') mv_value_ = tf.placeholder(tf.float32, shape=(), name = 'accuracy') r_value_ = tf.placeholder(tf.float32, shape=(), name = 'accuracy') tf.summary.scalar('combine_acc', combine_value_) tf.summary.scalar('i_acc', i_value_) tf.summary.scalar('mv_acc', mv_value_) tf.summary.scalar('r_acc', r_value_) print('Finish placeholder.') with tf.name_scope('flatten_input'): b_size = tf.shape(mv_placeholder)[0] flat_mv = tf.reshape(mv_placeholder, [b_size * FLAGS.num_segments, 224, 224, 3]) # Since we have mulitple segments in a single video flat_flow = tf.reshape(flow_placeholder, [b_size * FLAGS.num_segments, 224, 224, 3]) flat_i = tf.reshape(i_placeholder, [b_size * FLAGS.num_segments, 224, 224, 3]) flat_r = tf.reshape(r_placeholder, [b_size * FLAGS.num_segments, 224, 224, 3]) with tf.variable_scope('fc_var') as var_scope: mv_weights = { 'w1': _variable_with_weight_decay('wmv1', [2048 , 512 ], 0.0005), 'w2': _variable_with_weight_decay('wmv2', [512 , N_CLASS], 0.0005) } mv_biases = { 'b1': _variable_with_weight_decay('bmv1', [ 512 ], 0.00), 'b2': _variable_with_weight_decay('bmv2', [ N_CLASS ], 0.00) } i_weights = { 'w1': _variable_with_weight_decay('wi1', [2048 , 512 ], 0.0005), 'w2': _variable_with_weight_decay('wi2', [512 , N_CLASS], 0.0005) } i_biases = { 'b1': _variable_with_weight_decay('bi1', [ 512 ], 0.00), 'b2': _variable_with_weight_decay('bi2', [ N_CLASS ], 0.00) } r_weights = { 'w1': _variable_with_weight_decay('wr1', [2048 , 512 ], 0.0005), 'w2': _variable_with_weight_decay('wr2', [512 , N_CLASS], 0.0005) } r_biases = { 'b1': _variable_with_weight_decay('br1', [ 512 ], 0.00), 'b2': _variable_with_weight_decay('br2', [ N_CLASS ], 0.00) } with tf.variable_scope('fusion_var'): fusion = tf.get_variable('fusion', [3], initializer=tf.contrib.layers.xavier_initializer()) print('Finish Flatten.') with tf.device('/gpu:0'): with tf.name_scope('FLMG'): mv_res = tf.concat([flat_mv, flat_r], axis = -1) mv = slim.conv2d(mv_res, 8, kernel_size=[3, 3], scope = 'FLMG_1') mv = slim.conv2d(mv, 8, kernel_size=[3, 3], scope = 'FLMG_2') mv = slim.conv2d(mv, 6, kernel_size=[3, 3], scope = 'FLMG_3') mv = slim.conv2d(mv, 4, kernel_size=[3, 3], scope = 'FLMG_4') mv = slim.conv2d(mv, 2, kernel_size=[3, 3], scope = 'FLMG_5') mv = slim.conv2d(mv, 3, kernel_size=[3, 3], scope = 'FLMG_6') with tf.name_scope('FLMG_LOSS'): # The cost function -- l2 mse matrix_pow_2 = tf.pow(tf.subtract(mv, flat_flow), 2) matrix_norm = tf.reduce_sum(matrix_pow_2, axis = [1,2,3]) flmg_loss = tf.reduce_mean(matrix_norm) tf.summary.scalar('flmg_loss', flmg_loss) with slim.arg_scope(resnet_v1.resnet_arg_scope()): i_feature, _ = resnet_v1.resnet_v1_152(flat_i, num_classes=None, is_training=True, scope='i_resnet') mv_feature, _ = resnet_v1.resnet_v1_50(mv, num_classes=None, is_training=True, scope='mv_resnet') r_feature, _ = resnet_v1.resnet_v1_50(flat_r, num_classes=None, is_training=True, scope='r_resnet') with tf.name_scope('reshape_feature'): i_feature = tf.reshape(i_feature, [-1, 2048]) mv_feature = tf.reshape(mv_feature, [-1, 2048]) r_feature = tf.reshape(r_feature, [-1, 2048]) with tf.name_scope('inference_model'): i_sc, i_pred = model.inference_feature (i_feature, i_weights, i_biases, FLAGS.num_segments, N_CLASS, name = 'i_inf') mv_sc, mv_pred = model.inference_feature (mv_feature, mv_weights, mv_biases, FLAGS.num_segments, N_CLASS, name = 'mv_inf') r_sc, r_pred = model.inference_feature (r_feature, r_weights, r_biases, FLAGS.num_segments, N_CLASS, name = 'r_inf') combine_sc, pred_class = model.inference_fusion ( i_sc, mv_sc, r_sc, fusion) print('Finish Model.') with tf.name_scope('classiciation_loss'): one_hot_labels = tf.one_hot(label_placeholder, N_CLASS) mv_class_loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits = mv_sc, labels = one_hot_labels, dim=1)) i_class_loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits = i_sc, labels = one_hot_labels, dim=1)) r_class_loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits = r_sc, labels = one_hot_labels, dim=1)) tf.summary.scalar('mv_cls_loss', mv_class_loss) tf.summary.scalar('i_cls_loss', i_class_loss) tf.summary.scalar('r_cls_loss', r_class_loss) combine_loss = tf.reduce_sum(tf.nn.softmax_cross_entropy_with_logits(logits = combine_sc, labels = one_hot_labels, dim=1)) tf.summary.scalar('fuse_cls_loss', combine_loss) total_loss = combine_loss + i_class_loss + mv_class_loss + r_class_loss + flmg_loss tf.summary.scalar('tot_cls_loss', total_loss) with tf.name_scope('weigh_decay'): weight_loss = sum(tf.get_collection('losses')) tf.summary.scalar('eight_decay_loss', weight_loss) ''' with tf.name_scope('training_var_list'): mv_variable_list = list ( set(mv_weights.values()) | set(mv_biases.values()) ) mv_resnet_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='mv_resnet') i_variable_list = list ( set(i_weights.values()) | set(i_biases.values()) ) i_resnet_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='i_resnet') r_variable_list = list ( set(r_weights.values()) | set(r_biases.values()) ) r_resnet_variables = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='r_resnet') with tf.name_scope('summary_var'): _variable_summaries(mv_weights['w1']) _variable_summaries(i_weights['w2']) _variable_summaries(r_weights['w2']) _variable_summaries(mv_resnet_variables[0]) _variable_summaries(i_resnet_variables[0]) _variable_summaries(r_resnet_variables[0]) _variable_summaries(fusion) print('Finish variables.') ''' with tf.name_scope('optimizer'): ''' mv_fc_opt = tf.train.AdamOptimizer(FLAGS.mv_lr).minimize(mv_class_loss + weight_loss, var_list = mv_variable_list) mv_res_opt = tf.train.AdamOptimizer(FLAGS.resnet_lr).minimize(mv_class_loss, var_list = mv_resnet_variables) i_fc_opt = tf.train.AdamOptimizer(FLAGS.i_lr).minimize(i_class_loss + weight_loss, var_list = i_variable_list) i_res_opt = tf.train.AdamOptimizer(FLAGS.resnet_lr).minimize(i_class_loss, var_list = i_resnet_variables) r_fc_opt = tf.train.AdamOptimizer(FLAGS.r_lr).minimize(r_class_loss + weight_loss, var_list = r_variable_list) r_res_opt = tf.train.AdamOptimizer(FLAGS.resnet_lr).minimize(r_class_loss, var_list = r_resnet_variables) fusion_opt = tf.train.GradientDescentOptimizer(10e-6).minimize(combine_loss, var_list = fusion) ''' train_opt = tf.train.AdamOptimizer(FLAGS.tot_lr).minimize(total_loss, var_list = tf.trainable_variables()) print('Finish Optimizer.') with tf.name_scope('init_function'): init_var = tf.global_variables_initializer() with tf.name_scope('video_dataset'): train_data = dataset.buildTrainDataset_v2(FLAGS.train_list, FLAGS.data_path, FLAGS.num_segments, batch_size = FLAGS.batch_size, augment = False, shuffle = True, num_threads=2, buffer=100) test_data = dataset.buildTestDataset(FLAGS.valid_list, FLAGS.data_path, FLAGS.num_segments, batch_size = FLAGS.batch_size, num_threads = 2, buffer = 30) with tf.name_scope('dataset_iterator'): it = tf.contrib.data.Iterator.from_structure(train_data.output_types, train_data.output_shapes) next_data = it.get_next() init_data = it.make_initializer(train_data) it_test = tf.contrib.data.Iterator.from_structure(test_data.output_types, test_data.output_shapes) next_test_data = it_test.get_next() init_test_data = it_test.make_initializer(test_data) print('Finish Dataset.') restore_var = [v for v in tf.trainable_variables() if ('Adam' not in v.name)] first_restore_var = [v for v in tf.trainable_variables() if ('Adam' not in v.name and 'FLMG' not in v.name)] first_saver = tf.train.Saver(var_list=first_restore_var) my_saver = tf.train.Saver(var_list=restore_var, max_to_keep=5) config=tf.ConfigProto(allow_soft_placement=True, log_device_placement=False) sess = tf.Session(config=config) with tf.name_scope('writer'): merged = tf.summary.merge_all() if not tf.gfile.Exists(FLAGS.log_path): tf.gfile.MakeDirs(FLAGS.log_path) previous_runs = os.listdir(FLAGS.log_path) if len(previous_runs) == 0: run_number = 1 else: run_number = len(previous_runs) + 1 logdir = 'run_%02d' % run_number tf.gfile.MakeDirs(os.path.join(FLAGS.log_path, logdir)) writer = tf.summary.FileWriter(os.path.join(FLAGS.log_path, logdir), sess.graph) with tf.name_scope('saver'): if not tf.gfile.Exists(FLAGS.save_path): tf.gfile.MakeDirs(FLAGS.save_path) ''' i_saver = tf.train.Saver(i_variable_list) mv_saver = tf.train.Saver(mv_variable_list) r_saver = tf.train.Saver(r_variable_list) i_resnet_saver = tf.train.Saver(i_resnet_variables) mv_resnet_saver = tf.train.Saver(mv_resnet_variables) r_resnet_saver = tf.train.Saver(r_resnet_variables) ''' with tf.name_scope('intialization'): sess.run(init_var) sess.run(init_data) sess.run(init_test_data) #init_i_resent (sess) #init_mv_resent (sess) #init_r_resent(sess) ''' i_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'i_model.chkp'+FLAGS.steps)) mv_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'mv_model.chkp'+FLAGS.steps)) r_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'r_model.chkp'+FLAGS.steps)) i_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'i_resnet.chkp'+FLAGS.steps)) mv_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'mv_resnet.chkp'+FLAGS.steps)) r_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'r_resnet.chkp'+FLAGS.steps)) ''' try: my_saver.restore(sess, FLAGS.continue_training) except: # First train first_saver.restore(sess, FLAGS.continue_training) ''' i_resnet_saver = tf.train.Saver(i_resnet_variables) mv_resnet_saver = tf.train.Saver(mv_resnet_variables) r_resnet_saver = tf.train.Saver(r_resnet_variables) i_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'i_resnet.chkp'+FLAGS.steps)) mv_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'mv_resnet.chkp'+FLAGS.steps)) r_resnet_saver.restore(sess, os.path.join(FLAGS.saved_model_path, 'r_resnet.chkp'+FLAGS.steps)) ''' print('Finish Loading Pretrained Model.') ''' Main training loop ''' combine_acc = 0 i_acc = 0 mv_acc = 0 r_acc = 0 start_time = time.time() for step in range(FLAGS.max_steps): # Validation if (step) % 1000 == 0 and step > 0: combine_classes = [] mv_classes = [] i_classes = [] r_classes = [] gt_label = [] for i in range(100): ti_arr, tmv_arr, tr_arr, tlabel = sess.run(next_test_data) i_class, mv_class, r_class, com_class = sess.run([i_pred, mv_pred, r_pred, pred_class], feed_dict={mv_placeholder: tmv_arr, i_placeholder: ti_arr, r_placeholder: tr_arr , label_placeholder : tlabel }) combine_classes = np.append(combine_classes, com_class) mv_classes = np.append(mv_classes, mv_class) i_classes = np.append(i_classes, i_class) r_classes = np.append(r_classes, r_class) gt_label = np.append(gt_label, tlabel) combine_acc = np.sum((combine_classes == gt_label)) / gt_label.size i_acc = np.sum((i_classes == gt_label)) / gt_label.size mv_acc = np.sum((mv_classes == gt_label)) / gt_label.size r_acc = np.sum((r_classes == gt_label)) / gt_label.size print('Step %d finished with accuracy: %f , %f , %f, %f' % (step, i_acc, mv_acc, r_acc, combine_acc)) # Training procedure i_arr, mv_arr, r_arr, flow_arr, label = sess.run(next_data) summary, _, pred, loss1, loss2, loss3, loss4, loss5 = sess.run([merged, train_opt, pred_class, mv_class_loss, i_class_loss, r_class_loss, combine_loss, flmg_loss], feed_dict={mv_placeholder: mv_arr, i_placeholder: i_arr, flow_placeholder: flow_arr, r_placeholder: r_arr , label_placeholder : label, combine_value_: combine_acc, i_value_ : i_acc, mv_value_: mv_acc, r_value_ : r_acc}) if (step) % 10 == 0 : duration = time.time() - start_time print('Step %d: %.3f sec' % (step, duration), 'mv_loss:', loss1, 'i_loss:', loss2, 'r_loss:', loss3, 'fusion_loss:', loss4, 'flmg_loss:', loss5) print('GT:', label) print('Pred:', pred) writer.add_summary(summary, step) start_time = time.time() # Model Saving if (step) % 1000 == 0 and not step == 0 : ''' i_saver.save(sess, os.path.join(FLAGS.save_path, 'i_model.chkp'), global_step = step) mv_saver.save(sess, os.path.join(FLAGS.save_path, 'mv_model.chkp'), global_step = step) r_saver.save(sess, os.path.join(FLAGS.save_path, 'r_model.chkp'), global_step = step) i_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'i_resnet.chkp'), global_step = step) mv_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'mv_resnet.chkp'), global_step = step) r_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'r_resnet.chkp'), global_step = step) ''' my_saver.save(sess, os.path.join(FLAGS.save_path, 'all_net.chkp'), global_step = step) #if (step) % 10000 == 0 and not step == 0 : # i_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'i_resnet.chkp'), global_step = step) # mv_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'mv_resnet.chkp'), global_step = step) # r_resnet_saver.save(sess, os.path.join(FLAGS.save_path, 'r_resnet.chkp'), global_step = step) writer.close()
def build_model(self): """ :return: """ """ Helper Variables """ #self.global_step_tensor = tf.Variable(0, trainable=False, name='global_step') #self.global_step_inc = self.global_step_tensor.assign(self.global_step_tensor + 1) self.global_epoch_tensor = tf.Variable(0, trainable=False, name='global_epoch') self.global_epoch_inc = self.global_epoch_tensor.assign(self.global_epoch_tensor + 1) """ Inputs to the network """ with tf.variable_scope('inputs'): self.x, self.y, self.bi = self.data_loader.get_input() self.is_training = tf.placeholder(tf.bool, name='Training_flag') tf.add_to_collection('inputs', self.x) tf.add_to_collection('inputs', self.y) tf.add_to_collection('inputs', self.bi) tf.add_to_collection('inputs', self.is_training) """ Network Architecture """ with tf.variable_scope('network'): self.logits, end_points = resnet_v1.resnet_v1_50(inputs = self.x, num_classes = self.num_classes) self.logits = tf.squeeze(self.logits, axis=[1,2]) with tf.variable_scope('out'): #self.out = tf.squeeze(end_points['predictions'], axis=[1,2]) self.out = tf.nn.softmax(self.logits, dim=-1) tf.add_to_collection('out', self.out) print("Logits shape: ", self.logits.shape) print("predictions out shape: ", self.out.shape) print("network output argmax resnet") with tf.variable_scope('out_argmax'): self.out_argmax = tf.argmax(self.logits, axis=-1, output_type=tf.int64, name='out_argmax') #self.out_argmax = tf.squeeze(tf.argmax(self.out, 1), axis=[1]) print("Arg Max Shape: ", self.out_argmax.shape) with tf.variable_scope('loss-acc'): #one_hot_y = tf.one_hot(indices=self.y, depth=self.num_classes) self.loss = tf.losses.sparse_softmax_cross_entropy(labels = self.y, logits = self.logits) #probabilities = end_points['Predictions'] #accuracy, accuracy_update = tf.metrics.accuracy(labels = one_hot_y, predictions = self.out_argmax) #self.acc = tf.reduce_mean(tf.cast(tf.equal(self.y, self.out_argmax), tf.float32)) self.acc = self.evaluate_accuracy(self.y, self.out_argmax, self.is_training, self.config.patch_count) with tf.variable_scope('train_step'): update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS) with tf.control_dependencies(update_ops): self.train_step = self.optimizer.minimize(self.loss, global_step=self.global_step_tensor) tf.add_to_collection('train', self.train_step) tf.add_to_collection('train', self.loss) tf.add_to_collection('train', self.acc)
def top_feature_net(input, anchors, inds_inside, num_bases): stride = 8 arg_scope = resnet_v1.resnet_arg_scope(is_training=True) with slim.arg_scope(arg_scope): net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=16) block4 = end_points['resnet_v1_50/block4/unit_3/bottleneck_v1'] block3 = end_points['resnet_v1_50/block3/unit_5/bottleneck_v1'] block2 = end_points['resnet_v1_50/block2/unit_3/bottleneck_v1'] tf.summary.histogram('top_block4', block4) tf.summary.histogram('top_block3', block3) tf.summary.histogram('top_block2', block2) with tf.variable_scope("top_up") as sc: block4_ = conv2d_relu(block4, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='4') up_shape = tf.shape(block2) up4 = tf.image.resize_bilinear(block4_, [up_shape[1], up_shape[2]], name='up4') block3_ = conv2d_relu(block3, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='3') up3 = tf.image.resize_bilinear(block3_, [up_shape[1], up_shape[2]], name='up3') block2_ = conv2d_relu(block2, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='2') # up2 = upsample2d(block2_, factor = 2, has_bias=True, trainable=True, name='up2') up_34 = tf.add(up4, up3, name="up_add_3_4") up = tf.add(up_34, block2_, name="up_add_3_4_2") block = conv2d_relu(up, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', name='rgb_ft') with tf.variable_scope('rpn_top') as scope: up = conv2d_relu(block, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', name='2') scores = conv2d(up, num_kernels=2 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='score') probs = tf.nn.softmax(tf.reshape(scores, [-1, 2]), name='prob') deltas = conv2d(up, num_kernels=4 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='delta') deltasZ = conv2d(up, num_kernels=2 * num_bases, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='deltaZ') feature = block return feature, scores, probs, deltas #, rois, roi_scores,deltasZ, proposals_z, inside_inds_nms
def rgb_feature_net(input): arg_scope = resnet_v1.resnet_arg_scope(weight_decay=0.0) with slim.arg_scope(arg_scope): net, end_points = resnet_v1.resnet_v1_50(input, None, global_pool=False, output_stride=8) # pdb.set_trace() block4 = end_points['resnet_v1_50/block4'] block3 = end_points['resnet_v1_50/block3'] block2 = end_points['resnet_v1_50/block2'] # block1=end_points['resnet_v1_50/block1/unit_3/bottleneck_v1/conv1'] with tf.variable_scope("rgb_up") as sc: block4_ = conv2d_bn_relu(block4, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='4') up4 = upsample2d(block4_, factor=2, has_bias=True, trainable=True, name='up4') block3_ = conv2d_bn_relu(block3, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='3') up3 = upsample2d(block3_, factor=2, has_bias=True, trainable=True, name='up3') block2_ = conv2d_bn_relu(block2, num_kernels=256, kernel_size=(1, 1), stride=[1, 1, 1, 1], padding='SAME', name='2') up2 = upsample2d(block2_, factor=2, has_bias=True, trainable=True, name='up2') up_34 = tf.add(up4, up3, name="up_add_3_4") up = tf.add(up_34, up2, name="up_add_3_4_2") block = conv2d_bn_relu(up, num_kernels=256, kernel_size=(3, 3), stride=[1, 1, 1, 1], padding='SAME', name='rgb_ft') # block1_ = conv2d_bn_relu(block1, num_kernels=256, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='1') # up =tf.add(block1_, up_, name="up_add") # block = conv2d_bn_relu(block, num_kernels=512, kernel_size=(1,1), stride=[1,1,1,1], padding='SAME', name='2') #<todo> feature = upsample2d(block, factor = 4, ...) tf.summary.histogram('rgb_top_block', block) feature = block return feature
#print ("unitest for resnet") batch_size = 10 img_size = 256 img = cv2.imread( '/mnt/ilcompf8d0/user/weiyuewa/sources/pipeline1/tf_neural_renderer/img.png' ) # with tf.Session('') as sess: with tf.device('/gpu:0'): inputbatch = tf.expand_dims( tf.constant(img, dtype=tf.float32), axis=0) #tf.zeros([batch_size, img_size, img_size, 3]) logits, endpoints = resnet_v1.resnet_v1_50(inputbatch, 1000, is_training=False) config = tf.ConfigProto() config.gpu_options.allow_growth = True config.allow_soft_placement = True config.log_device_placement = False sess = tf.Session(config=config) variables_to_restore = [] a = [ name for name, _ in checkpoint_utils.list_variables( 'pretrained_model/resnet_v1_50.ckpt') ] # print a for var in slim.get_model_variables():
def _get_resnet_features(inputs): with slim.arg_scope(resnet_v1.resnet_arg_scope()): resnet_v1.resnet_v1_50(inputs, num_classes=None, is_training=True) return tf.get_default_graph().get_tensor_by_name( 'resnet_v1_50/block4/unit_3/bottleneck_v1/Relu:0')