final-battle

README.md

@@ -16,12 +16,12 @@ All the codes was writen by myself and tested under TensorFlow 1.6, Python 3.5,
 
 There are still other ways to further improve the performance but I didn't try those in this competition because of their limitations in applications, for example:
 
-- More large input image size
+- More larger input image size
 - More deeper backbone networks
 - Locate clothes first by detection networks
 - Multi-scale supervision for Stacked Hourglass Models
 - Extra-regressor to refine the location of keypoints
-- Multi-crop ensemble for single image predictions
+- Multi-crop or multi-scale ensemble for single image predictions
 
 If you find it's useful to your research or competitions, any contribution or star to this repo is welcomed.
 

ensemble_from_csv.py

@@ -32,7 +32,7 @@
 # 'sub_2_hg_4_256_64-half_epoch.csv',
 # 'sub_2_hg_8_256_64_v1-half_epoch.csv']#['cpn_2_320_160_1e-3.csv', 'sub_2_hg_4_256_64.csv', 'sub_2_cpn_320_100_1e-3.csv', 'sub_2_hg_8_256_64.csv']
 
-ensemble_subs = ['det-cpn-ohkm-384-96-lr1e-4-b10_predv1_finetune.csv', 'sext_cpn_v1_fine.csv']
+ensemble_subs = ['sext_cpn_flip.csv', 'detxt_cpn_flip.csv']
 
 
 def parse_comma_list(args):

eval_all_cpn_onepass.py

@@ -48,7 +48,7 @@
     'gpu_memory_fraction', 1., 'GPU memory fraction to use.')
 # scaffold related configuration
 tf.app.flags.DEFINE_string(
-    'data_dir', '../Datasets/tfrecords_test_stage1_b',#tfrecords_test tfrecords_test_stage1_b
+    'data_dir', '../Datasets/tfrecords_test',#tfrecords_test tfrecords_test_stage1_b
     'The directory where the dataset input data is stored.')
 tf.app.flags.DEFINE_string(
     'dataset_name', '{}_*.tfrecord', 'The pattern of the dataset name to load.')
@@ -112,6 +112,10 @@
   'seresnet50_cpn': {'backbone': seresnet_cpn.cascaded_pyramid_net, 'logs_sub_dir': 'logs_se_cpn'},
   'seresnext50_cpn': {'backbone': seresnet_cpn.xt_cascaded_pyramid_net, 'logs_sub_dir': 'logs_sext_cpn'},
   'detnext50_cpn': {'backbone': detxt_cpn.cascaded_pyramid_net, 'logs_sub_dir': 'logs_detxt_cpn'},
+  'large_seresnext_cpn': {'backbone': lambda inputs, output_channals, heatmap_size, istraining, data_format : seresnet_cpn.xt_cascaded_pyramid_net(inputs, output_channals, heatmap_size, istraining, data_format, net_depth=50),
+                        'logs_sub_dir': 'logs_large_sext_cpn'},
+  'large_detnext_cpn': {'backbone': lambda inputs, output_channals, heatmap_size, istraining, data_format : detxt_cpn.cascaded_pyramid_net(inputs, output_channals, heatmap_size, istraining, data_format, net_depth=50),
+                        'logs_sub_dir': 'logs_large_detxt_cpn'},
   'head_seresnext50_cpn': {'backbone': seresnet_cpn.head_xt_cascaded_pyramid_net, 'logs_sub_dir': 'logs_head_sext_cpn'},
 }
 

eval_script.sh

@@ -0,0 +1,16 @@
+#! /bin/bash
+
+# export CUDA_VISIBLE_DEVICES='0'
+# source /home/kapok/pyenv35/bin/activate
+# cd /media/rs/0E06CD1706CD0127/Kapok/Chi/fashionAI/Codes
+python eval_all_cpn_onepass.py --run_on_cloud=False --backbone=seresnext50_cpn
+python eval_all_cpn_onepass.py --run_on_cloud=False --backbone=detnext50_cpn
+python eval_all_cpn_onepass.py --run_on_cloud=False --backbone=large_seresnext_cpn --train_image_size=512 --heatmap_size=128
+python eval_all_cpn_onepass.py --run_on_cloud=False --backbone=large_detnext_cpn --train_image_size=512 --heatmap_size=128
+
+# for training
+python train_senet_cpn_onebyone.py --run_on_cloud=False
+python train_detxt_cpn_onebyone.py --run_on_cloud=False
+python train_large_xt_cpn_onebyone.py --run_on_cloud=False --backbone=detxt
+python train_large_xt_cpn_onebyone.py --run_on_cloud=False --backbone=sext
+

net/detxt_cpn.py

@@ -293,7 +293,7 @@ def dilated_se_next_bottleneck_block(inputs, input_filters, name_prefix, is_trai
     return tf.nn.relu(pre_act, name=name_prefix + '/relu')
 
 # the input image should in BGR order, note that this is not the common case in Tensorflow
-def sext_cpn_backbone(input_image, istraining, data_format, group=32):
+def sext_cpn_backbone(input_image, istraining, data_format, net_depth=50, group=32):
     bn_axis = -1 if data_format == 'channels_last' else 1
 
     if data_format == 'channels_last':
@@ -304,8 +304,11 @@ def sext_cpn_backbone(input_image, istraining, data_format, group=32):
         swaped_input_image = tf.stack([image_channels[2], image_channels[1], image_channels[0]], axis=1)
     #swaped_input_image = input_image
 
+    if net_depth not in [50, 101]:
+        raise TypeError('Only ResNeXt50 or ResNeXt101 is supprted now.')
+
     input_depth = [256, 512, 1024] # the input depth of the the first block is dummy input
-    num_units = [3, 4, 6]
+    num_units = [3, 4, 6] if net_depth==50 else [3, 4, 23]
     block_name_prefix = ['conv2_{}', 'conv3_{}', 'conv4_{}']
 
     if data_format == 'channels_first':
@@ -475,9 +478,9 @@ def global_net_sext_bottleneck_block(inputs, input_filters, is_training, data_fo
         pre_act = tf.add(residuals, rescaled_feat, name=name_prefix + '_add')
         return tf.nn.relu(pre_act, name=name_prefix + '/relu')
 
-def cascaded_pyramid_net(inputs, output_channals, heatmap_size, istraining, data_format):
+def cascaded_pyramid_net(inputs, output_channals, heatmap_size, istraining, data_format, net_depth=50):
     #with tf.variable_scope('resnet50', 'resnet50', values=[inputs]):
-    end_points = sext_cpn_backbone(inputs, istraining, data_format)
+    end_points = sext_cpn_backbone(inputs, istraining, data_format, net_depth=net_depth)
     pyramid_len = len(end_points)
     up_sampling = None
     pyramid_heatmaps = []