regular update

py/glu_ocr.py

@@ -14,7 +14,7 @@
 import mxnet.gluon as gluon
 import mxnet as mx
 
-vobsize = 3593
+vobsize = 100
 batch_size = 2
 model_prefix = '../params/ctc'
 img_w = 256
@@ -97,9 +97,15 @@ def train():
 				loss_ctc = ctc_loss(output, label)
 				loss_ctc = (label!=-1).sum(axis=1)*loss_ctc
 			loss_ctc.backward()
-			loss = loss_ctc.mean()
+			loss = loss_ctc.mean().asnumpy()
 			if i % 20 == 0:
 				print 'epoch %s\t\tbatch %s\t\tloss=%s'%(e, i, loss)
 
+'''
+"D:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Auxiliary\Build\vcvars64.bat" -vcvars_ver=14.11
+cmake -G "Visual Studio 15 2017 Win64" -T cuda=9.1,host=x64 -DUSE_CUDA=1 -DUSE_CUDNN=1 -DUSE_NVRTC=1 -DUSE_OPENCV=1 -DUSE_OPENMP=1 -DUSE_BLAS=open -DUSE_LAPACK=1 -DUSE_DIST_KVSTORE=0 -DCUDA_ARCH_LIST=Common -DCUDA_TOOLSET=9.1 -DCUDNN_INCLUDE=E:\Env\cuDNN\cuda\include -DCUDNN_LIBRARY=E:\Env\cuDNN\cuda\lib\x64\cudnn.lib "E:\Env\incubator-mxnet"
+
+'''
+
 if __name__ == '__main__':
 	train()

py/ocr.py

@@ -20,9 +20,9 @@
 import model
 
 vobsize = 3593
-batch_size = 2
+batch_size = 32
 model_prefix = '../params/ctc'
-ctx = mx.cpu(0)
+ctx = mx.gpu(0)
 img_w = 256
 img_h = 32
 seq_l = 20
@@ -70,14 +70,14 @@ def train():
 		dataiter,
 		optimizer = 'rmsprop',
 		optimizer_params = {'learning_rate':0.005},
-		eval_metric = 'loss',
+		eval_metric = 'acc',
 		batch_end_callback = mx.callback.Speedometer(batch_size, 5),
 		epoch_end_callback = mx.callback.do_checkpoint(model_prefix, 1),
 		num_epoch = 100,
 	)
 
 def test():
-	symbol = model.ctc(vobsize, False)
+	symbol = model.ctc(200, False)
 	dataiter = mx.io.NDArrayIter(data=mx.nd.normal(loc=0, scale=1, shape=(1, 1, img_w, img_h)), label=mx.nd.normal(loc=0, scale=1, shape=(1, seq_l)), batch_size=1, shuffle=True)
 	symbol = mx.mod.Module(symbol=symbol, context=mx.cpu(0), data_names=('data',), label_names=('label',))
 	symbol.bind(for_training=False, data_shapes=dataiter.provide_data)
@@ -87,4 +87,5 @@ def test():
 	print out.shape
 
 if __name__ == '__main__':
-	test()
+	print 'using mxnet version %s'%mx.__version__
+	train()

py/scratch.py

@@ -120,6 +120,7 @@ def init_data():
 		fs = fs[:20]
 		random.shuffle(fs)
 		idx = 0
+		iterstop = 0
 		while True:
 			try:
 				data = np.zeros((batch_size, 1, img_w, img_h))

py/test_utils.py

@@ -102,25 +102,13 @@ def foo():
 
 if __name__ == '__main__':
     in_var = mx.sym.Variable('data')
-
-    t1 = mx.symbol.reshape(in_var, (6,10))
-    t2 = mx.symbol.FullyConnected(t1, num_hidden=5, no_bias=True)
-    t3 = mx.symbol.reshape(t2, (3,2,5))
-
     labels_var = mx.sym.Variable('label')
-    ctc = mx.sym.contrib.ctc_loss(t3, labels_var)
+    ctc = mx.sym.contrib.ctc_loss(in_var, labels_var)
     loss = mx.symbol.MakeLoss(ctc)
-
-    arg_names = loss.list_arguments()
-    arg_shapes,_,_ = loss.infer_shape(data=(6,2,5), label=(2,3))
-    for name, shape in zip(arg_names, arg_shapes):
-        print '%s\t\t\t%s'%(name, shape)
-
+    arg_shapes,_,_ = loss.infer_shape(data=(6,2,1000), label=(2,3))
     arg_array = [mx.nd.normal(shape=shape, ctx=ctx) for shape in arg_shapes]
     exe = loss.bind(ctx=ctx, args=arg_array)
     exe.forward(is_train=True)
     exe.backward()
     outTest = exe.outputs[0]
-
     print '%s'%(outTest.asnumpy())
-    print '----------'

py/val_datagen.py

@@ -2,7 +2,7 @@
 
 import random
 import chardet
-import pygame
+# import pygame
 import os
 import numpy as np
 
@@ -86,7 +86,31 @@ def fee(self):
 	def fum(self):
 		return self.val
 
+def calc_norm(path):
+	array_list = []
+	norm, cnt = 0, 0
+	for _,_,fs in os.walk(path):
+		fs = sorted(fs)[:50]
+		for f in fs:
+			array = np.array(Image.open(path+f).convert('L'))
+			# array = 255*(array-np.min(array))/(np.max(array)-np.min(array))
+			array_list.append(array)
+	for i in range(len(array_list)):
+		for j in range(i):
+			a1 = array_list[i]
+			a2 = array_list[j]
+			norm += np.linalg.norm(a1-a2)
+			cnt += 1
+		# print '%s finished'%(i)
+	return norm/(cnt*(640*480))
+
 if __name__ == '__main__':
 	# init_dataiter()
-	t = T(3)
-	print t.foo().fee().fum()
+	# t = T(3)
+	# print t.foo().fee().fum()
+	norm_sal = calc_norm('E:/Paper/feature_validation/data/norm/saliency_512/')
+	norm_vgg = calc_norm('E:/Paper/feature_validation/data/norm/vgg_feature/')
+	print 'norm sal: %s'%norm_sal
+	print 'norm vgg: %s'%norm_vgg
+	# # norm sal: 9251865169.961721
+	# # norm vgg: 9803585796.040016