Merge branch 'master' of https://github.com/pengsun/DeepLearningTutorials

# Conflicts:
#	.gitignore
#	.idea/workspace.xml

Author: pengsun

.gitignore

@@ -13,4 +13,6 @@ html
 *.pyc
 *~
 *.swp
-/.idea/workspace.xml
+/.idea/*.xml
+/.idea/.name
+/.idea/DeepLearningTutorials-Peng.iml

.idea/.name

@@ -2,9 +2,9 @@
 import numpy as np
 from theano import tensor as T, function, shared
 
+rng = np.random.RandomState(1234)
 
 def init_theta():
-    rng = np.random.RandomState(1234)
     szAll = [(784, 512), (512, 256), (256, 10)]
     tmp = []
     for sz in szAll:

.idea/DeepLearningTutorials-Peng.iml

@@ -0,0 +1,63 @@
+import dataset, models, loss, optim
+import numpy as np
+from theano import tensor as T, function, shared
+
+rng = np.random.RandomState(1234)
+
+
+def init_theta_fc():
+    szAll = [(256, 128), (128, 10)]
+    tmp = []
+    for sz in szAll:
+        tmp.append(
+            np.asarray(
+                rng.uniform(
+                    low=-np.sqrt(6. / (sz[0] + sz[1])),
+                    high=np.sqrt(6. / (sz[0] + sz[1])),
+                    size=sz
+                ),
+                dtype='float32'
+            )
+        )
+        tmp.append(np.zeros((sz[1],), 'float32'))
+    sh_theta = []
+    for each in tmp:
+        sh_theta.append(shared(each, borrow=True))
+    return sh_theta
+
+
+def init_theta():
+    sz_flt = [(16, 1, 5, 5), (16, 16, 5, 5) ]
+    tmp = []
+    for sz in szAll:
+        tmp.append(
+            np.asarray(
+                rng.uniform(
+                    low=-np.sqrt(6. / (sz[0] + sz[1])),
+                    high=np.sqrt(6. / (sz[0] + sz[1])),
+                    size=sz
+                ),
+                dtype='float32'
+            )
+        )
+        tmp.append(np.zeros((sz[1],), 'float32'))
+    sh_theta = []
+    for each in tmp:
+        sh_theta.append(shared(each, borrow=True))
+    return sh_theta
+
+
+if __name__ == '__main__':
+    # config
+    itMax = 195*2
+    szBatch = 256
+    lr = 0.1
+    vaFreq = 20
+
+    import tr_va_te
+    tr_va_te.run(itMax=itMax,
+                 szBatch=szBatch,
+                 lr=lr,
+                 vaFreq=vaFreq,
+                 init_theta=init_theta,
+                 mo_create=models.create_mlp)

.idea/misc.xml

@@ -0,0 +1,43 @@
+from theano import tensor as T, shared, function
+import numpy as np
+
+rng = np.random.RandomState(1234)
+
+# fprop
+#p = T.matrix(name='p', dtype='float32')
+p = shared(
+    np.asarray(
+        rng.uniform(low=-1, high=1, size=(3, 4)),
+        dtype='float32'
+    ),
+    borrow=True
+)
+I = T.matrix(name='I', dtype='float32')
+z = p * I
+u = z**2
+v = p + u
+ell = T.mean(v)
+# bprop
+dp = T.grad(ell, wrt=p)
+# the graph
+fg = function(
+    inputs=[I],
+    outputs=[ell, v, dp],
+    updates=[(p, p - np.float32(0.1)*dp)]
+)
+
+# fire
+pval = np.asarray(
+    rng.uniform(low=-1, high=1, size=(3, 4)),
+    dtype='float32'
+)
+Ival = np.asarray(
+    rng.uniform(low=-1, high=1, size=(3, 4)),
+    dtype='float32'
+)
+[ellval, vval, dpval] = fg(Ival)
+
+pass
+
+
+