[GitHub] [incubator-mxnet] david-seiler commented on a change in pull request #14617: PDF operators for the random samplers, and also the Dirichlet

[GitBox <gi...@apache.org>]

david-seiler commented on a change in pull request #14617: PDF operators for the random samplers, and also the Dirichlet
URL: https://github.com/apache/incubator-mxnet/pull/14617#discussion_r281726103
 
 

 ##########
 File path: tests/python/unittest/test_random.py
 ##########
 @@ -290,11 +315,84 @@ def check_with_device(device, dtype):
                 for check_name, check_func, tol in symbdic['checks']:
                     assert np.abs(check_func(samples, params)) < tol, "symbolic test: %s check for `%s` did not pass" % (check_name, name)
 
-@with_seed()
+        # check pdfs with only a subset of the generated samples
+        un1 = np.resize(un1, (un1.shape[0], un1.shape[1], pdfshape[0], pdfshape[1]))
+        print(name)
+        symbol  = symbdic['pdfsymbol']
+        pdffunc = symbdic['pdffunc']
+        v0 = mx.sym.Variable('v0')
+        v1 = mx.sym.Variable('v1')
+        v2 = mx.sym.Variable('v2')
+        p1 = np.array(symbdic['inputs'][0][1])
+        p2 = None if single_param else np.array(symbdic['inputs'][1][1])
+        # Move samples away from boundaries of support
+        if name == 'gamma' or name == 'exponential':
+           un1 = np.maximum(un1, 1e-1)
+        if name == 'uniform':
+           un1 = np.minimum(np.maximum(un1.reshape((un1.shape[0],un1.shape[1],-1)), p1.reshape((p1.shape[0],p1.shape[1],-1))+1e-4),
+                            p2.reshape((p2.shape[0],p2.shape[1],-1))-1e-4).reshape(un1.shape) 
+        for use_log in [False, True]:
+            test_pdf = symbol(v0, v1, is_log=use_log) if single_param else symbol(v0, v1, v2, is_log=use_log)
+            forw_atol  = 1e-7 if dtype != np.float16 else 1e-3
+            forw_rtol  = 1e-4 if dtype != np.float16 else 5e-2
+            backw_atol = 1e-3
+            backw_rtol = 5e-2
+            if single_param:
+                res = pdffunc(un1.reshape((un1.shape[0],un1.shape[1],-1)),
+                    p1.reshape((p1.shape[0],p1.shape[1],-1))).reshape(un1.shape)
+                if use_log: 
+                    res = np.log(res)
+                check_symbolic_forward(test_pdf, [un1, p1], [res], atol=forw_atol, rtol=forw_rtol, dtype=dtype)
+                if dtype == np.float64:
+                  grad_nodes = ['v1'] if symbdic['discrete'] else ['v0', 'v1']
+                  check_numeric_gradient(test_pdf, [un1, p1], grad_nodes=grad_nodes, atol=backw_atol, rtol=backw_rtol, dtype=dtype)
+            else:
+                res = pdffunc(un1.reshape((un1.shape[0],un1.shape[1],-1)),
+                    p1.reshape((p1.shape[0],p1.shape[1],-1)),
+                    p2.reshape((p2.shape[0],p2.shape[1],-1))).reshape(un1.shape)
+                if use_log:
+                    res = np.log(res)
+                check_symbolic_forward(test_pdf, [un1, p1, p2], [res], atol=forw_atol, rtol=forw_rtol, dtype=dtype)
+                if dtype == np.float64:
+                  grad_nodes = ['v1', 'v2'] if symbdic['discrete'] else ['v0', 'v1', 'v2']
+                  check_numeric_gradient(test_pdf, [un1, p1, p2], grad_nodes=grad_nodes, atol=backw_atol, rtol=backw_rtol, dtype=dtype)
+
+@with_seed(1000)
+def test_dirichlet():
+    num_classes = 2
+    num = 100
+    alpha = np.random.uniform(low=0.5, high=10, size=(4, num_classes))
+
+    samples = []
+    results = []
+    for a in alpha:
+        v = ss.dirichlet.rvs(a, size=num)
+        samples.append(v)
+        results.append(ss.dirichlet.logpdf(v.transpose(), a))
+    samples = np.concatenate(samples, axis=0).reshape((2, 2, num, num_classes))
+    results = np.concatenate(results, axis=0).reshape((2, 2, num))
+
+    alpha = alpha.reshape((2, 2, num_classes))
+
+    for dtype in [np.float32, np.float64]:
+        forw_atol  = 1e-5
+        forw_rtol  = 1e-4
+        backw_atol = 1e-5 if dtype == np.float64 else 1e-3
+        backw_rtol = 1e-4 if dtype == np.float64 else 5e-2
+        for use_log in [False, True]:
+            print("use_log",use_log)
+            v0 = mx.sym.Variable('v0')
+            v1 = mx.sym.Variable('v1')
+            test_pdf = mx.sym.random_pdf_dirichlet(v0, v1, is_log=use_log)
+            res = results if use_log else np.exp(results)
+            check_symbolic_forward(test_pdf, [samples, alpha], [res], atol=forw_atol, rtol=forw_rtol, dtype=dtype)
+            if dtype == 'float64':
+                check_numeric_gradient(test_pdf, [samples, alpha], numeric_eps=1e-7, atol=backw_atol, rtol=backw_rtol, dtype=dtype)
+
 def test_random():
-    check_with_device(mx.context.current_context(), 'float16')
-    check_with_device(mx.context.current_context(), 'float32')
-    check_with_device(mx.context.current_context(), 'float64')
+    check_with_device(mx.context.current_context(), np.float16)
+    check_with_device(mx.context.current_context(), np.float32)
+    check_with_device(mx.context.current_context(), np.float64)
 
 Review comment:
   Done

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services