Unlock 30% off on Manual Testing Annual Plans this Holiday Season.

Copied to Clipboard!

How to use freemem method in autotest

Best Python code snippet using autotest_python

test_memory.py

Source:test_memory.py

...17# The GC need to be enabled for those tests to work correctly.18if not getattr(mode_with_gpu.linker, 'allow_gc', False):19    mode_with_gpu.linker = copy.copy(mode_with_gpu.linker)20    mode_with_gpu.linker.allow_gc = True21def freemem(extra_alloc=0):22    """23    Return the free memory on the gpu in megabytes.24    """25    gc.collect()26    gc.collect()27    gc.collect()28    n_mallocs = cuda.cuda_ndarray.cuda_ndarray.outstanding_mallocs()29    if hasattr(cuda.cuda_ndarray.cuda_ndarray, "theano_allocated"):30        theano_alloc = cuda.cuda_ndarray.cuda_ndarray.theano_allocated()31        return ("(n malloc/theano mem allocated in KB)",32                n_mallocs + extra_alloc,33                int(theano_alloc / 1024))34    return ("n malloc on the gpu", n_mallocs + extra_alloc)35    # I don't use the following by default as if there is other stuff running36    # on the GPU, this won't work.37    mem_info = cuda.cuda_ndarray.cuda_ndarray.mem_info()38    gpu_used = (mem_info[1] - mem_info[0]) / 1024 ** 239    mem_info_msg = "(n malloc/gpu mem used in MB)"40    return (mem_info_msg, n_mallocs, int(gpu_used))41def test_memory():42    """43    We test that we do not keep link to memory between Theano function call44    and during Theano compilation45    The origin of this code come from Aaron Vandenoord and Sander Dieleman.46    I have their autorisation to put this in Theano with the Theano license.47    note::48        This test can fail if there is other process running on the gpu.49    """50    shapes = (200, 100)51    # more_alloc1 was different for each dtype in the past.52    # more_alloc2 is still currently not the same for both dtype.53    # when dtype is float32, the computation is done on the gpu.54    # This insert constant on the gpu during compilation55    # that raise the number of alloc.56    # When dtype is float64, only the shared is on the gpu and it is transferd57    # to the cpu for computation. So no extra alloc after compilation.58    # more_alloc1 if after the first compilation, more_alloc2 after the second.59    for dtype, more_alloc1, more_alloc2 in [("float32", 0, 3),60                                            ("float64", 0, 0)]:61        print(dtype)62        test_params = np.asarray(np.random.randn(np.prod(shapes)), dtype)63        some_vector = tensor.vector('some_vector', dtype=dtype)64        some_matrix = some_vector.reshape(shapes)65        mem1 = freemem()66        print("Before shared variable", mem1)67        variables = cuda.shared_constructor(np.ones((shapes[1],),68                                                    dtype='float32'))69        derp = tensor.sum(tensor.dot(some_matrix[:shapes[0]], variables))70        print("Shared took ",71              np.prod(variables.get_value(72                  borrow=True,73                  return_internal_type=True).shape) *74              4 / 1024,75              "kB")76        mem2 = freemem()77        print("Before compilation", mem2)78        mem2_1 = freemem(extra_alloc=more_alloc1)79        mem2_2 = freemem(extra_alloc=more_alloc2)80        obj = theano.function([some_vector], derp, mode=mode_with_gpu)81        mem3 = freemem()82        print("After function compilation 1", mem3)83        assert mem2_1 == mem3, (mem2_1, mem3, dtype)84        grad_derp = tensor.grad(derp, some_vector)85        grad = theano.function([some_vector], grad_derp, mode=mode_with_gpu)86        mem4 = freemem()87        print("After function compilation 2", mem4)88        assert mem2_2 == mem4, (mem2_2, mem4, dtype)89        for i in range(3):90            obj(test_params)91            print("After function evaluation 1", freemem())92            assert mem2_2 == freemem(), (mem2_2, freemem())93            grad(test_params)94            print("After function evaluation 2", freemem())95            assert mem2_2 == freemem(), (mem2_2, freemem())96        del obj97        # print "After deleting function 1", freemem()98        # assert mem2 == freemem(), (mem2, freemem())99        del grad100        print("After deleting function 2", freemem())101        assert mem2 == freemem(), (mem2, freemem())102        del derp, variables, grad_derp103        print("After deleting shared variable and ref to it", freemem())104        assert mem1 == freemem(), (mem1, freemem())105@theano.configparser.change_flags(**{'vm.lazy': True})106def test_memory_lazy():107    """As test_memory, but with the ifelse op.108    We need to test it as the ifelse op with the [c]vm create op not109    executed in the graph. This mess with [c]vm gc implementation.110    """111    shapes = (50, 100)112    # more_alloc1 is not the same for both dtype.113    # when dtype is float32, the computation is done on the gpu.114    # This insert constant on the gpu during compilation115    # that raise the number of alloc.116    # When dtype is float64, only the shared is on the gpu and it is transferd117    # to the cpu for computation. So no extra alloc after compilation.118    # more_alloc1 if after the first compilation119    for dtype, more_alloc1 in [("float32", 1),120                               ("float64", 0)]:121        print(dtype)122        test_params = np.asarray(np.random.randn(np.prod(shapes)), dtype)123        some_vector = tensor.vector('some_vector', dtype=dtype)124        some_matrix = some_vector.reshape(shapes)125        branch_select = tensor.iscalar()126        mem1 = freemem()127        print("Before shared variable", mem1)128        variables = cuda.shared_constructor(np.ones((shapes[1],),129                                                    dtype='float32'))130        derp = tensor.sum(tensor.dot(some_matrix[:shapes[0]], variables))131        derp = ifelse.IfElse(1)(branch_select,132                                derp, some_matrix[:shapes[0]].sum())133        derp += 1134        print("Shared took ",135              np.prod(variables.get_value(136                  borrow=True,137                  return_internal_type=True).shape) *138              4 / 1024,139              "kB")140        mem2 = freemem()141        print("Before compilation", mem2)142        mem2_1 = freemem(extra_alloc=more_alloc1)143        obj = theano.function([some_vector, branch_select], derp,144                              mode=mode_with_gpu)145        # theano.printing.debugprint(obj, print_type=True)146        mem3 = freemem()147        print("After function compilation 1", mem3)148        assert mem2_1 == mem3, (mem2_1, mem3)149        for i in range(3):150            obj(test_params, 1)151            print("After function evaluation branch true", freemem())152            assert mem2_1 == freemem(), (mem2_1, freemem())153            obj(test_params, 0)154            print("After function evaluation branch false", freemem())155            assert mem2_1 == freemem(), (mem2_1, freemem())156        del obj157        print("After deleting function 1", freemem())158        assert mem2 == freemem(), (mem2, freemem())159        del derp, variables160        print("After deleting shared variable and ref to it", freemem())...

Automation Testing Tutorials

Learn to execute automation testing from scratch with LambdaTest Learning Hub. Right from setting up the prerequisites to run your first automation test, to following best practices and diving deeper into advanced test scenarios. LambdaTest Learning Hubs compile a list of step-by-step guides to help you be proficient with different test automation frameworks i.e. Selenium, Cypress, TestNG etc.