I'm trying to get a basic DBM model working using Pylearn. I'm starting with the model in pylearn2/scripts/tutorials/dbm_demo/rbm.yaml. I've replaced the MNIST dataset with the BlackBoxDataset, and removed the bias_from_marginals line due to the out of\r\n range values. When I build the model, I get some pretty hairy error messages, but it seems to boil down to creating a bad kernel which is failing to compile. The error message is:
\r\nmod.cu(65): error: expected a ")"\r\n
The entire kernel code is printed out on error, and the section with the error looks like this:
\r\n58 y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16));\r\n
59 y1 -= (y1 < 0 || y1 >= M2) ? M2 : 0;
60 y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16));
61 y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
62 y2 += x23;
63 y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
64 y2 += y1;
65 y2 -= (y2 < 0 or y2 >= M2) ? M2 : 0;
66
67 x23 = x22;
68 x22 = x21;
69 x21 = y2;
So it appears that on line 65 it's placing an "or" instead of a "||". Is this a problem anyone's seen before?
\r\n
\r\n
The full stack trace is as follows:
\r\n['nvcc', '-shared', '-g', '-O3', '-arch=sm_20', '-Xlinker', '/DEBUG', '-m64', '-Xcompiler', '-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,/Zi', '-IC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray', '-IAUTO\\\\include', '-IC:\\\\Python27\\\\lib\\\\site-packages\\\\numpy\\\\core\\\\include', '-IC:\\\\Python27\\\\include', '-IC:\\\\Python27\\\\lib\\\\site-packages\\\\theano\\\\sandbox\\\\cuda', '-o', 'C:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\tmphlgfm0\\\\465ad7f8e1d438d913f57bdca2bb9eb1.pyd', 'mod.cu', '-LC:\\\\Python27\\\\libs', '-LAUTO\\\\lib', '-LC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray', '-LAUTO\\\\lib', '-LAUTO\\\\lib64', '-LC:\\\\Python27', '-lpython27', '-lcudart', '-lcublas', '-lcuda_ndarray']\r\n
Traceback (most recent call last):
File "D:\\Projects\\Pylearn\\main.py", line 8, in <module>
train_obj.main_loop()
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\train.py", line 109, in main_loop
self.algorithm.setup(model=self.model, dataset=self.dataset)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\training_algorithms\\sgd.py", line 217, in setup
grads, updates = self.cost.get_gradients(model, X, ** fixed_var_descr.fixed_vars)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\costs\\cost.py", line 207, in get_gradients
result = cost.get_gradients(model, X, Y_to_pass, ** kwargs)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\costs\\dbm.py", line 328, in get_gradients
layer_to_chains = model.make_layer_to_state(self.num_chains)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\models\\dbm\\__init__.py", line 343, in make_layer_to_state
states = [ layer.make_state(num_examples, rng) for layer in layers ]
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\models\\dbm\\__init__.py", line 1507, in make_state
(h_state , h_sample)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function.py", line 221, in function
profile=profile)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\pfunc.py", line 506, in pfunc
on_unused_input=on_unused_input)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function_module.py", line 1339, in orig_function
defaults)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function_module.py", line 1167, in create
_fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\link.py", line 383, in make_thunk
output_storage = output_storage)[:3]
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\vm.py", line 800, in make_all
for node in order]
File "C:\\Python27\\lib\\site-packages\\theano\\sandbox\\cuda\\__init__.py", line 244, in make_thunk
compute_map, no_recycling)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\op.py", line 580, in make_thunk
output_storage=node_output_storage)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 913, in make_thunk
keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 856, in __compile__
keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 1279, in cthunk_factory
key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cmodule.py", line 975, in module_from_key
module = compile_steps.next()
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 1201, in compile_cmodule_by_step
preargs=preargs)
File "C:\\Python27\\lib\\site-packages\\theano\\sandbox\\cuda\\nvcc_compiler.py", line 391, in compile_str
'for cmd', ' '.join(cmd))
Exception: ('nvcc return status', 2, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_20 -Xlinker /DEBUG -m64 -Xcompiler -DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,/Zi -IC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray -IAUTO\\\\include -IC:\\\\Python27\\\\lib\\\\site-packages\\\\numpy\\\\core\\\\include -IC:\\\\Python27\\\\include -IC:\\\\Python27\\\\lib\\\\site-packages\\\\theano\\\\sandbox\\\\cuda -o C:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\tmphlgfm0\\\\465ad7f8e1d438d913f57bdca2bb9eb1.pyd mod.cu -LC:\\\\Python27\\\\libs -LAUTO\\\\lib -LC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray -LAUTO\\\\lib -LAUTO\\\\lib64 -LC:\\\\Python27 -lpython27 -lcudart -lcublas -lcuda_ndarray', '[*1 -> GPU_mrg_uniform{CudaNdarrayType(float32, matrix),no_inplace}(<CudaNdarrayType(float32, vector)>, <TensorType(int32, vector)>), *1::1]')
\r\n
\r\n
\r\n
", "messages": [ { "id": "23967", "postDate": "05/07/2013 00:46:58", "content": "
I'm trying to get a basic DBM model working using Pylearn. I'm starting with the model in pylearn2/scripts/tutorials/dbm_demo/rbm.yaml. I've replaced the MNIST dataset with the BlackBoxDataset, and removed the bias_from_marginals line due to the out of\r\n range values. When I build the model, I get some pretty hairy error messages, but it seems to boil down to creating a bad kernel which is failing to compile. The error message is:
\r\nmod.cu(65): error: expected a ")"\r\n
The entire kernel code is printed out on error, and the section with the error looks like this:
\r\n58 y1 = ((x21 & MASK2) << i15) + (MULT2 * (x21 >> i16));\r\n
59 y1 -= (y1 < 0 || y1 >= M2) ? M2 : 0;
60 y2 = ((x23 & MASK2) << i15) + (MULT2 * (x23 >> i16));
61 y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
62 y2 += x23;
63 y2 -= (y2 < 0 || y2 >= M2) ? M2 : 0;
64 y2 += y1;
65 y2 -= (y2 < 0 or y2 >= M2) ? M2 : 0;
66
67 x23 = x22;
68 x22 = x21;
69 x21 = y2;
So it appears that on line 65 it's placing an "or" instead of a "||". Is this a problem anyone's seen before?
\r\n
\r\n
The full stack trace is as follows:
\r\n['nvcc', '-shared', '-g', '-O3', '-arch=sm_20', '-Xlinker', '/DEBUG', '-m64', '-Xcompiler', '-DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,/Zi', '-IC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray', '-IAUTO\\\\include', '-IC:\\\\Python27\\\\lib\\\\site-packages\\\\numpy\\\\core\\\\include', '-IC:\\\\Python27\\\\include', '-IC:\\\\Python27\\\\lib\\\\site-packages\\\\theano\\\\sandbox\\\\cuda', '-o', 'C:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\tmphlgfm0\\\\465ad7f8e1d438d913f57bdca2bb9eb1.pyd', 'mod.cu', '-LC:\\\\Python27\\\\libs', '-LAUTO\\\\lib', '-LC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray', '-LAUTO\\\\lib', '-LAUTO\\\\lib64', '-LC:\\\\Python27', '-lpython27', '-lcudart', '-lcublas', '-lcuda_ndarray']\r\n
Traceback (most recent call last):
File "D:\\Projects\\Pylearn\\main.py", line 8, in <module>
train_obj.main_loop()
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\train.py", line 109, in main_loop
self.algorithm.setup(model=self.model, dataset=self.dataset)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\training_algorithms\\sgd.py", line 217, in setup
grads, updates = self.cost.get_gradients(model, X, ** fixed_var_descr.fixed_vars)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\costs\\cost.py", line 207, in get_gradients
result = cost.get_gradients(model, X, Y_to_pass, ** kwargs)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\costs\\dbm.py", line 328, in get_gradients
layer_to_chains = model.make_layer_to_state(self.num_chains)
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\models\\dbm\\__init__.py", line 343, in make_layer_to_state
states = [ layer.make_state(num_examples, rng) for layer in layers ]
File "C:\\Python27\\lib\\site-packages\\pylearn2-0.1dev-py2.7.egg\\pylearn2\\models\\dbm\\__init__.py", line 1507, in make_state
(h_state , h_sample)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function.py", line 221, in function
profile=profile)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\pfunc.py", line 506, in pfunc
on_unused_input=on_unused_input)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function_module.py", line 1339, in orig_function
defaults)
File "C:\\Python27\\lib\\site-packages\\theano\\compile\\function_module.py", line 1167, in create
_fn, _i, _o = self.linker.make_thunk(input_storage=input_storage_lists)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\link.py", line 383, in make_thunk
output_storage = output_storage)[:3]
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\vm.py", line 800, in make_all
for node in order]
File "C:\\Python27\\lib\\site-packages\\theano\\sandbox\\cuda\\__init__.py", line 244, in make_thunk
compute_map, no_recycling)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\op.py", line 580, in make_thunk
output_storage=node_output_storage)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 913, in make_thunk
keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 856, in __compile__
keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 1279, in cthunk_factory
key=key, fn=self.compile_cmodule_by_step, keep_lock=keep_lock)
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cmodule.py", line 975, in module_from_key
module = compile_steps.next()
File "C:\\Python27\\lib\\site-packages\\theano\\gof\\cc.py", line 1201, in compile_cmodule_by_step
preargs=preargs)
File "C:\\Python27\\lib\\site-packages\\theano\\sandbox\\cuda\\nvcc_compiler.py", line 391, in compile_str
'for cmd', ' '.join(cmd))
Exception: ('nvcc return status', 2, 'for cmd', 'nvcc -shared -g -O3 -arch=sm_20 -Xlinker /DEBUG -m64 -Xcompiler -DCUDA_NDARRAY_CUH=90c42b258281d604500ad1a5624b6654,-D NPY_ARRAY_ENSURECOPY=NPY_ENSURECOPY,-D NPY_ARRAY_ALIGNED=NPY_ALIGNED,-D NPY_ARRAY_WRITEABLE=NPY_WRITEABLE,-D NPY_ARRAY_UPDATE_ALL=NPY_UPDATE_ALL,-D NPY_ARRAY_C_CONTIGUOUS=NPY_C_CONTIGUOUS,-D NPY_ARRAY_F_CONTIGUOUS=NPY_F_CONTIGUOUS,/Zi -IC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray -IAUTO\\\\include -IC:\\\\Python27\\\\lib\\\\site-packages\\\\numpy\\\\core\\\\include -IC:\\\\Python27\\\\include -IC:\\\\Python27\\\\lib\\\\site-packages\\\\theano\\\\sandbox\\\\cuda -o C:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\tmphlgfm0\\\\465ad7f8e1d438d913f57bdca2bb9eb1.pyd mod.cu -LC:\\\\Python27\\\\libs -LAUTO\\\\lib -LC:\\\\Users\\\\dspineo\\\\AppData\\\\Local\\\\Theano\\\\compiledir_Windows-7-6.1.7601-SP1-Intel64_Family_6_Model_58_Stepping_9_GenuineIntel-2.7.3\\\\cuda_ndarray -LAUTO\\\\lib -LAUTO\\\\lib64 -LC:\\\\Python27 -lpython27 -lcudart -lcublas -lcuda_ndarray', '[*1 -> GPU_mrg_uniform{CudaNdarrayType(float32, matrix),no_inplace}(<CudaNdarrayType(float32, vector)>, <TensorType(int32, vector)>), *1::1]')
\r\n
\r\n
\r\n
", "rawMarkdown": "", "votes": null }, { "id": "23968", "postDate": "05/07/2013 01:04:06", "content": "
Looks like a bug in theano. Can you e-mail me the full yaml file?
\r\nAlso, if you're not using the latest version of theano, try updating. It's possible this bug has already been fixed in the theano github repository.
", "rawMarkdown": "", "votes": null }, { "id": "23972", "postDate": "05/07/2013 02:29:46", "content": "Updating to the latest fixed it. I was using the Gohlke build of 0.6.0rc3 previously.
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 23968, "author_name": "iangoodfellow", "author_url": "", "post_date": "05/07/2013 01:04:06", "content": "Looks like a bug in theano. Can you e-mail me the full yaml file?
\r\nAlso, if you're not using the latest version of theano, try updating. It's possible this bug has already been fixed in the theano github repository.
", "votes": null, "replies": [] }, { "id": 23972, "author_name": "dpineo", "author_url": "", "post_date": "05/07/2013 02:29:46", "content": "Updating to the latest fixed it. I was using the Gohlke build of 0.6.0rc3 previously.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "23967": "", "23968": "", "23972": "" }, "source": "meta" }