You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I have encountered an issue similar to 717 and 727.
At the bottom of the post, I put the exact error that I get and some logs.
Analyzing in debug, I think that the problem is during the graph creation (traverse.py listup_nodes) that adds Im2Col layer not supported by fallback backend (accordingly to answer to 717), but I can't be more specific, because I don't have deep understanding of webdnn graph representation.
These are two models very similar: one with some separable convolutions and the other only with standard convolutions. sep_vs_std_conv.zip
Only the second one is correctly converted on all backends, while the first is converted to webgpu, webgl and wasm (although it takes a lot of time). Both of them use keras fallback converter, because in TF converter "DepthwiseConv2dNative is not supported yet".
Error: [convert_keras.py] Failed generating descriptor for fallback backend Traceback (most recent call last): File "./bin/convert_keras.py", line 102, in main graph_exec_data = generate_descriptor(backend, graph, constant_encoder_name=args.encoding) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/interface/generator.py", line 107, in generate_descriptor return generator(graph, **kwargs) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/fallback/generator.py", line 75, in generate return FallbackDescriptorGenerator.generate(graph, **kwargs) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/fallback/generator.py", line 64, in generate kernels=cls.generate_kernels(graph, memory_layout), File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/interface/generator.py", line 57, in generate_kernels raise NotImplementedError(f"[{cls.__name__}] Operator {op} is not handled by any generator handler") NotImplementedError: [FallbackDescriptorGenerator] Operator <Im2Col inputs={'im': <Variable Variable10 shape=(1, 44, 44, 8), order=[N, H, W, C]>}, outputs={'col': <Variable Variable12 shape=(1, 44, 44, 3, 3, 8), order=[N, H, W, KH, KW, C]>}> is not handled by any generator handler
Subset (related to the first separable convolution) of debug outputs in traverse.py dump function:
I have encountered an issue similar to 717 and 727.
At the bottom of the post, I put the exact error that I get and some logs.
Analyzing in debug, I think that the problem is during the graph creation (traverse.py listup_nodes) that adds Im2Col layer not supported by fallback backend (accordingly to answer to 717), but I can't be more specific, because I don't have deep understanding of webdnn graph representation.
These are two models very similar: one with some separable convolutions and the other only with standard convolutions.
sep_vs_std_conv.zip
Only the second one is correctly converted on all backends, while the first is converted to webgpu, webgl and wasm (although it takes a lot of time). Both of them use keras fallback converter, because in TF converter "DepthwiseConv2dNative is not supported yet".
Error:
[convert_keras.py] Failed generating descriptor for fallback backend Traceback (most recent call last): File "./bin/convert_keras.py", line 102, in main graph_exec_data = generate_descriptor(backend, graph, constant_encoder_name=args.encoding) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/interface/generator.py", line 107, in generate_descriptor return generator(graph, **kwargs) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/fallback/generator.py", line 75, in generate return FallbackDescriptorGenerator.generate(graph, **kwargs) File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/fallback/generator.py", line 64, in generate kernels=cls.generate_kernels(graph, memory_layout), File "/Users/michele/Anaconda2/anaconda/envs/webdnn_converter/lib/python3.6/site-packages/webdnn-1.2.2-py3.6.egg/webdnn/backend/interface/generator.py", line 57, in generate_kernels raise NotImplementedError(f"[{cls.__name__}] Operator {op} is not handled by any generator handler") NotImplementedError: [FallbackDescriptorGenerator] Operator <Im2Col inputs={'im': <Variable Variable10 shape=(1, 44, 44, 8), order=[N, H, W, C]>}, outputs={'col': <Variable Variable12 shape=(1, 44, 44, 3, 3, 8), order=[N, H, W, KH, KW, C]>}> is not handled by any generator handler
Subset (related to the first separable convolution) of debug outputs in traverse.py dump function:
Keras model architecture:
The text was updated successfully, but these errors were encountered: