search

FrozenGene / tvm-tutorial

TVM tutorial
65 stars 10 forks source link

tflite frontend fail with NCHW #1

Open Li-wuhan opened 4 years ago

Li-wuhan commented 4 years ago

relay.frontend.from_tflite() fail, seems NCHW not supported. Have a try with modifying run_tvm.py to NHWC and can relay.frontend.from_tflite() can pass, is it to say we cannot support NCHW when convert TFLite?

root@76802c36f63b:~/tvm_test/tvm-tutorial# python3 run_tvm.py Traceback (most recent call last):

File "run_tvm.py", line 73, in tvm_output = run(model_file, image_data)

File "run_tvm.py", line 32, in run func, params = relay.frontend.from_tflite(tflite_model, shape_dict={input_tensor: input_shape}, dtype_dict={input_tensor: input_dtype})

File "/root/tvm/python/tvm/relay/frontend/tflite.py", line 1386, in from_tflite mod = _module.Module.from_expr(func)

File "/root/tvm/python/tvm/relay/module.py", line 233, in from_expr return _module.Module_FromExpr(expr, funcs, defs)

File "/root/tvm/python/tvm/_ffi/_ctypes/function.py", line 211, in call raise get_last_ffi_error()

tvm._ffi.base.TVMError: Traceback (most recent call last): [bt] (7) /root/tvm/build/libtvm.so(TVMFuncCall+0x65) [0x7f10176b17d5] [bt] (6) /root/tvm/build/libtvm.so(+0x11b5c0e) [0x7f10175c9c0e] [bt] (5) /root/tvm/build/libtvm.so(tvm::relay::ModuleNode::FromExpr(tvm::relay::Expr const&, tvm::Map<tvm::relay::GlobalVar, tvm::relay::Function, void, void> const&, tvm::Map<tvm::relay::GlobalTypeVar, tvm::relay::TypeData, void, void> const&)+0x167) [0x7f10175c7017] [bt] (4) /root/tvm/build/libtvm.so(tvm::relay::ModuleNode::Add(tvm::relay::GlobalVar const&, tvm::relay::Function const&, bool)+0x58b) [0x7f10175c665b] [bt] (3) /root/tvm/build/libtvm.so(tvm::relay::InferType(tvm::relay::Function const&, tvm::relay::Module const&, tvm::relay::GlobalVar const&)+0x259) [0x7f10174d32f9] [bt] (2) /root/tvm/build/libtvm.so(tvm::relay::TypeInferencer::Infer(tvm::relay::Expr)+0x71) [0x7f10174d2b01] [bt] (1) /root/tvm/build/libtvm.so(tvm::relay::ErrorReporter::RenderErrors(tvm::relay::Module const&, bool)+0x181e) [0x7f101759d4ae] [bt] (0) /root/tvm/build/libtvm.so(dmlc::LogMessageFatal::~LogMessageFatal()+0x43) [0x7f1016efef23] File "/root/tvm/src/relay/ir/error.cc", line 133 TVMError: Error(s) have occurred. The program has been annotated with them:

In main: v0.0.4 fn (%input: Tensor[(1, 3, 224, 224), float32], %v_param_1: Tensor[(3, 3, 3, 32), float32], %v_param_2: Tensor[(32), float32], %v_param_3: Tensor[(3, 3, 32, 1), float32], %v_param_4: Tensor[(32), float32], %v_param_5: Tensor[(1, 1, 32, 64), float32], %v_param_6: Tensor[(64), float32], %v_param_7: Tensor[(3, 3, 64, 1), float32], %v_param_8: Tensor[(64), float32], %v_param_9: Tensor[(1, 1, 64, 128), float32], %v_param_10: Tensor[(128), float32], %v_param_11: Tensor[(3, 3, 128, 1), float32], %v_param_12: Tensor[(128), float32], %v_param_13: Tensor[(1, 1, 128, 128), float32], %v_param_14: Tensor[(128), float32], %v_param_15: Tensor[(3, 3, 128, 1), float32], %v_param_16: Tensor[(128), float32], %v_param_17: Tensor[(1, 1, 128, 256), float32], %v_param_18: Tensor[(256), float32], %v_param_19: Tensor[(3, 3, 256, 1), float32], %v_param_20: Tensor[(256), float32], %v_param_21: Tensor[(1, 1, 256, 256), float32], %v_param_22: Tensor[(256), float32], %v_param_23: Tensor[(3, 3, 256, 1), float32], %v_param_24: Tensor[(256), float32], %v_param_25: Tensor[(1, 1, 256, 512), float32], %v_param_26: Tensor[(512), float32], %v_param_27: Tensor[(3, 3, 512, 1), float32], %v_param_28: Tensor[(512), float32], %v_param_29: Tensor[(1, 1, 512, 512), float32], %v_param_30: Tensor[(512), float32], %v_param_31: Tensor[(3, 3, 512, 1), float32], %v_param_32: Tensor[(512), float32], %v_param_33: Tensor[(1, 1, 512, 512), float32], %v_param_34: Tensor[(512), float32], %v_param_35: Tensor[(3, 3, 512, 1), float32], %v_param_36: Tensor[(512), float32], %v_param_37: Tensor[(1, 1, 512, 512), float32], %v_param_38: Tensor[(512), float32], %v_param_39: Tensor[(3, 3, 512, 1), float32], %v_param_40: Tensor[(512), float32], %v_param_41: Tensor[(1, 1, 512, 512), float32], %v_param_42: Tensor[(512), float32], %v_param_43: Tensor[(3, 3, 512, 1), float32], %v_param_44: Tensor[(512), float32], %v_param_45: Tensor[(1, 1, 512, 512), float32], %v_param_46: Tensor[(512), float32], %v_param_47: Tensor[(3, 3, 512, 1), float32], %v_param_48: Tensor[(512), float32], %v_param_49: Tensor[(1, 1, 512, 1024), float32], %v_param_50: Tensor[(1024), float32], %v_param_51: Tensor[(3, 3, 1024, 1), float32], %v_param_52: Tensor[(1024), float32], %v_param_53: Tensor[(1, 1, 1024, 1024), float32], %v_param_54: Tensor[(1024), float32], %v_param_55: Tensor[(1, 1, 1024, 1001), float32], %v_param_56: Tensor[(1001), float32]) { %0 = nn.pad(%input, pad_width=[[0, 0], [0, 1], [0, 1], [0, 0]]); %1 = nn.conv2d(%0, %v_param_1, strides=[2, 2], channels=32, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWIO") in particular dimension 2 conflicts 224 does not match 3; unable to unify: Tensor[(3, 3, 224, 32), float32] and Tensor[(3, 3, 3, 32), float32]; ; %2 = nn.bias_add(%1, %v_param_2, axis=3); %3 = clip(%2, a_min=0f, a_max=6f); %4 = nn.pad(%3, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %5 = nn.conv2d(%4, %v_param_3, groups=32, channels=32, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %6 = nn.bias_add(%5, %v_param_4, axis=3); %7 = clip(%6, a_min=0f, a_max=6f); %8 = nn.conv2d(%7, %v_param_5, channels=64, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %9 = nn.bias_add(%8, %v_param_6, axis=3); %10 = clip(%9, a_min=0f, a_max=6f); %11 = nn.pad(%10, pad_width=[[0, 0], [0, 1], [0, 1], [0, 0]]); %12 = nn.conv2d(%11, %v_param_7, strides=[2, 2], groups=64, channels=64, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %13 = nn.bias_add(%12, %v_param_8, axis=3); %14 = clip(%13, a_min=0f, a_max=6f); %15 = nn.conv2d(%14, %v_param_9, channels=128, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %16 = nn.bias_add(%15, %v_param_10, axis=3); %17 = clip(%16, a_min=0f, a_max=6f); %18 = nn.pad(%17, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %19 = nn.conv2d(%18, %v_param_11, groups=128, channels=128, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %20 = nn.bias_add(%19, %v_param_12, axis=3); %21 = clip(%20, a_min=0f, a_max=6f); %22 = nn.conv2d(%21, %v_param_13, channels=128, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %23 = nn.bias_add(%22, %v_param_14, axis=3); %24 = clip(%23, a_min=0f, a_max=6f); %25 = nn.pad(%24, pad_width=[[0, 0], [0, 1], [0, 1], [0, 0]]); %26 = nn.conv2d(%25, %v_param_15, strides=[2, 2], groups=128, channels=128, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %27 = nn.bias_add(%26, %v_param_16, axis=3); %28 = clip(%27, a_min=0f, a_max=6f); %29 = nn.conv2d(%28, %v_param_17, channels=256, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %30 = nn.bias_add(%29, %v_param_18, axis=3); %31 = clip(%30, a_min=0f, a_max=6f); %32 = nn.pad(%31, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %33 = nn.conv2d(%32, %v_param_19, groups=256, channels=256, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %34 = nn.bias_add(%33, %v_param_20, axis=3); %35 = clip(%34, a_min=0f, a_max=6f); %36 = nn.conv2d(%35, %v_param_21, channels=256, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %37 = nn.bias_add(%36, %v_param_22, axis=3); %38 = clip(%37, a_min=0f, a_max=6f); %39 = nn.pad(%38, pad_width=[[0, 0], [0, 1], [0, 1], [0, 0]]); %40 = nn.conv2d(%39, %v_param_23, strides=[2, 2], groups=256, channels=256, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %41 = nn.bias_add(%40, %v_param_24, axis=3); %42 = clip(%41, a_min=0f, a_max=6f); %43 = nn.conv2d(%42, %v_param_25, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %44 = nn.bias_add(%43, %v_param_26, axis=3); %45 = clip(%44, a_min=0f, a_max=6f); %46 = nn.pad(%45, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %47 = nn.conv2d(%46, %v_param_27, groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %48 = nn.bias_add(%47, %v_param_28, axis=3); %49 = clip(%48, a_min=0f, a_max=6f); %50 = nn.conv2d(%49, %v_param_29, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %51 = nn.bias_add(%50, %v_param_30, axis=3); %52 = clip(%51, a_min=0f, a_max=6f); %53 = nn.pad(%52, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %54 = nn.conv2d(%53, %v_param_31, groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %55 = nn.bias_add(%54, %v_param_32, axis=3); %56 = clip(%55, a_min=0f, a_max=6f); %57 = nn.conv2d(%56, %v_param_33, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %58 = nn.bias_add(%57, %v_param_34, axis=3); %59 = clip(%58, a_min=0f, a_max=6f); %60 = nn.pad(%59, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %61 = nn.conv2d(%60, %v_param_35, groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %62 = nn.bias_add(%61, %v_param_36, axis=3); %63 = clip(%62, a_min=0f, a_max=6f); %64 = nn.conv2d(%63, %v_param_37, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %65 = nn.bias_add(%64, %v_param_38, axis=3); %66 = clip(%65, a_min=0f, a_max=6f); %67 = nn.pad(%66, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %68 = nn.conv2d(%67, %v_param_39, groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %69 = nn.bias_add(%68, %v_param_40, axis=3); %70 = clip(%69, a_min=0f, a_max=6f); %71 = nn.conv2d(%70, %v_param_41, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %72 = nn.bias_add(%71, %v_param_42, axis=3); %73 = clip(%72, a_min=0f, a_max=6f); %74 = nn.pad(%73, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %75 = nn.conv2d(%74, %v_param_43, groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %76 = nn.bias_add(%75, %v_param_44, axis=3); %77 = clip(%76, a_min=0f, a_max=6f); %78 = nn.conv2d(%77, %v_param_45, channels=512, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %79 = nn.bias_add(%78, %v_param_46, axis=3); %80 = clip(%79, a_min=0f, a_max=6f); %81 = nn.pad(%80, pad_width=[[0, 0], [0, 1], [0, 1], [0, 0]]); %82 = nn.conv2d(%81, %v_param_47, strides=[2, 2], groups=512, channels=512, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %83 = nn.bias_add(%82, %v_param_48, axis=3); %84 = clip(%83, a_min=0f, a_max=6f); %85 = nn.conv2d(%84, %v_param_49, channels=1024, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %86 = nn.bias_add(%85, %v_param_50, axis=3); %87 = clip(%86, a_min=0f, a_max=6f); %88 = nn.pad(%87, pad_width=[[0, 0], [1, 1], [1, 1], [0, 0]]); %89 = nn.conv2d(%88, %v_param_51, groups=1024, channels=1024, kernel_size=[3, 3], data_layout="NHWC", kernel_layout="HWOI"); %90 = nn.bias_add(%89, %v_param_52, axis=3); %91 = clip(%90, a_min=0f, a_max=6f); %92 = nn.conv2d(%91, %v_param_53, channels=1024, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %93 = nn.bias_add(%92, %v_param_54, axis=3); %94 = clip(%93, a_min=0f, a_max=6f); %95 = nn.avg_pool2d(%94, pool_size=[7, 7], strides=[2, 2], layout="NHWC"); %96 = nn.conv2d(%95, %v_param_55, channels=1001, kernel_size=[1, 1], data_layout="NHWC", kernel_layout="HWIO"); %97 = nn.bias_add(%96, %v_param_56, axis=3); %98 = reshape(%97, newshape=[1, 1001]); nn.softmax(%98, axis=1) }

FrozenGene commented 4 years ago

Yes. The upstream TVM TFLite frontend has been modified into NHWC. Please see this PR: https://github.com/dmlc/tvm/pull/3141 to know more background and details.

FrozenGene commented 4 years ago

Welcome pr to modify this repo's tutorial. :-)