I’m trying to use a working ‘channel stacked’ model with relay. This dies on the first Conv2D after a Concatenate and then on a Dense after a Flatten. It seems to compute the shapes incorrectly.
Most interesting line (1/2) in error:
%2 = nn.conv2d(%1, %v_param_1, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) in particular dimension 1 conflicts 100 does not match 2; unable to unify: Tensor[(64, 100, 3, 3), float32]andTensor[(64, 2, 3, 3), float32]; ;
Line (2/2) in error - later:
%6 = nn.dense(%5, %v_param_3, units=256) in particular dimension 1 conflicts (int64)3200 does not match 160000; unable to unify: Tensor[(256, 3200), float32] and Tensor[(256, 160000), float32]; ;
Btw, I also tried this w/ my inputs as 1,50,50 (as opposed to 50,50,1 below) and the errors were similar.
This seemed vaguely similar to another error I had seen here on Keras - so I used their format to reproduce…
Thanks in advance!!! p
Here’s my example:
import numpy as np import keras from keras.layers import Input, Concatenate, Conv2D, Flatten, Dense from keras.models import Model # Simple Keras model X1 = Input(shape=(50, 50, 1), name="X1") X2 = Input(shape=(50, 50, 1), name="X2") x = Concatenate()([X1,X2]) x = Conv2D(64,kernel_size=(3,3),padding='same')(x) # 64 filters model = keras.Model(inputs = [X1,X2], outputs = x, name='generator') # normally more of the model here... x = Flatten()(x) x = Dense(256,activation='relu')(x) # The keras model works fine d1 = np.zeros([1,50,50,1]) d2 = np.zeros([1,50,50,1]) P = model.predict([d1,d2]) import tvm import tvm.relay as relay shape_dict = {'X1': [1,50,50,1], 'X2': [1,50,50,1] } mod, params = relay.frontend.from_keras(model, shape_dict)
Output of last: --------------------------------------------------------------------------- TVMError Traceback (most recent call last) in 3 4 shape_dict = {‘X1’: [1,50,50,1], ‘X2’: [1,50,50,1] } ----> 5 mod, params = relay.frontend.from_keras(model, shape_dict)
~/repos/tvm/python/tvm/relay/frontend/keras.py in from_keras(model, shape, layout)
917 func = _expr.Function(analysis.free_vars(outexpr), outexpr)
918 params = {k:_nd.array(np.array(v, dtype=np.float32)) for k, v in etab.params.items()}
--> 919 return IRModule.from_expr(func), params
~/repos/tvm/python/tvm/ir/module.py in from_expr(expr, functions, type_defs)
221 funcs = functions if functions is not None else {}
222 defs = type_defs if type_defs is not None else {}
--> 223 return _ffi_api.Module_FromExpr(expr, funcs, defs)
224
225 def _import(self, file_to_import):
~/repos/tvm/python/tvm/_ffi/_ctypes/packed_func.py in __call__(self, *args)
211 self.handle, values, tcodes, ctypes.c_int(num_args),
212 ctypes.byref(ret_val), ctypes.byref(ret_tcode)) != 0:
--> 213 raise get_last_ffi_error()
214 _ = temp_args
215 _ = args
TVMError: Traceback (most recent call last):
[bt] (8) /home/ace/repos/tvm/build/libtvm.so(TVMFuncCall+0x65) [0x7f4ba7023715]
[bt] (7) /home/ace/repos/tvm/build/libtvm.so(+0x2971ba) [0x7f4ba68481ba]
[bt] (6) /home/ace/repos/tvm/build/libtvm.so(tvm::IRModule::FromExpr(tvm::RelayExpr const&, tvm::Map<tvm::GlobalVar, tvm::BaseFunc, void, void> const&, tvm::Map<tvm::GlobalTypeVar, tvm::TypeData, void, void> const&)+0x23b) [0x7f4ba684447b]
[bt] (5) /home/ace/repos/tvm/build/libtvm.so(tvm::IRModuleNode::Add(tvm::GlobalVar const&, tvm::BaseFunc const&, bool)+0x43e) [0x7f4ba684405e]
[bt] (4) /home/ace/repos/tvm/build/libtvm.so(tvm::RunTypeCheck(tvm::IRModule const&, tvm::GlobalVar const&, tvm::relay::Function)+0x376) [0x7f4ba683e756]
[bt] (3) /home/ace/repos/tvm/build/libtvm.so(tvm::relay::InferType(tvm::relay::Function const&, tvm::IRModule const&, tvm::GlobalVar const&)+0x1dc) [0x7f4ba6e9c59c]
[bt] (2) /home/ace/repos/tvm/build/libtvm.so(tvm::relay::TypeInferencer::Infer(tvm::RelayExpr)+0x73) [0x7f4ba6e9bd63]
[bt] (1) /home/ace/repos/tvm/build/libtvm.so(tvm::ErrorReporter::RenderErrors(tvm::IRModule const&, bool)+0x1e16) [0x7f4ba6832b16]
[bt] (0) /home/ace/repos/tvm/build/libtvm.so(+0x23ef61) [0x7f4ba67eff61]
File "/home/ace/repos/tvm/src/ir/error.cc", line 133
TVMError:
Error(s) have occurred. The program has been annotated with them:
In `main`:
v0.0.4
fn (%X1: Tensor[(1, 50, 50, 1), float32], %X2: Tensor[(1, 50, 50, 1), float32], %v_param_1: Tensor[(64, 2, 3, 3), float32], %v_param_2: Tensor[(64), float32], %v_param_3: Tensor[(256, 160000), float32], %v_param_4: Tensor[(256), float32]) {
%0 = (%X1, %X2);
%1 = concatenate(%0, axis=1);
--> %2 = nn.conv2d(%1, %v_param_1, padding=[1, 1, 1, 1], channels=64, kernel_size=[3, 3]) in particular dimension 1 conflicts 100 does not match 2; unable to unify: `Tensor[(64, 100, 3, 3), float32]` and `Tensor[(64, 2, 3, 3), float32]`; ;
%3 = nn.bias_add(%2, %v_param_2);
%4 = transpose(%3, axes=[0, 2, 3, 1]);
%5 = nn.batch_flatten(%4);
--> %6 = nn.dense(%5, %v_param_3, units=256) in particular dimension 1 conflicts (int64)3200 does not match 160000; unable to unify: `Tensor[(256, 3200), float32]` and `Tensor[(256, 160000), float32]`; ;
%7 = nn.bias_add(%6, %v_param_4);
nn.relu(%7)
}