jikechao opened a new pull request, #14724:
URL: https://github.com/apache/tvm/pull/14724
This pr fixes 2 questions.
1. The wrong documentation about attribute `layout` in `from_keras()`.
By analysis the source code
[Here](https://github.com/jikechao/tvm/blob/main_abc/python/tvm/relay/frontend/keras.py#:~:text=assert%20layout%20in,NHWC%20or%20NDHWC%22),
we can find that `input_layout` has 4 valid values (e.g, 'NWC', 'NCHW', NHWC
and NDHWC"). However, in the [source code
comments](https://github.com/jikechao/tvm/blob/main_abc/python/tvm/relay/frontend/keras.py#:~:text=layout%3A%20str,better%20across%20TVM.)
and [official
documentation](https://tvm.apache.org/docs/reference/api/python/relay/frontend.html?highlight=from_keras#tvm.relay.frontend.from_keras:~:text=layout%20(str)%20%E2%80%93%20One%20of%20%E2%80%98NCHW%E2%80%99%20or%20%E2%80%98NHWC%E2%80%99%2C%20indicates%20how%20data%20should%20be%20arranged%20in%20the%20output%20model.%20Default%20layout%20is%20%E2%80%98NCHW%E2%80%99%20as%20it%20in%20general%20performs%20better%20across%20TVM.)
, it says only 'nchw' and 'nhwc' are supported. They are conflicted.
2. Add exception checking to reject the mismatched `layout` and
`input_shape` immediately.
For the following script: relay can convert the keras model to RelayIR
successful, It crashed when executed the last statements (e.g., `model =
model.evaluate()`) in the script and throw: "Check failed: (!axes.defined() ||
static_cast<int>(axes.size()) == ndim) is false: Dimension mismatch: axes has 4
elements, but data.ndim = 3". This exception checking is too late, and the
crash message will increase the difficulty of debugging. Thus, I added an
exception checking in keras importer and give a clear exception message.
```
import tvm
import tvm.relay as relay
import numpy as np
from tensorflow import keras
from tensorflow.keras import layers, models
input_shape = (1, 2, 3)
input_data = np.random.random(input_shape)
x = layers.Input(shape=input_shape[1:], dtype='float32')
layer = keras.layers.Flatten()
layer.set_weights(layer.get_weights())
y = layer(x)
model = models.Model(x, y)
shape_dict = {'input_1': input_shape}
mod, params = relay.frontend.from_keras(model, shape_dict) # convert to
RelayIR successfully.
print(mod)
with tvm.transform.PassContext(opt_level=3):
model = relay.build_module.create_executor("vm", mod, tvm.cpu(0),
'llvm', params)
model = model.evaluate() # crash here!!!
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
