整个代码和错误可以在这里看到- https://www.kaggle.com/pradhyo/keras-style-transfer-different-losses/data?scriptVersionId=9726696（页面中的错误大约是60％-黑色背景的较长部分）
即使wd_float（）返回float，我认为下面的问题是_loss似乎是两倍。
我尝试从wd_float（）返回int，还尝试在style_loss_wasserstein（）中强制转换为int，以查看发生了什么，并且错误变为“ pyfunc_0返回的第0个值是int，但是期望float”。所以我最好的猜测是tf.py_func正在某种程度上将wd_float的输出转换为double。
from scipy.stats import wasserstein_distance
# https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.wasserstein_distance.html
# wasserstein_distance is supposed to return float based on docs
def wd_float(x, y):
_wd = wasserstein_distance(x, y)
return float(_wd)
def style_loss_wasserstein():
...
_loss = tf.py_func(wd_float, [style_features, combination_features], tf.float32)
_loss = K.cast(_loss, dtype='float32')
print(f"first loss: {_loss}")
return tf.convert_to_tensor(_loss, tf.float32)
...
from keras import backend as K
from scipy.optimize import fmin_l_bfgs_b
class Evaluator(object):
...
def eval_loss_and_grads(self, x):
...
outs = self.f_outputs([x]) # line causing the error
loss_value = outs[0]
...
def loss(self, x):
assert self.loss_value is None
loss_value, grad_values = self.eval_loss_and_grads(x)
self.loss_value = loss_value
self.grad_values = grad_values
return self.loss_value
outputs = [style_loss_wasserstein]
f_outputs = K.function([combination_image], outputs)
evaluator = Evaluator(f_outputs, img_nrows, img_ncols)
x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),
fprime=evaluator.grads, maxfun=20)
这是堆栈跟踪
---------------------------------------------------------------------------
InvalidArgumentError Traceback (most recent call last)
<ipython-input-24-fcc21e1fc999> in <module>()
13 loss_fn,
14 iterations,
---> 15 save_every)
16
17 # Display each output iteration for a style
<ipython-input-16-2397e37b5e4f> in style_transfer(base_image_path, style_reference_image_path, result_prefix, loss_fn, iterations, save_every)
47 for i in range(iterations):
48 x, min_val, info = fmin_l_bfgs_b(evaluator.loss, x.flatten(),
---> 49 fprime=evaluator.grads, maxfun=20)
50 print('Iteration ' + str(i) + ' loss value:', min_val)
51 # save current generated image
/opt/conda/lib/python3.6/site-packages/scipy/optimize/lbfgsb.py in fmin_l_bfgs_b(func, x0, fprime, args, approx_grad, bounds, m, factr, pgtol, epsilon, iprint, maxfun, maxiter, disp, callback, maxls)
197
198 res = _minimize_lbfgsb(fun, x0, args=args, jac=jac, bounds=bounds,
--> 199 **opts)
200 d = {'grad': res['jac'],
201 'task': res['message'],
/opt/conda/lib/python3.6/site-packages/scipy/optimize/lbfgsb.py in _minimize_lbfgsb(fun, x0, args, jac, bounds, disp, maxcor, ftol, gtol, eps, maxfun, maxiter, iprint, callback, maxls, **unknown_options)
333 # until the completion of the current minimization iteration.
334 # Overwrite f and g:
--> 335 f, g = func_and_grad(x)
336 elif task_str.startswith(b'NEW_X'):
337 # new iteration
/opt/conda/lib/python3.6/site-packages/scipy/optimize/lbfgsb.py in func_and_grad(x)
283 else:
284 def func_and_grad(x):
--> 285 f = fun(x, *args)
286 g = jac(x, *args)
287 return f, g
/opt/conda/lib/python3.6/site-packages/scipy/optimize/optimize.py in function_wrapper(*wrapper_args)
291 def function_wrapper(*wrapper_args):
292 ncalls[0] += 1
--> 293 return function(*(wrapper_args + args))
294
295 return ncalls, function_wrapper
<ipython-input-14-687064fde378> in loss(self, x)
29 def loss(self, x):
30 assert self.loss_value is None
---> 31 loss_value, grad_values = self.eval_loss_and_grads(x)
32 self.loss_value = loss_value
33 self.grad_values = grad_values
<ipython-input-14-687064fde378> in eval_loss_and_grads(self, x)
19 else:
20 x = x.reshape((1, self.img_nrows, self.img_ncols, 3))
---> 21 outs = self.f_outputs([x])
22 loss_value = outs[0]
23 if len(outs[1:]) == 1:
/opt/conda/lib/python3.6/site-packages/Keras-2.2.4-py3.6.egg/keras/backend/tensorflow_backend.py in __call__(self, inputs)
2919 return self._legacy_call(inputs)
2920
-> 2921 return self._call(inputs)
2922 else:
2923 if py_any(is_tensor(x) for x in inputs):
/opt/conda/lib/python3.6/site-packages/Keras-2.2.4-py3.6.egg/keras/backend/tensorflow_backend.py in _call(self, inputs)
2877 fetched = self._callable_fn(*array_vals, run_metadata=self.run_metadata)
2878 else:
-> 2879 fetched = self._callable_fn(*array_vals)
2880 return fetched[:len(self.outputs)]
2881
/opt/conda/lib/python3.6/site-packages/tensorflow/python/client/session.py in __call__(self, *args, **kwargs)
1437 ret = tf_session.TF_SessionRunCallable(
1438 self._session._session, self._handle, args, status,
-> 1439 run_metadata_ptr)
1440 if run_metadata:
1441 proto_data = tf_session.TF_GetBuffer(run_metadata_ptr)
/opt/conda/lib/python3.6/site-packages/tensorflow/python/framework/errors_impl.py in __exit__(self, type_arg, value_arg, traceback_arg)
526 None, None,
527 compat.as_text(c_api.TF_Message(self.status.status)),
--> 528 c_api.TF_GetCode(self.status.status))
529 # Delete the underlying status object from memory otherwise it stays alive
530 # as there is a reference to status from this from the traceback due to
InvalidArgumentError: 0-th value returned by pyfunc_0 is double, but expects float
[[{{node PyFunc}} = PyFunc[Tin=[DT_FLOAT, DT_FLOAT], Tout=[DT_FLOAT], token="pyfunc_0", _device="/job:localhost/replica:0/task:0/device:CPU:0"](Reshape_20/_609, Reshape_21/_611)]]
[[{{node PyFunc/_613}} = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_103_PyFunc", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]()]]
答案 0 :(得分：0)
我认为这里的问题是TensorFlow参数值将tensorflow图保留为int，但是您尝试使用float返回图。可以将tf.py_func中的参数从tf.float32更改为tf.int（或tf.int32，无论在pyhton中如何工作）都可以解决问题。
答案 1 :(得分：0)
今天我也遇到了类似的问题，但是在TF2中的自定义损失函数中。
对我来说，解决方案是在使用自定义numpy函数返回数据之前，先使用np.ndarray.astype函数将数据转换为np.float32类型，例如：
def custom_loss_function_loss2(y_true, y_pred):
loss = np.random.random(50)
return np.ndarray.astype(loss, np.float32)
@tf.function
def custom_loss_function(y_true, y_pred):
loss1 = tf.keras.losses.binary_crossentropy(y_true, y_pred)
loss2 = tf.numpy_function(custom_loss_function_loss2, [y_true, y_pred], np.float32)
return loss1 + loss2
希望以后能以相同的错误帮助其他人。
答案 2 :(得分：0)
当我在tensorflow中使用py_func函数时，我也遇到了这个问题。
我通过将np.float修改为np.float32来解决