tensor forest estimator value error at fitting the training part - tensorflow

Question

Code :
from sklearn import cross_validation as cv
import numpy as np
from tensorflow.contrib.learn.python.learn.estimators import estimator
from tensorflow.contrib.tensor_forest.python import tensor_forest
X = np.array([[ 74., 166., 331., 161., 159., 181., 180.],
[ 437., 427., 371., 361., 393., 465., 464.],
[ 546., 564., 588., 595., 536., 537., 520.],
[ 89., 89., 87., 87., 108., 113., 111.],
[ 75., 90., 74., 89., 130., 140., 135.]])
Y = np.array([[ 51., 43., 29., 43., 43., 41., 42.],
[ 22., 23., 26., 27., 25., 19., 19.],
[ 7., 7., 5., 5., 9., 8., 10.],
[ 55., 54., 55., 53., 51., 51., 51.],
[ 58., 57., 57., 58., 55., 55., 55.]])
train_X, test_X, train_Y, test_Y = cv.train_test_split(X, Y,
test_size=0.50, random_state=42)
def build_estimator() :
params = tensor_forest.ForestHParams(num_classes=7, num_features=7,
num_trees=30, max_nodes=100)
graph_builder_class = tensor_forest.RandomForestGraphs
graph_builder_class = tensor_forest.TrainingLossForest
return estimator.SKCompat(random_forest.TensorForestEstimator(
params, graph_builder_class=graph_builder_class,
model_dir=None))
est = build_estimator()
train_X = train_X.astype(dtype = np.float32)
train_Y = train_Y.astype(dtype = np.float32)
est = est.fit(x=train_X, y=train_Y, batch_size = 100)
My both input and output shape is [Number_of_samples,7]. It perfectly runs with scikitlearn random forest classifier. However for tf learn, I get the following error when fitting the estimator :
INFO:tensorflow:Constructing forest with params =
INFO:tensorflow:{'valid_leaf_threshold': 1, 'split_after_samples':
250, 'num_output_columns': 8, 'feature_bagging_fraction': 1.0,
'split_initializations_per_input': 1, 'bagged_features': None,
'min_split_samples': 5, 'max_nodes': 100, 'num_features': 7,
'num_trees': 30, 'num_splits_to_consider': 7, 'base_random_seed': 0,
'num_outputs': 1, 'dominate_fraction': 0.99, 'max_fertile_nodes': 50,
'bagged_num_features': 7, 'dominate_method': 'bootstrap',
'bagging_fraction': 1.0, 'regression': False, 'num_classes': 7}
ValueErrorTraceback (most recent call last)
in ()
----> 1 est = est.fit(x=train_X, y=train_Y, batch_size = 100)
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in fit(self, x, y, batch_size, steps, max_steps, monitors) 1351
steps=steps, 1352 max_steps=max_steps,
-> 1353 monitors=all_monitors) 1354 return self 1355
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/client/random_forest.pyc
in fit(self, x, y, input_fn, steps, batch_size, monitors, max_steps)
262 elif input is not None:
263 self._estimator.fit(input_fn=input_fn, steps=steps, monitors=monitors,
--> 264 max_steps=max_steps)
265 else:
266 raise ValueError('fit: Must provide either both x and y or input_fn.')
/usr/local/lib/python2.7/dist-packages/tensorflow/python/util/deprecation.pyc
in new_func(*args, **kwargs)
278 _call_location(), decorator_utils.get_qualified_name(func),
279 func.module, arg_name, date, instructions)
--> 280 return func(*args, **kwargs)
281 new_func.doc = _add_deprecated_arg_notice_to_docstring(
282 func.doc, date, instructions)
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in fit(self, x, y, input_fn, steps, batch_size, monitors, max_steps)
424 hooks.append(basic_session_run_hooks.StopAtStepHook(steps, max_steps))
425
--> 426 loss = self._train_model(input_fn=input_fn, hooks=hooks)
427 logging.info('Loss for final step: %s.', loss)
428 return self
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in _train_model(self, input_fn, hooks)
932 features, labels = input_fn()
933 self._check_inputs(features, labels)
--> 934 model_fn_ops = self._call_legacy_get_train_ops(features, labels)
935 ops.add_to_collection(ops.GraphKeys.LOSSES, model_fn_ops.loss)
936 all_hooks.extend([
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in _call_legacy_get_train_ops(self, features, labels) 1001 1002
def _call_legacy_get_train_ops(self, features, labels):
-> 1003 train_ops = self._get_train_ops(features, labels) 1004 if isinstance(train_ops, model_fn_lib.ModelFnOps): # Default
signature 1005 return train_ops
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in _get_train_ops(self, features, labels) 1160 ModelFnOps
object. 1161 """
-> 1162 return self._call_model_fn(features, labels, model_fn_lib.ModeKeys.TRAIN) 1163 1164 def
_get_eval_ops(self, features, labels, metrics):
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/learn/python/learn/estimators/estimator.pyc
in _call_model_fn(self, features, labels, mode) 1131 if
'model_dir' in model_fn_args: 1132 kwargs['model_dir'] =
self.model_dir
-> 1133 model_fn_results = self._model_fn(features, labels, **kwargs) 1134 1135 if isinstance(model_fn_results, model_fn_lib.ModelFnOps):
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/client/random_forest.pyc
in _model_fn(features, labels)
128 if labels is not None:
129 training_loss = graph_builder.training_loss(
--> 130 features, labels, name=LOSS_NAME)
131 training_graph = control_flow_ops.group(
132 graph_builder.training_graph(
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/python/tensor_forest.pyc
in training_loss(self, features, labels, name)
559
560 def training_loss(self, features, labels, name='training_loss'):
--> 561 return array_ops.identity(self._get_loss(features, labels), name=name)
562
563
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/python/tensor_forest.pyc
in _get_loss(self, features, labels)
548 self._loss = control_flow_ops.cond(
549 self.average_size() > 0, _average_loss,
--> 550 lambda: constant_op.constant(sys.maxsize, dtype=dtypes.float32))
551
552 return self._loss
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.pyc
in cond(pred, fn1, fn2, name) 1757 context_t =
CondContext(pred, pivot_1, branch=1) 1758 context_t.Enter()
-> 1759 orig_res, res_t = context_t.BuildCondBranch(fn1) 1760 context_t.ExitResult(res_t) 1761 context_t.Exit()
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/control_flow_ops.pyc
in BuildCondBranch(self, fn) 1658 def BuildCondBranch(self, fn):
1659 """Add the subgraph defined by fn() to the graph."""
-> 1660 r = fn() 1661 original_r = r 1662 result = []
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/python/tensor_forest.pyc
in _average_loss()
544 probs = self.inference_graph(features)
545 return math_ops.reduce_sum(self.loss_fn(
--> 546 probs, labels)) / math_ops.to_float(array_ops.shape(labels)[0])
547
548 self._loss = control_flow_ops.cond(
/usr/local/lib/python2.7/dist-packages/tensorflow/contrib/tensor_forest/python/tensor_forest.pyc
in _loss(probs, targets)
508 def _loss(probs, targets):
509 if targets.get_shape().ndims > 1:
--> 510 targets = array_ops.squeeze(targets, squeeze_dims=[1])
511 one_hot_labels = array_ops.one_hot(
512 math_ops.to_int32(targets),
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/array_ops.pyc
in squeeze(input, axis, name, squeeze_dims) 2270 if
np.isscalar(axis): 2271 axis = [axis]
-> 2272 return gen_array_ops._squeeze(input, axis, name) 2273 2274
/usr/local/lib/python2.7/dist-packages/tensorflow/python/ops/gen_array_ops.pyc
in _squeeze(input, squeeze_dims, name) 3496 """ 3497 result
= _op_def_lib.apply_op("Squeeze", input=input,
-> 3498 squeeze_dims=squeeze_dims, name=name) 3499 return result 3500
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/op_def_library.pyc
in apply_op(self, op_type_name, name, **keywords)
761 op = g.create_op(op_type_name, inputs, output_types, name=scope,
762 input_types=input_types, attrs=attr_protos,
--> 763 op_def=op_def)
764 if output_structure:
765 outputs = op.outputs
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc
in create_op(self, op_type, inputs, dtypes, input_types, name, attrs,
op_def, compute_shapes, compute_device) 2395
original_op=self._default_original_op, op_def=op_def) 2396 if
compute_shapes:
-> 2397 set_shapes_for_outputs(ret) 2398 self._add_op(ret) 2399
self._record_op_seen_by_control_dependencies(ret)
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc
in set_shapes_for_outputs(op) 1755 shape_func =
_call_cpp_shape_fn_and_require_op 1756
-> 1757 shapes = shape_func(op) 1758 if shapes is None: 1759 raise RuntimeError(
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/ops.pyc
in call_with_requiring(op) 1705 1706 def
call_with_requiring(op):
-> 1707 return call_cpp_shape_fn(op, require_shape_fn=True) 1708 1709 _call_cpp_shape_fn_and_require_op =
call_with_requiring
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/common_shapes.pyc
in call_cpp_shape_fn(op, input_tensors_needed,
input_tensors_as_shapes_needed, debug_python_shape_fn,
require_shape_fn)
608 res = _call_cpp_shape_fn_impl(op, input_tensors_needed,
609 input_tensors_as_shapes_needed,
--> 610 debug_python_shape_fn, require_shape_fn)
611 if not isinstance(res, dict):
612 # Handles the case where _call_cpp_shape_fn_impl calls unknown_shape(op).
/usr/local/lib/python2.7/dist-packages/tensorflow/python/framework/common_shapes.pyc
in _call_cpp_shape_fn_impl(op, input_tensors_needed,
input_tensors_as_shapes_needed, debug_python_shape_fn,
require_shape_fn)
673 missing_shape_fn = True
674 else:
--> 675 raise ValueError(err.message)
676
677 if missing_shape_fn:
ValueError: Can not squeeze dim[1], expected a dimension of 1, got 7
for 'cond/Squeeze' (op: 'Squeeze') with input shapes: [?,7].

Answer 1

Right now, when using TrainingLossForest it assumes a classification problem. I will fix this internally, but for now you can get around it with something like:
from tensorflow.contrib.losses.python.losses import loss_ops
def _loss_fn(values, targets):
return loss_ops.mean_squared_error(values, targets)
def _builder_class(params, **kwargs):
return tensor_forest.TrainingLossForest(
params, loss_fn=_loss_fn, **kwargs)
TensorForestEstimator(..., graph_builder_class=_builder_class)
Or whatever loss function you want (MSE here as example). Also, simply using (graph_builder_class = tensor_forest.RandomForestGraphs) should suffice, but the loss is the number of nodes, so training stops when the forest stops growing or reaches max_nodes, which is maybe not what you want.