TensorFlow Split and LRN operator support for Alexnet Model #1572
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -7,6 +7,7 @@ | |
| import numpy as np | ||
|
|
||
| import tvm | ||
| from nnvm import NNVMError | ||
| from .. import symbol as _sym | ||
| from .. import graph as _graph | ||
| from .. compiler import graph_util | ||
|
|
@@ -133,7 +134,7 @@ def _impl(inputs, attr, params): | |
| attr['strides'] = (attr['strides'][1], attr['strides'][2]) | ||
|
|
||
| # Fix padding | ||
| input_shape = attr['_input_shapes'][inputs[0]] | ||
| attr['padding'] = attr['padding'].decode("utf-8") | ||
|
|
||
| if attr['padding'] == 'VALID': | ||
|
|
@@ -142,11 +143,11 @@ def _impl(inputs, attr, params): | |
| stride_h, stride_w = attr['strides'] | ||
| kernel_h, kernel_w = attr['kernel_shape'] | ||
| if attr['data_format'] == 'NHWC': | ||
| in_h = input_shape[1] | ||
| in_w = input_shape[2] | ||
| else: | ||
| in_h = input_shape[2] | ||
| in_w = input_shape[3] | ||
|
|
||
| pad_v = _get_pad_pair(in_h, kernel_h, stride_h) | ||
| pad_h = _get_pad_pair(in_w, kernel_w, stride_w) | ||
|
|
@@ -171,28 +172,31 @@ def _impl(inputs, attr, params): | |
| def _conv(opname): | ||
| def _impl(inputs, attr, params): | ||
| attr['data_format'] = attr['data_format'].decode("utf-8") | ||
| input_shape = attr['_input_shapes'][inputs[0]] | ||
|
|
||
| # Extract kernel shape from params | ||
| if inputs[1] in attr['_input_shapes']: | ||
| weight_shape = tuple(attr['_input_shapes'][inputs[1]]) | ||
| else: | ||
| weight_shape = params[inputs[1].list_output_names()[0]].shape | ||
|
|
||
| if attr['data_format'] == 'NHWC': | ||
| kernel_h, kernel_w, _, depth_mult = weight_shape | ||
| attr['kernel_shape'] = (weight_shape[0], weight_shape[1]) | ||
| if opname == 'conv': | ||
| attr['channels'] = weight_shape[3] | ||
| else: | ||
| attr['channels'] = input_shape[3] * depth_mult | ||
|
|
||
| if 'dilations' in attr: | ||
| attr['dilations'] = (attr['dilations'][0], attr['dilations'][1]) | ||
| elif attr['data_format'] == 'NCHW': | ||
| depth_mult, _, kernel_h, kernel_w = weight_shape | ||
| attr['kernel_shape'] = (weight_shape[2], weight_shape[3]) | ||
| if opname == 'conv': | ||
| attr['channels'] = weight_shape[1] | ||
| else: | ||
| attr['channels'] = input_shape[1] * depth_mult | ||
|
|
||
| if 'dilations' in attr: | ||
| attr['dilations'] = (attr['dilations'][2], attr['dilations'][3]) | ||
|
|
@@ -215,11 +219,11 @@ def _impl(inputs, attr, params): | |
| stride_h, stride_w = attr['strides'] | ||
| kernel_h, kernel_w = attr['kernel_shape'] | ||
| if attr['data_format'] == 'NHWC': | ||
| in_h = input_shape[1] | ||
| in_w = input_shape[2] | ||
| else: | ||
| in_h = input_shape[2] | ||
| in_w = input_shape[3] | ||
|
|
||
| pad_v = _get_pad_pair(in_h, kernel_h, stride_h) | ||
| pad_h = _get_pad_pair(in_w, kernel_w, stride_w) | ||
|
|
@@ -428,10 +432,20 @@ def _impl(inputs, attr, params): | |
| ignores=['index_type', 'T'])(new_inputs, attr) | ||
| return _impl | ||
|
|
||
| def _split(): | ||
| def _impl(inputs, attr, params): | ||
| pop_node = inputs.pop(0) | ||
| axis = params[pop_node.list_output_names()[0]].asnumpy()[0] | ||
| return AttrCvt(op_name="split", | ||
| ignores=['num_split'], | ||
|
Contributor
There was a problem hiding this comment. Suggest to use transforms instead of ignores & extras.
Member
There was a problem hiding this comment. Have you reached consensus the changes here is good? |
||
| extras={'indices_or_sections':attr['num_split'], | ||
| 'axis': axis})(inputs, attr) | ||
| return _impl | ||
|
|
||
| def _lrn(): | ||
| def _impl(inputs, attr, params): | ||
| attr_new = {} | ||
| depth_radius = attr.get('depth_radius', 2) | ||
| size = (depth_radius * 2) + 1 | ||
| attr_new['axis'] = 3 # Fix axis, NHWC format | ||
| attr_new['size'] = size | ||
|
|
@@ -493,7 +507,7 @@ def _impl(inputs, attr, params): | |
| new_axis_mask = int(attr.get('new_axis_mask', 0)) | ||
| shrink_axis_mask = int(attr.get('shrink_axis_mask', 0)) | ||
| data_shape = attr['_input_shapes'][inputs[0]] | ||
| data_dim = len(data_shape) | ||
| stride_dim = len(stride) | ||
|
|
||
| def _transform_mask(stride_dim, ellipsis_mask): | ||
|
|
@@ -523,26 +537,26 @@ def _transform_mask(stride_dim, ellipsis_mask): | |
| + new_axes_after_ellipsis), data_dim) | ||
| for i in range(final_index, to_index): | ||
| m_begin[final_index] = 0 | ||
| m_end[final_index] = data_shape[final_index] | ||
| m_stride[final_index] = 1 | ||
| final_index += 1 | ||
| elif not mask & new_axis_mask: | ||
| if final_index == len(m_begin): | ||
| break | ||
| if mask & begin_mask: | ||
| m_begin[final_index] = data_shape[final_index] \ | ||
| if stride[index] < 0 else 0 | ||
| elif begin[index]: | ||
| m_begin[final_index] = begin[index] | ||
| if mask & end_mask: | ||
| m_end[final_index] = 0 if stride[index] < 0 \ | ||
| else data_shape[final_index] | ||
| elif end[index]: | ||
| m_end[final_index] = end[index] | ||
| m_stride[final_index] = stride[index] | ||
| if mask & shrink_axis_mask: | ||
| #Tensorflow make axis with shrink_axis_mask as dimension 1 | ||
| m_begin[final_index] = data_shape[final_index] + begin[index] \ | ||
| if begin[index] < 0 else begin[index] | ||
| m_end[final_index] = begin[index] + 1 | ||
| m_stride[final_index] = 1 | ||
|
|
@@ -603,8 +617,8 @@ def _impl(inputs, in_state_c, in_state_h, attr, params): | |
| forget_bias = attr.pop('forget_bias') | ||
| input_shape = attr['_input_shapes'][inputs[0]] | ||
| weight_shape = attr['_input_shapes'][inputs[3]] | ||
| batch_size, input_size = input_shape[0], input_shape[1] | ||
| num_hidden_layers = weight_shape[1] | ||
| num_hidden = num_hidden_layers // 4 | ||
|
|
||
| in_data = _sym.reshape(in_data, | ||
|
|
@@ -695,6 +709,7 @@ def _impl(inputs, attr, params): | |
| 'Fill' : _fill(), | ||
| 'GatherV2' : _gather_v2(), | ||
| 'StridedSlice' : _stridedSlice(), | ||
| 'Split' : _split(), | ||
| 'LRN' : _lrn(), | ||
| 'Pad' : _pad('Pad'), | ||
| 'PadV2' : _pad('PadV2'), | ||
|
|
@@ -798,8 +813,8 @@ def _LSTMBlockCellWrapper(inputs, attr, params, | |
| """LSTM cell warapper to prepare the inputs""" | ||
| input_shape = attr['_input_shapes'][inputs[0]] | ||
| weight_shape = attr['_input_shapes'][inputs[3]] | ||
| batch_size = input_shape[0] | ||
| num_hidden = weight_shape[1] // 4 | ||
|
|
||
| if layer == 0: | ||
| #Create initial states placeholder in case of first layer | ||
|
|
@@ -982,24 +997,31 @@ def from_tensorflow(self, graph): | |
| # Pass the node name too in attr | ||
| attr["_node_name"] = node.name | ||
|
|
||
| inputs = [] | ||
| for node_input_name in node.input: | ||
| node_input_key = node_input_name.split(':') | ||
| slot_num = 0 | ||
| if len(node_input_key) > 1: | ||
| slot_num = int(node_input_key[1]) | ||
| node_input_key = node_input_key[0] | ||
|
|
||
| try: | ||
| try: | ||
| input_sym = self._nodes[node_input_key].__getitem__(slot_num) | ||
| except NNVMError: | ||
| # TODO: Fancy node name with invalid slot should discard and | ||
| # retrieve node input with zero'th(default) index. | ||
| # eg: Node name:- 'Model/RNN/cell_0/RnnCell:6', in this case | ||
| # name had fancy name convention and discard slot-id. | ||
| input_sym = self._nodes[node_input_key].__getitem__(0) | ||
|
|
||
| inputs.append(input_sym) | ||
| input_shapes[input_sym] = self._output_shapes[ | ||
| node_input_key].__getitem__(slot_num) | ||
| attr['_input_shapes'] = input_shapes | ||
| except KeyError: | ||
|
Contributor
There was a problem hiding this comment. This except is for the same statement below where the Suggest to make one try block with multiple except instead of nested.
Contributor
Author
There was a problem hiding this comment. as address to review comment, we need nested try block to handle fancy node name convention logic.
Contributor
There was a problem hiding this comment. It's the same statement which cause 2 exceptions in two different cases, one try on
Contributor
Author
There was a problem hiding this comment. First try block we are trying to access the slot base input and after caught NNVMError we are trying to access default(0) location input, in this case if we caught error should be thrown to user. |
||
| # TODO: Need to find clean way to handle '^CheckNumerics' | ||
| pass | ||
|
|
||
| inputs = self._fix_extranodes(node.op, attr, inputs) | ||
|
|
||
|
|
||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
|
|
@@ -71,17 +71,19 @@ def run_tvm_graph(graph_def, input_data, input_node, output_shape, output_dtype) | |
| def run_tf_graph(sess, input_data, input_node, output_node): | ||
| """ Generic function to execute tensorflow """ | ||
|
|
||
| if isinstance(input_data, list): | ||
| input_dict = {} | ||
| for i, e in enumerate(input_node): | ||
| input_dict[e] = input_data[i] | ||
| else: | ||
| input_dict = {input_node: input_data} | ||
|
|
||
| if isinstance(output_node, list): | ||
| tensor = [sess.graph.get_tensor_by_name(node_name) for node_name in output_node] | ||
| return sess.run(tensor, input_dict) | ||
|
|
||
| tensor = sess.graph.get_tensor_by_name(output_node) | ||
| return sess.run(tensor, input_dict) | ||
|
|
||
|
|
||
| def compare_tf_with_tvm(in_data, in_name, out_name, init_global_variables=False): | ||
|
|
@@ -787,29 +789,72 @@ def _get_sample(data, state): | |
| np.testing.assert_allclose(tf_samples, tvm_samples, rtol=1e-5, atol=1e-5) | ||
| assert(tvm_sample_str == tf_sample_str) | ||
|
|
||
|
|
||
| ####################################################################### | ||
| # Split | ||
| # ------ | ||
| def _test_split(ip_shape, num_split, axis): | ||
| tf.reset_default_graph() | ||
| dtype = 'float32' | ||
| in_data = tf.placeholder(dtype, ip_shape, name="in_data") | ||
| tf.split(in_data, num_split, axis=axis, name="split") | ||
| np_data = np.random.uniform(size=ip_shape).astype(dtype) | ||
| with tf.Session() as sess: | ||
| final_graph_def = tf.graph_util.convert_variables_to_constants( | ||
| sess, | ||
| sess.graph.as_graph_def(add_shapes=True), | ||
| ['split']) | ||
|
|
||
| tf_out_node_names = ['split:%g' % i for i in range(num_split)] | ||
| tf_output = run_tf_graph(sess, [np_data], ['in_data:0'], tf_out_node_names) | ||
| tvm_out_shapes = [tf_output[i].shape for i in range(num_split)] | ||
| tvm_out_dtypes = [tf_output[i].dtype for i in range(num_split)] | ||
| tvm_output = run_tvm_graph(final_graph_def, [np_data], ['in_data'], | ||
| tvm_out_shapes, tvm_out_dtypes) | ||
| np.testing.assert_allclose(tf_output, tvm_output, atol=1e-5, rtol=1e-5) | ||
| sess.close() | ||
|
|
||
| def test_forward_split(): | ||
| '''test split operator''' | ||
| _test_split((2, 3), 2, axis=0) | ||
| _test_split((6, 3), 3, axis=0) | ||
| _test_split((5, 9, 3), 3, axis=1) | ||
| _test_split((2,5,3,9), 3, axis=3) | ||
|
|
||
|
|
||
| ####################################################################### | ||
| # LRN (Local Response Normalization) | ||
| # ---------------------------------- | ||
| def _test_lrn(ip_shape, depth_radius=2, alpha=1e-05, beta=0.75, bias=1.0): | ||
|
Contributor
There was a problem hiding this comment. Ref. we don't need to explicitly set the default values. TF can handle them if not set. Suggest to ignore to allow TF default values play role here. |
||
| tf.reset_default_graph() | ||
| dtype = 'float32' | ||
| in_data = tf.placeholder(dtype, ip_shape, name="in_data") | ||
| tf.nn.local_response_normalization(in_data, | ||
| depth_radius=depth_radius, | ||
| alpha=alpha, | ||
| beta=beta, | ||
| bias=bias, | ||
| name="local_response_normalization") | ||
| np_data = np.random.uniform(size=ip_shape).astype(dtype) | ||
| with tf.Session() as sess: | ||
|
Contributor
There was a problem hiding this comment. #1546 adds Suggest to use it. |
||
| final_graph_def = tf.graph_util.convert_variables_to_constants( | ||
| sess, | ||
| sess.graph.as_graph_def(add_shapes=True), | ||
| ['local_response_normalization']) | ||
| tf_output = run_tf_graph(sess, [np_data], ['in_data:0'], | ||
| 'local_response_normalization:0') | ||
| tvm_output = run_tvm_graph(final_graph_def, [np_data], ['in_data'], | ||
| tf_output.shape, dtype) | ||
| np.testing.assert_allclose(tf_output, tvm_output, atol=1e-5, rtol=1e-5) | ||
| sess.close() | ||
|
|
||
| def test_forward_lrn(): | ||
| '''test local_response_normalization operator''' | ||
| _test_lrn((4, 4, 4, 4)) | ||
| _test_lrn((4,3,5,6), 1, 1e-05, 0.5) | ||
| _test_lrn((1, 3, 20, 20), 2, 1e-05, 0.5, 1.0) | ||
| _test_lrn((1, 3, 20, 20), 2, 1e-05, 0.75, 2.0) | ||
|
|
||
|
|
||
| ####################################################################### | ||
| # l2_normalize | ||
|
|
@@ -855,6 +900,7 @@ def test_forward_l2_normalize(): | |
| test_forward_resize_bilinear() | ||
| test_forward_pad() | ||
| test_forward_lstm() | ||
| test_forward_split() | ||
| test_forward_stridedslice() | ||
| test_forward_gather() | ||
| test_forward_ptb() | ||
|
|
||
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ref. Split
tensorflow has two variations for split as
num_splitandsize_split.Please handle if possible or raise exception accordingly.
If handling size_split please handle
numattribute if possible or raise exception accordingly.There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Split had num_split attribute and SplitV had size_splits attribute. As of this PR supported Split operator