Trying to experiment with ncs : can't compile simple model (IndexError: list index out of range)

idata · ‎11-05-2018

Hi,

I am trying to get a first really simple model on the stick (just to check everything works). this is the model code :

import tensorflow as tf

tf.reset_default_graph()
X = tf.placeholder(dtype=tf.float32, shape=(1,2,2), name = "in037")
Y = tf.constant([[[1,1],[1,1]]], dtype=tf.float32)
Z = tf.math.add(X, Y, name="out037")
init_op = tf.global_variables_initializer()

with tf.Session() as sess:
  #sess.run(init_op)
  tf.train.export_meta_graph('test/graph.meta')

When I compile the graph : mvNCCompile graph.meta -s 1 -in=in037 -on=out037 -is 2 2 -o Test.graph,

this is what I get :

! > /usr/lib/python3.7/site-packages/networkx/classes/graph.py:23: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working

! > from collections import Mapping

! > /usr/lib/python3.7/site-packages/networkx/classes/reportviews.py:95: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working

! > from collections import Mapping, Set, Iterable

! > /usr/lib/python3.7/site-packages/skimage/init.py:60: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses

! > import imp

! > /usr/lib/python3.7/site-packages/google/protobuf/internal/containers.py:182: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working

! > MutableMapping = collections.MutableMapping

! > /usr/lib/python3.7/site-packages/google/protobuf/internal/containers.py:340: DeprecationWarning: Using or importing the ABCs from 'collections' instead of from 'collections.abc' is deprecated, and in 3.8 it will stop working

! > collections.MutableSequence.register(BaseContainer)

! > /usr/bin/ncsdk/Controllers/Parsers/TensorFlowParser/Convolution.py:46: SyntaxWarning: assertion is always true, perhaps remove parentheses?

! > assert(False, "Layer type not supported by Convolution: " + obj.type)

! > /usr/bin/ncsdk/Controllers/Parsers/Phases.py:617: DeprecationWarning: invalid escape sequence \

! > """

! > /usr/bin/ncsdk/Controllers/Parsers/Phases.py:893: DeprecationWarning: invalid escape sequence \

! > """

! > /usr/bin/ncsdk/Controllers/Parsers/Phases.py:322: SyntaxWarning: assertion is always true, perhaps remove parentheses?

! > assert(len(pred) == 1, "Slice not supported to have >1 predecessors")

! > mvNCCompile v02.00, Copyright @ Intel Corporation 2017

! >

! > __ WARNING: No Weights found. using filler weights __

! > /usr/bin/ncsdk/Controllers/TensorFlowParser.py:173: DeprecationWarning: invalid escape sequence \d

! > shape: (1, 2, 2)

! > res.shape: (1, 2, 2)

! > TensorFlow output shape: (1, 2, 2)

! > Traceback (most recent call last):

! > File "/usr/bin/mvNCCompile", line 206, in

! > create_graph(args.network, args.image, args.inputnode, args.outputnode, args.outfile, args.nshaves, args.inputsize, args.weights, args.explicit_concat, args.ma2480, args.scheduler, args.new_parser, args.cpp, args)

! > File "/usr/bin/mvNCCompile", line 185, in create_graph

! > load_ret = load_network(args, parser, myriad_config)

! > File "/usr/bin/ncsdk/Controllers/Scheduler.py", line 156, in load_network

! > network.optimize()

! > File "/usr/bin/ncsdk/Models/Network.py", line 250, in optimize

! > self.convert_network_input_to_yxz()

! > File "/usr/bin/ncsdk/Models/Network.py", line 337, in convert_network_input_to_yxz

! > if self.stageslist[0].op in [StageType.fully_connected_layer, StageType.convolution, StageType.max_pooling,

! > IndexError: list index out of range

The deprecation warnings doesn't really bothers me. The "no weights found" neither since it's just an "add" graph. But what is strange is the IndexError, since (as I read in the sources), self.stageslist should always contain at least one element (the input), and it seems detected since I read "shape: (1, 2, 2)" in the output… any ideas ?

idata · ‎11-05-2018

I just remembered there was something about the input dimension, so I tried this (I added the --new-parser option since I get the same error without), and I get :

import tensorflow as tf

tf.reset_default_graph()
X = tf.placeholder(dtype=tf.float32, shape=(1,2,2,1), name = "in037")
Y = tf.constant([[[[1],[1]],[[1],[1]]]], dtype=tf.float32)
Z = tf.math.add(X, Y, name="out037")
init_op = tf.global_variables_initializer()

with tf.Session() as sess:
  sess.run(init_op)
  tf.train.export_meta_graph('test/graph.meta')

mvNCCompile graph.meta -s 12 -in=in037 -on=out037 -o Test.graph --new-parser

__ WARNING: No Weights found. using filler weights __

graph.meta

output tensor shape (1, 2, 2, 1)

Fusing DeptwiseConv + Pointwise Convolution into plain Convolution

Fusing Add and Batch after Convolution

Fusing Pad and Convolution2D

Fusing BatchNorm and Scale after Convolution

Replacing BN with Bias&Scale

Fusing Permute and Flatten

Fusing Eltwise and Relu

Eliminate layers that have been parsed as NoOp

Evaluating input and weigths for each hw layer

Traceback (most recent call last):

File "/usr/bin/mvNCCompile", line 206, in

create_graph(args.network, args.image, args.inputnode, args.outputnode, args.outfile, args.nshaves, args.inputsize, args.weights, args.explicit_concat, args.ma2480, args.scheduler, args.new_parser, args.cpp, args)

File "/usr/bin/mvNCCompile", line 185, in create_graph

load_ret = load_network(args, parser, myriad_config)

File "/usr/bin/ncsdk/Controllers/Scheduler.py", line 118, in load_network

parsedLayers = fixTensors(parsedLayers, scheduler, arguments.ma2480)

File "/usr/bin/ncsdk/Controllers/Optimizer.py", line 160, in fixTensors

return fixTensorImported(parsedLayers, scheduler, myriadX)

File "/usr/bin/ncsdk/Controllers/ConversionPlacement.py", line 230, in fixTensors

assert(len(producerNode) == 1)

AssertionError