Rounds any large float constants to the specified number of levels.
| 30 | |
| 31 | // Rounds any large float constants to the specified number of levels. |
| 32 | Status RoundWeights(const GraphDef& input_graph_def, |
| 33 | const TransformFuncContext& context, |
| 34 | GraphDef* output_graph_def) { |
| 35 | int32 num_steps; |
| 36 | TF_RETURN_IF_ERROR( |
| 37 | context.GetOneInt32Parameter("num_steps", 256, &num_steps)); |
| 38 | TF_RETURN_IF_ERROR(ReplaceMatchingOpTypes( |
| 39 | input_graph_def, {"Const"}, |
| 40 | [num_steps](const NodeMatch& match, const std::set<string>& input_nodes, |
| 41 | const std::set<string>& output_nodes, |
| 42 | std::vector<NodeDef>* new_nodes) { |
| 43 | const NodeDef& old_const_node = match.node; |
| 44 | if (!old_const_node.attr().count("dtype")) { |
| 45 | return errors::InvalidArgument("No 'dtype' attribute for Const node ", |
| 46 | old_const_node.name()); |
| 47 | } |
| 48 | if (!old_const_node.attr().count("value")) { |
| 49 | return errors::InvalidArgument("No 'value' attribute for Const node ", |
| 50 | old_const_node.name()); |
| 51 | } |
| 52 | const DataType old_dtype = old_const_node.attr().at("dtype").type(); |
| 53 | Tensor old_tensor; |
| 54 | if (!old_tensor.FromProto(old_const_node.attr().at("value").tensor())) { |
| 55 | return errors::InvalidArgument("Decoding Tensor failed for node", |
| 56 | old_const_node.name()); |
| 57 | } |
| 58 | const size_t num_elements = old_tensor.NumElements(); |
| 59 | // If this isn't a float constant, or it's too small, then reuse the |
| 60 | // same node with no changes. The size is important because small |
| 61 | // constants tend to be used for more accuracy-sensitive calculations, |
| 62 | // and the benefit of shrinking them is very marginal. |
| 63 | if ((old_dtype != DT_FLOAT) || (num_elements < 16)) { |
| 64 | new_nodes->push_back(old_const_node); |
| 65 | return Status::OK(); |
| 66 | } |
| 67 | const float* old_values = old_tensor.flat<float>().data(); |
| 68 | float min = std::numeric_limits<float>::max(); |
| 69 | float max = std::numeric_limits<float>::min(); |
| 70 | for (int i = 0; i < num_elements; ++i) { |
| 71 | const float value = old_values[i]; |
| 72 | min = std::min(min, value); |
| 73 | max = std::max(max, value); |
| 74 | } |
| 75 | // min_value == max_value is a tricky case. It can occur for general |
| 76 | // tensors, and of course for scalars. The quantized ops cannot deal |
| 77 | // with this case, so we set max_value to something else. |
| 78 | // It's a tricky question what is the numerically best solution to |
| 79 | // deal with this degeneracy. |
| 80 | // TODO(petewarden): Better use a tolerance than a hard comparison? |
| 81 | if (min == max) { |
| 82 | if (std::abs(min) < 0.000001f) { |
| 83 | max = min + 1.0f; |
| 84 | } else if (min > 0) { |
| 85 | max = 2.0f * min; |
| 86 | } else { |
| 87 | min = 2.0f * max; |
| 88 | } |
| 89 | } |