}
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/biases/read"
  op: "Identity"
  input: "BoxPredictor_4/ClassPredictor/biases"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@BoxPredictor_4/ClassPredictor/biases"
      }
    }
  }
}
node {

    fc1 = gen_mnist_ops.new_fully_connected(reshape, self.weights['f3'],
                                            self.biases['b3'], 'RELU')
    # output shape: [-1, 10]
    return gen_mnist_ops.new_fully_connected(fc1, self.weights['f4'],
                                             self.biases['b4'])


def main(strategy):
  """Trains an MNIST model using the given tf.distribute.Strategy."""

 private:
  // Whether the value is used as a bias input of another op. Here we assume
  // bias is used immediately by the user. This assumption is always correct
  // after constant folding.
  bool UsedAsBias(Value value) {
    for (auto &use : value.getUses()) {
      auto biases = TFL::GetOpQuantSpec(use.getOwner())->biases_params;
      if (biases.find(use.getOperandNumber()) != biases.end()) return true;
    }
    return false;
  }

      }
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/biases/read"
  op: "Identity"
  input: "BoxPredictor_4/ClassPredictor/biases"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@BoxPredictor_4/ClassPredictor/biases"
      }
    }
  }
}
node {

      }
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/biases/read"
  op: "Identity"
  input: "BoxPredictor_4/ClassPredictor/biases"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@BoxPredictor_4/ClassPredictor/biases"
      }
    }
  }
}
node {

      }
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/biases/read"
  op: "Identity"
  input: "BoxPredictor_4/ClassPredictor/biases"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@BoxPredictor_4/ClassPredictor/biases"
      }
    }
  }
}
node {

          GetQuantScaleSpec(user);
      const BiasParamsMap biases = spec->biases_params;

      // The quantization parameters of a `weight` shouldn't be determined by
      // other values. So any constants which are not bias, an operand of an
      // op with same scale requirements, and haven't been quantized are
      // weights.
      if (!biases.contains(operand_num) &&
          !scale_spec->has_same_scale_requirement &&

        return;

      // We only fold int32/int64 for Const and i32 for QConst if not specify
      // all constants flag. (Since they're more like "configs" or i32 biases.)
      // We will fold every const ops (and q_const ops) if we speicfy the
      // fold_all_constants_flag.
      if (!fold_all_constants_flag_) {
        if (!IsConstOrQConstInt(op)) return;
      }

	// randomly. We may want to either pick a much larger value
	// (AFL uses 32768, paired with a similar impl to chooseLen
	// which biases towards smaller lengths that grow over time),
	// or set the max based on characteristics of the corpus
	// (libFuzzer sets a min/max based on the min/max size of
	// entries in the corpus and then picks uniformly from
	// that range).
	n := m.chooseLen(4096)
	if len(b)+n >= cap(b) {

			// instead of the original value (see https://go.dev/issue/51526), so
			// they must be treated as int32 instead.
			//
			// We arbitrarily draw the line at UTF-8 validity, which biases toward the
			// "rune" interpretation. (However, we accept either format as input.)
			if utf8.ValidRune(t) {
				fmt.Fprintf(b, "rune(%q)\n", t)
			} else {
				fmt.Fprintf(b, "int32(%v)\n", t)
			}