[Quantization] Make calibration faster and more memory usage friendly (…

…apache#4589)

* Use memory efficient calibrate

* Fixed indexing

* add cpp kl stub

* ported KL cpp from mxnet

* Fixed std::distance arguments order

* remove python implementation

* fix lint and indent

* fix indent

* refactoring

* fix lint

* fix for i386

python/tvm/relay/quantize/_calibrate.py

@@ -33,24 +33,7 @@
 from .kl_divergence import _find_scale_by_kl
 
 
-def collect_stats(mod, dataset):
-    """Given an annotated graph, create a profile graph to collect profile data from the
-    calibration dataset. This pass collects simulated_quantize op input into a tuple.
-    Simulated_quantize ops are rewritten to identity mode. The tuple is the output of the profile
-    graph.
-
-    Parameters
-    ----------
-    mod: Module
-        The simulation graph after annotation.
-
-    Returns
-    -------
-    ret: list of ndarray
-        List of output data of each layer
-    """
-
-    logging.info("collecting statistics for calibration...")
+def _get_profile_runtime(mod):
     func = mod['main']
     func = _quantize.CreateStatsCollector(func)
 
@@ -63,30 +46,61 @@ def collect_stats(mod, dataset):
 
     with _transform.build_config(opt_level=3):
         graph, lib, params = _build_module.build(func, target=target)
-    outputs = []
     runtime = graph_runtime.create(graph, lib, ctx)
     runtime.set_input(**params)
 
+    return runtime
+
+
+def collect_stats(mod, dataset, chunk_by=-1):
+    """Given an annotated graph, create a profile graph to collect profile data from the
+    calibration dataset. This pass collects simulated_quantize op input into a tuple.
+    Simulated_quantize ops are rewritten to identity mode. The tuple is the output of the profile
+    graph.
+
+    Parameters
+    ----------
+    mod: Module
+        The simulation graph after annotation.
+
+    dataset: Iterable[NDArray]
+        The calibration dataset.
+
+    chunk_by: optional, int
+        The size of chunk to be returned in one iteration. It is meant to be
+        used for reducing memory usage. If not specified, return samples for
+        all layers in one chunk.
+
+    Returns
+    -------
+    ret: Iterable[list of ndarray]
+        List of output data of each layer, chunked by the chunk_by parameter
+    """
+    logging.info("collecting statistics for calibration...")
+    runtime = _get_profile_runtime(mod)
     num_outputs = runtime.get_num_outputs()
-    outputs = [[] for i in range(num_outputs)]
+    chunk_by = num_outputs if chunk_by == -1 else chunk_by
 
-    for batch in dataset:
-        runtime.set_input(**batch)
-        runtime.run()
-        for i in range(num_outputs):
-            output = runtime.get_output(i).asnumpy()
-            outputs[i].append(output)
-    for i in range(num_outputs):
-        outputs[i] = np.concatenate(outputs[i]).reshape(-1)
-    return outputs
+    for i in range(0, num_outputs, chunk_by):
+        outputs = [[] for i in range(min(chunk_by, num_outputs - i))]
+        for batch in dataset:
+            runtime.set_input(**batch)
+            runtime.run()
+            for j in range(i, min(i+chunk_by, num_outputs)):
+                outputs[j-i].append(runtime.get_output(j).asnumpy())
+        yield [np.concatenate(output).reshape(-1) for output in outputs]
 
 
-def _kl_scale(stats):
-    with mp.Pool() as pool:
+def _kl_scale(mod, dataset):
+    cfg = quantize.current_qconfig()
+    chunk_by = cfg.calibrate_chunk_by
+    scales = []
+    for samples in collect_stats(mod, dataset, chunk_by):
         logging.info("finding threshold with kl for calibration...")
-        scales = list(pool.map(_find_scale_by_kl, stats))
+        with mp.Pool() as pool:
+            scales += list(pool.map(_find_scale_by_kl, samples))
 
-    def func(sq_call):  # pylint: disable=unused-argument
+    def func(_):
         scale = scales[func.scale_idx]
         func.scale_idx += 1
         return scale
@@ -168,13 +182,12 @@ def calibrate(dataset=None):
     ret: Function
         The module pass function.
     """
-    def wrapped_func(mod, ctx): # pylint: disable=unused-argument
+    def wrapped_func(mod, _):
         """make transform.module pass happy"""
         cfg = quantize.current_qconfig()
 
         if cfg.calibrate_mode == 'kl_divergence':
-            stats = collect_stats(mod, dataset)
-            input_scale_func = _kl_scale(stats)
+            input_scale_func = _kl_scale(mod, dataset)
         elif cfg.calibrate_mode == 'global_scale':
             input_scale_func = _global_scale
         else:

python/tvm/relay/quantize/kl_divergence.py

@@ -16,36 +16,14 @@
 # under the License.
 """Find optimal scale for quantization by minimizing KL-divergence"""
 
-try:
-    from scipy import stats
-except ImportError:
-    stats = None
-
+import ctypes
 import numpy as np
 
-
-def _smooth_distribution(p, eps=0.0001):
-    """Given a discrete distribution (may have not been normalized to 1),
-    smooth it by replacing zeros with eps multiplied by a scaling factor and taking the
-    corresponding amount off the non-zero values.
-    Ref: http://hanj.cs.illinois.edu/cs412/bk3/KL-divergence.pdf
-    """
-    is_zeros = (p == 0).astype(np.float32)
-    is_nonzeros = (p != 0).astype(np.float32)
-    n_zeros = is_zeros.sum()
-    n_nonzeros = p.size - n_zeros
-    if not n_nonzeros:
-        raise ValueError('The discrete probability distribution is malformed. All entries are 0.')
-    eps1 = eps * float(n_zeros) / float(n_nonzeros)
-    assert eps1 < 1.0, 'n_zeros=%d, n_nonzeros=%d, eps1=%f' % (n_zeros, n_nonzeros, eps1)
-    hist = p.astype(np.float32)
-    hist += eps * is_zeros + (-eps1) * is_nonzeros
-    assert (hist <= 0).sum() == 0
-    return hist
+from . import _quantize
 
 
-# pylint: disable=invalid-name
-def _find_scale_by_kl(arr, quantized_dtype='int8', num_bins=8001, num_quantized_bins=255):
+def _find_scale_by_kl(arr, quantized_dtype='int8',
+                      num_bins=8001, num_quantized_bins=255):
     """Given a tensor, find the optimal threshold for quantizing it.
     The reference distribution is `q`, and the candidate distribution is `p`.
     `q` is a truncated version of the original distribution.
@@ -54,73 +32,21 @@ def _find_scale_by_kl(arr, quantized_dtype='int8', num_bins=8001, num_quantized_
     http://on-demand.gputechconf.com/gtc/2017/presentation/s7310-8-bit-inference-with-tensorrt.pdf
     """
     assert isinstance(arr, np.ndarray)
-    assert stats is not None, "scipy needs to be installed for \
-    utilizing kl calibration during quantization"
-
     min_val = np.min(arr)
     max_val = np.max(arr)
-    th = max(abs(min_val), abs(max_val))
+    thres = max(abs(min_val), abs(max_val))
 
     if min_val >= 0 and quantized_dtype in ['uint8']:
         # We need to move negative bins to positive bins to fit uint8 range.
         num_quantized_bins = num_quantized_bins * 2 + 1
 
-    hist, hist_edges = np.histogram(arr, bins=num_bins, range=(-th, th))
-    zero_bin_idx = num_bins // 2
-    num_half_quantized_bins = num_quantized_bins // 2
-
-    thresholds = np.zeros(num_bins // 2 + 1 - num_quantized_bins // 2)
-    divergence = np.zeros_like(thresholds)
-    quantized_bins = np.zeros(num_quantized_bins, dtype=np.int32)
-    # i means the number of bins on half axis excluding the zero bin.
-    for i in range(num_quantized_bins // 2,
-                   num_bins // 2 + 1):
-        p_bin_idx_start = zero_bin_idx - i
-        p_bin_idx_stop = zero_bin_idx + i + 1
-        thresholds[i - num_half_quantized_bins] = hist_edges[p_bin_idx_stop]
-        sliced_nd_hist = hist[p_bin_idx_start:p_bin_idx_stop]
-
-        # generate reference distribution p
-        p = sliced_nd_hist.copy()
-        assert p.size % 2 == 1
-        assert p.size >= num_quantized_bins
-        # put left outlier count in p[0]
-        left_outlier_count = np.sum(hist[0:p_bin_idx_start])
-        p[0] += left_outlier_count
-        # put right outlier count in p[-1]
-        right_outlier_count = np.sum(hist[p_bin_idx_stop:])
-        p[-1] += right_outlier_count
-        # is_nonzeros[k] indicates whether hist[k] is nonzero
-        is_nonzeros = (p != 0).astype(np.int32)
+    def get_pointer(arr, ctypes_type):
+        ptr = arr.ctypes.data_as(ctypes.POINTER(ctypes_type))
+        return ctypes.cast(ptr, ctypes.c_void_p)
 
-        # calculate how many bins should be merged to generate quantized distribution q
-        num_merged_bins = sliced_nd_hist.size // num_quantized_bins
-        # merge hist into num_quantized_bins bins
-        for j in range(num_quantized_bins):
-            start = j * num_merged_bins
-            stop = start + num_merged_bins
-            quantized_bins[j] = sliced_nd_hist[start:stop].sum()
-        quantized_bins[-1] += sliced_nd_hist[num_quantized_bins * num_merged_bins:].sum()
-        # expand quantized_bins into p.size bins
-        q = np.zeros(sliced_nd_hist.size, dtype=np.float32)
-        for j in range(num_quantized_bins):
-            start = j * num_merged_bins
-            if j == num_quantized_bins - 1:
-                stop = len(is_nonzeros)
-            else:
-                stop = start + num_merged_bins
-            norm = is_nonzeros[start:stop].sum()
-            if norm != 0:
-                q[start:stop] = float(quantized_bins[j]) / float(norm)
-        q[p == 0] = 0
-        p = _smooth_distribution(p)
-        # There is a chance that q is an invalid probability distribution.
-        try:
-            q = _smooth_distribution(q)
-        except ValueError:
-            divergence[i - num_half_quantized_bins] = float("inf")
-        divergence[i - num_half_quantized_bins] = stats.entropy(p, q)
+    hist, hist_edges = np.histogram(arr, bins=num_bins, range=(-thres, thres))
+    hist_ptr = get_pointer(hist.astype(np.int32), ctypes.c_int)
+    hist_edges_ptr = get_pointer(hist_edges, ctypes.c_float)
 
-    min_divergence_idx = np.argmin(divergence)
-    opt_th = thresholds[min_divergence_idx]
-    return opt_th
+    return _quantize.FindScaleByKLMinimization(hist_ptr, hist_edges_ptr,
+                                               num_bins, num_quantized_bins)

python/tvm/relay/quantize/quantize.py

@@ -81,7 +81,8 @@ class QConfig(NodeBase):
         "do_simulation": False,
         "round_for_shift": True,
         "debug_enabled_ops": None,
-        "rounding": "UPWARD"
+        "rounding": "UPWARD",
+        "calibrate_chunk_by": -1,
     }
 
     # pylint: disable=no-member

src/relay/pass/quantize/calibrate.cc

@@ -26,12 +26,122 @@
 #include <tvm/relay/analysis.h>
 #include <tvm/relay/expr_functor.h>
 #include <tvm/relay/op.h>
+#include <numeric>
 #include "./quantize.h"
 
 namespace tvm {
 namespace relay {
 namespace quantize {
 
+// KL divergence minimization code is adapted from MXNet.
+// The original one is in incubator-mxnet/src/operator/quantization/calibrate.cc
+static std::vector<float> SmoothDistribution(const std::vector<float>& p,
+                                             const float eps = 0.0001) {
+  std::vector<size_t> is_zeros(p.size());
+  std::vector<size_t> is_nonzeros(p.size());
+  {
+    auto it = p.begin();
+    std::generate(is_zeros.begin(), is_zeros.end(),
+                  [&it]() { return static_cast<size_t>(*(it++) == 0.f); });
+  }
+  {
+    auto it = p.begin();
+    std::generate(is_nonzeros.begin(), is_nonzeros.end(),
+                  [&it]() { return static_cast<size_t>(*(it++) != 0.f); });
+  }
+  size_t n_zeros = std::accumulate(is_zeros.begin(), is_zeros.end(), 0);
+  size_t n_nonzeros = p.size() - n_zeros;
+  if (!n_nonzeros) {
+    // The discrete probability distribution is malformed. All entries are 0.
+    return std::vector<float>();
+  }
+  float eps1 = eps * static_cast<float>(n_zeros) / static_cast<float>(n_nonzeros);
+  if (eps1 >= 1.0) return std::vector<float>();
+  auto ret = p;
+  for (size_t i = 0; i < p.size(); i++) {
+    ret[i] += eps * is_zeros[i] - eps1 * is_nonzeros[i];
+  }
+  return ret;
+}
+
+static float ComputeEntropy(float* p, float* q, size_t size) {
+  float p_sum = std::accumulate(p, p+size, 0.f);
+  float q_sum = std::accumulate(q, q+size, 0.f);
+  float ret = 0;
+  for (size_t i = 0; i < size; i++) {
+    CHECK(p[i] > 0 && q[i] > 0);
+    p[i] /= p_sum;
+    q[i] /= q_sum;
+    if (p[i] && q[i]) ret += p[i] * std::log(p[i] / q[i]);
+  }
+  return ret;
+}
+
+float MinimizeKL(const std::vector<int>& hist,
+                 const std::vector<float>& hist_edges,
+                 int num_bins, int num_quantized_bins) {
+  const int zero_bin_idx = num_bins / 2;
+  const int num_half_quantized_bins = num_quantized_bins / 2;
+  std::vector<float> thresholds(num_bins / 2 + 1 - num_quantized_bins / 2, 0.f);
+  std::vector<float> divergence(thresholds.size(), 0.f);
+  std::vector<float> quantized_bins(num_quantized_bins, 0);
+  for (int i = num_quantized_bins / 2; i < zero_bin_idx + 1; ++i) {
+    const int p_bin_idx_start = zero_bin_idx - i;
+    const int p_bin_idx_stop = zero_bin_idx + i + 1;
+    thresholds[i - num_half_quantized_bins] = hist_edges[p_bin_idx_stop];
+
+    std::vector<int> sliced_nd_hist(p_bin_idx_stop - p_bin_idx_start);
+    std::vector<float> p(sliced_nd_hist.size());
+    p[0] = 0;
+    p.back() = 0;
+    for (int j = 0; j < num_bins; j++) {
+      if (j <= p_bin_idx_start) {
+        p[0] += hist[j];
+      } else if (j >= p_bin_idx_stop) {
+        p.back() += hist[j];
+      } else {
+        sliced_nd_hist[j - p_bin_idx_start] = hist[j];
+        p[j - p_bin_idx_start] = hist[j];
+      }
+    }
+    // calculate how many bins should be merged to generate quantized distribution q
+    const auto num_merged_bins = sliced_nd_hist.size() / num_quantized_bins;
+    for (int j = 0; j < num_quantized_bins; j++) {
+      const int start = j * num_merged_bins;
+      const int stop = (j + 1) * num_merged_bins;
+      quantized_bins[j] =
+          std::accumulate(sliced_nd_hist.begin() + start, sliced_nd_hist.begin() + stop, 0);
+    }
+    quantized_bins.back() += std::accumulate(
+        sliced_nd_hist.begin() + static_cast<int>(num_quantized_bins * num_merged_bins),
+        sliced_nd_hist.end(), 0);
+    // expand quantized_bins into p.size bins
+    std::vector<float> q(sliced_nd_hist.size(), 0);
+    for (int j = 0; j < num_quantized_bins; j++) {
+      const int start = j * num_merged_bins;
+      const int stop = (j == num_quantized_bins - 1) ? q.size() : ((j + 1) * num_merged_bins);
+      int norm = std::count_if(sliced_nd_hist.begin() + start, sliced_nd_hist.begin() + stop,
+                               [](size_t i) { return i != 0; });
+      if (norm) {
+        for (int k = start; k < stop; k++) {
+          if (p[k]) q[k] = quantized_bins[j] / norm;
+        }
+      }
+    }
+    p = SmoothDistribution(p);
+    q = SmoothDistribution(q);
+
+    if (!q.size()) {
+      divergence[i - num_half_quantized_bins] = std::numeric_limits<float>::infinity();
+    } else {
+      divergence[i - num_half_quantized_bins] = ComputeEntropy(p.data(), q.data(), p.size());
+    }
+  }
+  auto min_divergence_idx = std::distance(divergence.begin(),
+                                          std::min_element(divergence.begin(), divergence.end()));
+  return thresholds[min_divergence_idx];;
+}
+
 class StatsCollector : private ExprMutator {
  public:
   StatsCollector() : simulated_quantize_op_(Op::Get("relay.op.annotation.simulated_quantize")) {}
@@ -95,6 +205,18 @@ Expr CreateStatsCollector(const Expr& expr) {
 TVM_REGISTER_API("relay._quantize.CreateStatsCollector")
 .set_body_typed(CreateStatsCollector);
 
+
+TVM_REGISTER_API("relay._quantize.FindScaleByKLMinimization")
+.set_body([](TVMArgs args, TVMRetValue *ret) {
+  int* hist_ptr = static_cast<int*>(static_cast<void*>(args[0]));
+  float* hist_edges_ptr = static_cast<float*>(static_cast<void*>(args[1]));
+  int num_bins = args[2];
+  int num_quantized_bins = args[3];
+  std::vector<int> hist(hist_ptr, hist_ptr + num_bins);
+  std::vector<float> hist_edges(hist_edges_ptr, hist_edges_ptr + num_bins + 1);
+  ret[0] = MinimizeKL(hist, hist_edges, num_bins, num_quantized_bins);
+});
+
 }  // namespace quantize
 }  // namespace relay
 }  // namespace tvm