Add set_data_types_from_keras_model() and set_accum_from_keras_model() for automatic precision inference

hls4ml/model/profiling.py

@@ -1,16 +1,30 @@
 from hls4ml.model.hls_model import HLSModel
 from hls4ml.model.hls_layers import IntegerPrecisionType, FixedPrecisionType
-import matplotlib.pyplot as plt
 import numpy as np
-import pandas
-import seaborn as sb
 import uuid
 import os
 import shutil
 from collections import defaultdict
-
 from hls4ml.model.hls_model import HLSModel
 
+try:
+ import matplotlib.pyplot as plt
+ __plt_present__ = True
+except ImportError:
+ __plt_present__ = False
+
+try:
+ import pandas
+ __pandas_present__ = True
+except ImportError:
+ __pandas_present__ = False
+
+try:
+ import seaborn as sb
+ __seaborn_present__ = True
+except ImportError:
+ __seaborn_present__ = False
+
 try:
  from tensorflow import keras
  import qkeras
@@ -25,6 +39,24 @@
  __torch_profiling_enabled__ = False
 
 
+def _check_plt():
+ if not __plt_present__:
+ raise RuntimeError("matplotlib could not be imported. Have you installed it by installing hls4ml[profiling] "
+ "or otherwise?")
+
+
+def _check_pandas():
+ if not __pandas_present__:
+ raise RuntimeError("pandas could not be imported. Have you installed it by installing hls4ml[profiling] "
+ "or otherwise?")
+
+
+def _check_seaborn():
+ if not __seaborn_present__:
+ raise RuntimeError("seaborn could not be imported. Have you installed it by installing hls4ml[profiling] "
+ "or otherwise?")
+
+
 def get_unoptimized_hlsmodel(model):
  from hls4ml.converters import convert_from_config
 
@@ -61,6 +93,9 @@ def array_to_summary(x, fmt='boxplot'):
  return y
 
 def boxplot(data, fmt='longform'):
+ _check_plt()
+ _check_seaborn()
+
  if fmt == 'longform':
  f = plt.figure() #figsize=(3, 3))
  hue = 'layer' if 'layer' in data.keys() else None
@@ -93,6 +128,9 @@ def boxplot(data, fmt='longform'):
  return None
 
 def histogram(data, fmt='longform'):
+ _check_plt()
+ _check_seaborn()
+
  f = plt.figure()
  from matplotlib.ticker import MaxNLocator
  n = len(data) if fmt == 'summary' else len(data['weight'].unique())
@@ -115,6 +153,8 @@ def histogram(data, fmt='longform'):
  'histogram' : histogram}
 
 def types_boxplot(data, fmt='longform'):
+ _check_plt()
+
  from matplotlib.patches import PathPatch
  from matplotlib.patches import Rectangle
  ax = plt.gca()
@@ -140,6 +180,9 @@ def types_boxplot(data, fmt='longform'):
  ax.add_patch(rectangle)
 
 def types_histogram(data, fmt='longform'):
+ _check_plt()
+ _check_seaborn()
+
  ax = plt.gca()
  layers = np.array(ax.get_legend_handles_labels()[1])
  colors = sb.color_palette("husl", len(layers))
@@ -160,6 +203,8 @@ def ap_fixed_WIF(dtype):
  return W, I, F
 
 def types_hlsmodel(model):
+ _check_pandas()
+
  suffix = ['w', 'b']
  data = {'layer' : [], 'low' : [], 'high' : []}
  # Plot the default precision
@@ -183,6 +228,8 @@ def types_hlsmodel(model):
  return data
 
 def activation_types_hlsmodel(model):
+ _check_pandas()
+
  data = {'layer' : [], 'low' : [], 'high' : []}
  # Get the default precision
  default_precision = model.config.model_precision['default']
@@ -225,10 +272,35 @@ def weights_hlsmodel(model, fmt='longform', plot='boxplot'):
  data[-1]['weight'] = l
 
  if fmt == 'longform':
+ _check_pandas()
  data = pandas.DataFrame(data)
  return data
 
 
+def activations_hlsmodel(model, X, fmt='summary', plot='boxplot'):
+ if fmt == 'longform':
+ raise NotImplemented
+ elif fmt == 'summary':
+ data = []
+
+ _, trace = model.trace(np.ascontiguousarray(X))
+
+ if len(trace) == 0:
+ raise RuntimeError("HLSModel must have tracing on for at least 1 layer (this can be set in its config)")
+
+ for layer in trace.keys():
+ print(" {}".format(layer))
+
+ if fmt == 'summary':
+ y = trace[layer].flatten()
+ y = abs(y[y != 0])
+
+ data.append(array_to_summary(y, fmt=plot))
+ data[-1]['weight'] = layer
+
+ return data
+
+
 def _keras_batchnorm(layer):
  weights = layer.get_weights()
  epsilon = layer.epsilon
@@ -255,30 +327,6 @@ def _keras_layer(layer):
  })
 
 
-def activations_hlsmodel(model, X, fmt='summary', plot='boxplot'):
- if fmt == 'longform':
- raise NotImplemented
- elif fmt == 'summary':
- data = []
-
- _, trace = model.trace(np.ascontiguousarray(X))
-
- if len(trace) == 0:
- raise RuntimeError("HLSModel must have tracing on for at least 1 layer (this can be set in its config)")
-
- for layer in trace.keys():
- print(" {}".format(layer))
-
- if fmt == 'summary':
- y = trace[layer].flatten()
- y = abs(y[y != 0])
-
- data.append(array_to_summary(y, fmt=plot))
- data[-1]['weight'] = layer
-
- return data
-
-
 def weights_keras(model, fmt='longform', plot='boxplot'):
  if fmt == 'longform':
  data = {'x' : [], 'layer' : [], 'weight' : []}
@@ -305,6 +353,7 @@ def weights_keras(model, fmt='longform', plot='boxplot'):
  data[-1]['weight'] = l
 
  if fmt == 'longform':
+ _check_pandas()
  data = pandas.DataFrame(data)
  return data
 
@@ -332,6 +381,7 @@ def activations_keras(model, X, fmt='longform', plot='boxplot'):
  data[-1]['weight'] = layer.name
 
  if fmt == 'longform':
+ _check_pandas()
  data = pandas.DataFrame(data)
  return data
 
@@ -364,6 +414,7 @@ def weights_torch(model, fmt='longform', plot='boxplot'):
  data[-1]['weight'] = l
 
  if fmt == 'longform':
+ _check_pandas()
  data = pandas.DataFrame(data)
  return data
 
@@ -392,6 +443,7 @@ def activations_torch(model, X, fmt='longform', plot='boxplot'):
  data[-1]['weight'] = lname
 
  if fmt == 'longform':
+ _check_pandas()
  data = pandas.DataFrame(data)
  return data
 
@@ -423,6 +475,7 @@ def numerical(model=None, hls_model=None, X=None, plot='boxplot'):
  respectively. (Optimizations are applied to an HLSModel by hls4ml,
  a post-optimization HLSModel is a final model)
  """
+ _check_plt()
  wp, wph, ap, aph = None, None, None, None
 
  hls_model_present = hls_model is not None and isinstance(hls_model, HLSModel)
@@ -579,6 +632,8 @@ def get_ymodel_keras(keras_model, X):
 
 def _norm_diff(ymodel, ysim):
  """Calculate the square root of the sum of the squares of the differences"""
+ _check_plt()
+
  diff = {}
 
  for key in list(ysim.keys()):
@@ -601,6 +656,7 @@ def _dist_diff(ymodel, ysim):
  meaning "very difference".
  If difference < original value then the normalized difference would be difference/original.
  """
+ _check_plt()
 
  diff = {}
 
@@ -662,7 +718,8 @@ def compare(keras_model, hls_model, X, plot_type = "dist_diff"):
  matplotlib figure
  plot object of the histogram depicting the difference in each layer's output
  """
-
+ _check_plt()
+
  #Take in output from both models
  #Note that each y is a dictionary with structure {"layer_name": flattened ouput array}
  ymodel = get_ymodel_keras(keras_model, X)

hls4ml/utils/__init__.py

@@ -1,5 +1,5 @@
 from __future__ import absolute_import
 
-from hls4ml.utils.config import config_from_keras_model
+from hls4ml.utils.config import config_from_keras_model, set_data_types_from_keras_model, set_accum_from_keras_model
 from hls4ml.utils.plot import plot_model
 from hls4ml.utils.example_models import fetch_example_model, fetch_example_list