Merge pull request #129 from cinemascience/MLBatchUpdates

adding batch ML functionality

pycinema/filters/MLTFPredictor.py

@@ -23,16 +23,17 @@ def __init__(self):
 
     def _update(self):
         images = self.inputs.images.get()
-        model = self.inputs.trainedModel.get()
-        # get required input properties from model 
+        modelList = self.inputs.trainedModel.get()
+        # get required input properties from first model
+        model = modelList[0]
         width = model.layers[0].input_shape[1]
         height = model.layers[0].input_shape[2]
         channels = model.layers[0].input_shape[3]
         if channels == 1:
           gray_req = True
         else: #channels == 3 or 4
           gray_req = False
-        
+
         result = []
         # iterate over all the images in the input images
         for image in self.inputs.images.get():
@@ -45,13 +46,17 @@ def _update(self):
           data = np.array(data) / 255
           data = data.reshape((1, width, height, channels))
 
-          # at the moment, assuming only one predicted output
-          # from the network
-          one_hot = model.predict(data, verbose = 0)
-          predictedValue = np.argmax(one_hot)
-          img.meta['PredictedValue'] = float(predictedValue)
+          pv_iter = 0
+          for model in modelList:
+            # at the moment, assuming only one predicted output
+            # from the network
+            one_hot = model.predict(data, verbose = 0)
+            predictedValue = np.argmax(one_hot)
+            img.meta['PredictedValue_' + str(pv_iter)] = float(predictedValue)
+            pv_iter += 1
+
           result.append(img)
 
         self.outputs.images.set(result)
 
-        return 1;
+        return 1;

pycinema/filters/MLTFReader.py

@@ -2,7 +2,9 @@
 import logging as log
 
 from os.path import exists
+import os
 import tensorflow as tf
+import csv
 
 class MLTFReader(Filter):
 
@@ -12,35 +14,60 @@ def __init__(self):
             'path': ''
           },
           outputs={
-            'model': None
+            'model': []
           }
         )
 
     def _update(self):
 
         modelPath = self.inputs.path.get()
+        #modelDirectoryPath = modelDirectoryPath + "/" #in case missing
 
         if not modelPath: #if the path is empty
-            log.error('ML Model path empty')
-            self.outputs.model.set(None)
+            log.error('ML Model or data.csv path empty')
+            self.outputs.models.set([])
             return 0
 
         if not exists(modelPath):
-            log.error('[ERROR] ML Model not found:', modelPath)
-            self.outputs.model.set(None)
+            log.error('[ERROR] ML Model or data.csv not found:', modelPath)
+            self.outputs.models.set([])
             return 0
 
+        # load_model supports two formats that a tensorflow 
+        # model can be saved as: .keras and HDF5
         try:
-            # load_model supports the three formats that a tensorflow 
-            # model can be saved as: .keras, HDF5 and SavedModel 
-            model = tf.keras.models.load_model(modelPath)
+            models = []
+            # if the path directly points to a TF model
+            if modelPath.endswith(".h5") or modelPath.endswith(".keras"):
+                model = tf.keras.models.load_model(modelPath)
+                models.append(model)
+
+            else:
+                # if the path points to a data.csv file 
+                # the csv file must have the first column as numerical order of files
+                # the second column is the path to each model
+                # parse through models directory and load a list of models
+                table = []
+                with open(modelPath, 'r+') as csvfile:
+                    rows = csv.reader(csvfile, delimiter=',')
+                    for row in rows:  
+                        table.append(row) 
+                table = table[1:]
+                numModels = len(table)
+
+                models = [None] * numModels
+                for row in table:
+                    parent = os.path.dirname(modelPath) + "/"
+                    filePath = os.path.join(parent, row[1])
+                    model = tf.keras.models.load_model(filePath)      
+                    models[int(row[0])] = model
 
             #check if training configuration exists, if not give error
         except:
-            log.error('[ERROR] Unable to open ML Model')
-            self.outputs.model.set(None)
+            log.error('[ERROR] Unable to open ML Model Directory')
+            self.outputs.models.set([])
             return 0
 
-        self.outputs.model.set(model)
+        self.outputs.model.set(models)
 
-        return 1
+        return 1

testing/gold/MLOutputTest.csv

@@ -1,11 +1,11 @@
-TrueValue,FILE,id,PredictedValue
-0,data/mnist.cdb/image/0.png,0,0.0
-1,data/mnist.cdb/image/1.png,1,1.0
-2,data/mnist.cdb/image/2.png,2,2.0
-3,data/mnist.cdb/image/3.png,3,3.0
-4,data/mnist.cdb/image/4.png,4,4.0
-5,data/mnist.cdb/image/5.png,5,5.0
-6,data/mnist.cdb/image/6.png,6,6.0
-7,data/mnist.cdb/image/7.png,7,7.0
-8,data/mnist.cdb/image/8.png,8,7.0
-9,data/mnist.cdb/image/9.png,9,9.0
+id,TrueValue,FILE,PredictedValue_0
+0,0,data/mnist.cdb/image/0.png,0.0
+1,1,data/mnist.cdb/image/1.png,1.0
+2,2,data/mnist.cdb/image/2.png,2.0
+3,3,data/mnist.cdb/image/3.png,3.0
+4,4,data/mnist.cdb/image/4.png,4.0
+5,5,data/mnist.cdb/image/5.png,5.0
+6,6,data/mnist.cdb/image/6.png,6.0
+7,7,data/mnist.cdb/image/7.png,7.0
+8,8,data/mnist.cdb/image/8.png,7.0
+9,9,data/mnist.cdb/image/9.png,9.0