Merge pull request #454 from cpaxton/feature/scripts-conditional

Updates to scripts and to conditional image tests

.gitignore

@@ -26,4 +26,13 @@
 *.h5
 *.so
 *.ttt
-*.json
+*.json
+
+# Tags file
+tags
+
+# Slurm files
+slurm-*.out
+
+# orig files from vim
+*.orig

commands/multi_secondaries_no_ros.sh

@@ -0,0 +1,44 @@
+#!/usr/bin/env bash
+echo "------------------------------------"
+echo "Running $0"
+echo "This will train the following:"
+echo " - V(x)"
+echo " - p(action | x)"
+echo " - pi(x, action) ~ goal_pose"
+echo " - pi(x, action) ~ u"
+echo "------------------------------------"
+./costar_models/scripts/ctp_model_tool \
+  --model secondary --data_file data.h5f \
+  --epochs 50 \
+  --lr 0.001  --dropout_rate 0.1 \
+  --features multi --batch_size 64 \
+  --steps_per_epoch 500 --submodel value $1
+
+./costar_models/scripts/ctp_model_tool \
+  --model secondary --data_file data.h5f \
+  --epochs 50 \
+  --lr 0.001  --dropout_rate 0.1 --features multi \
+  --batch_size 64 --steps_per_epoch 500 --submodel next $1
+
+./costar_models/scripts/ctp_model_tool \
+  --model secondary --data_file data.h5f \
+  --epochs 50 \
+  --lr 0.001  --dropout_rate 0.1 --features multi \
+  --batch_size 64 --steps_per_epoch 500 --submodel q $1
+
+./costar_models/scripts/ctp_model_tool \
+  --model secondary --data_file data.h5f \
+  --epochs 100 \
+  --lr 0.001  --dropout_rate 0.1 \
+  --success_only \
+  --features multi --batch_size 64 \
+  --steps_per_epoch 500 --submodel actor $1
+
+ ./costar_models/scripts/ctp_model_tool \
+   --model secondary --data_file data.h5f \
+  --epochs 200 \
+   --lr 0.001  --dropout_rate 0.1 \
+   --features multi --batch_size 64 \
+   --success_only \
+   --steps_per_epoch 500 --submodel pose $1
+

costar_models/python/costar_models/abstract.py

@@ -19,9 +19,12 @@ class AbstractAgentBasedModel(object):
     '''
 
     def makeName(self, prefix, submodel=None):
-        name = os.path.join(self.model_directory, prefix) + "_model"
+        dir = self.model_directory
+        if submodel is not None and self.reqs_directory is not None:
+            dir = self.reqs_directory
+        name = os.path.join(dir, prefix) + "_model"
         if self.features is not None:
-            name += "_%s"%self.features   
+            name += "_%s"%self.features
         if submodel is not None:
             name += "_%s.h5f"%submodel
         return name
@@ -65,7 +68,7 @@ def __init__(self, taskdef=None,
             use_batchnorm=1,
             hypothesis_dropout=False,
             dense_representation=True,
-            skip_connections=0,
+            skip_connections=1,
             use_noise=False,
             load_pretrained_weights=False,
             retrain=True,
@@ -81,8 +84,12 @@ def __init__(self, taskdef=None,
             clip_weights=0, use_wasserstein=False,
             option_num=None, # for policy model
             unique_id="", # for status file
-            task=None, robot=None, model="", model_directory="./", *args,
-            **kwargs):
+            task=None,
+            robot=None,
+            model="",
+            model_directory="./",
+            reqs_directory=None,
+            *args, **kwargs):
 
         if lr == 0 or lr < 1e-30:
             raise RuntimeError('You probably did not mean to set ' + \
@@ -124,6 +131,9 @@ def __init__(self, taskdef=None,
         self.clipnorm = float(clipnorm)
         self.taskdef = taskdef
         self.model_directory = os.path.expanduser(model_directory)
+        self.reqs_directory = None
+        if reqs_directory is not None:
+            self.reqs_directory = os.path.expanduser(reqs_directory)
         self.name = self.makeName(self.name_prefix)
         self.num_generator_files = num_generator_files
         self.dropout_rate = dropout_rate
@@ -143,7 +153,7 @@ def __init__(self, taskdef=None,
         self.save_model = save_model
         self.hidden_size = hidden_size
         self.option_num = option_num
-        
+
         if self.noise_dim < 1:
             self.use_noise = False
 
@@ -160,7 +170,7 @@ def __init__(self, taskdef=None,
         # Unique id for status file
         self.unique_id = unique_id
 
-        
+
 
         # default: store the whole model here.
         # NOTE: this may not actually be where you want to save it.
@@ -175,6 +185,7 @@ def __init__(self, taskdef=None,
         print("Task =", self.task)
         print("Model type =", model)
         print("Model directory =", self.model_directory)
+        print("Reqs directory =", self.reqs_directory)
         print("Models saved with prefix =", self.name)
         print("Unique id for status file =", self.unique_id)
         print("-----------------------------------------------------------")
@@ -239,9 +250,9 @@ def trainFromGenerators(self, train_generator, test_generator, data=None):
     def _getData(self, *args, **kwargs):
         '''
         This function should process all the data you need for a generator.
-        ''' 
+        '''
         raise NotImplementedError('_getData() requires a dataset.')
-        
+
     def trainGenerator(self, dataset):
         return self._yieldLoop(dataset.sampleTrain)
 
@@ -254,7 +265,7 @@ def testGenerator(self, dataset):
 
     def _yieldLoop(self, sampleFn):
       '''
-      This helper function runs in a loop infinitely, executing some callable 
+      This helper function runs in a loop infinitely, executing some callable
       to extract a set of feature information from a dataset file, and then
       performs any necessary preprocessing on it.
 
@@ -430,7 +441,7 @@ def __init__(self, taskdef, *args, **kwargs):
         self.baseline = None
         # All low-level policies pi(x,o) --> u
         self.policies = []
-       
+
     def _makeSupervisor(self, feature):
         '''
         This needs to create a supervisor. This one maps from input to the

costar_models/python/costar_models/conditional_image.py

@@ -47,7 +47,6 @@ def __init__(self, *args, **kwargs):
         self.rep_size = 256
         self.num_transforms = 3
         self.transform_model = None
-        self.skip_connections = False
         self.save_encoder_decoder = self.retrain
 
         if self.use_noise:
@@ -73,22 +72,15 @@ def _makePredictor(self, features):
         label_in = Input((1,))
         ins = [img0_in, img_in]
 
-        if self.skip_connections:
-            encoder = self._makeImageEncoder2(img_shape)
-            decoder = self._makeImageDecoder2(self.hidden_shape)
-        else:
-            encoder = self._makeImageEncoder(img_shape)
-            decoder = self._makeImageDecoder(self.hidden_shape)
+        encoder = self._makeImageEncoder(img_shape)
+        decoder = self._makeImageDecoder(self.hidden_shape)
 
         LoadEncoderWeights(self, encoder, decoder)
 
         # =====================================================================
         # Load the arm and gripper representation
-        if self.skip_connections:
-            h, s32, s16, s8 = encoder([img0_in, img_in])
-        else:
-            h = encoder([img_in])
-            h0 = encoder(img0_in)
+        h = encoder([img_in])
+        h0 = encoder(img0_in)
 
         if self.validate:
             self.loadValidationModels(arm_size, gripper_size, h0, h)
@@ -132,7 +124,7 @@ def _makePredictor(self, features):
         if self.no_disc:
             disc_wt = 0.
         else:
-            disc_wt = 1e-4
+            disc_wt = 1e-3
         if self.no_disc:
             train_predictor = Model(ins + [label_in],
                     [image_out, image_out2] + enc_outs)
@@ -185,26 +177,31 @@ def loadValidationModels(self, arm_size, gripper_size, h0, h):
         arm_gripper = Concatenate()([arm_in, gripper_in])
         label_in = Input((1,))
 
-        self.value_model = GetValueModel(h, self.num_options, 64,
+        print(">>> VALUE MODEL")
+        self.value_model = GetValueModel(h, self.num_options, 128,
                 self.decoder_dropout_rate)
         self.value_model.compile(loss="mae", optimizer=self.getOptimizer())
         self.value_model.load_weights(self.makeName("secondary", "value"))
 
+        print(">>> NEXT MODEL")
         self.next_model = GetNextModel(h, self.num_options, 128,
                 self.decoder_dropout_rate)
         self.next_model.compile(loss="mae", optimizer=self.getOptimizer())
         self.next_model.load_weights(self.makeName("secondary", "next"))
 
+        print(">>> ACTOR MODEL")
         self.actor = GetActorModel(h, self.num_options, arm_size, gripper_size,
                 self.decoder_dropout_rate)
         self.actor.compile(loss="mae",optimizer=self.getOptimizer())
         self.actor.load_weights(self.makeName("secondary", "actor"))
 
+        print(">>> POSE MODEL")
         self.pose_model = GetPoseModel(h, self.num_options, arm_size, gripper_size,
                 self.decoder_dropout_rate)
         self.pose_model.compile(loss="mae",optimizer=self.getOptimizer())
         self.pose_model.load_weights(self.makeName("secondary", "pose"))
 
+        print(">>> Q MODEL")
         self.q_model = GetNextModel(h, self.num_options, 128,
                 self.decoder_dropout_rate)
         self.q_model.compile(loss="mae", optimizer=self.getOptimizer())
@@ -238,6 +235,10 @@ def pnext(self, hidden0, hidden, prev_option):
         #p /= np.sum(p)
         return p
 
+    def q(self, hidden0, hidden, prev_option):
+        p = self.q_model.predict([hidden0, hidden, prev_option])
+        return p
+
     def value(self, hidden0, hidden):
         #v = self.value_model.predict([h0, hidden, prev_option])
         v = self.value_model.predict([hidden0, hidden])

costar_models/python/costar_models/conditional_image_gan.py

@@ -46,7 +46,6 @@ def __init__(self, *args, **kwargs):
         self.rep_size = 256
         self.num_transforms = 3
         self.do_all = True
-        self.skip_connections = False
         self.save_encoder_decoder = self.retrain
         self.noise_iters = 2
 
@@ -71,12 +70,8 @@ def _makePredictor(self, features):
         next_option2_in = Input((1,), name="next_option2_in")
         ins = [img0_in, img_in, next_option_in, next_option2_in]
 
-        if self.skip_connections:
-            encoder = self._makeImageEncoder2(img_shape)
-            decoder = self._makeImageDecoder2(self.hidden_shape)
-        else:
-            encoder = self._makeImageEncoder(img_shape)
-            decoder = self._makeImageDecoder(self.hidden_shape)
+        encoder = self._makeImageEncoder(img_shape)
+        decoder = self._makeImageDecoder(self.hidden_shape)
 
         LoadEncoderWeights(self, encoder, decoder, gan=True)
 
@@ -87,11 +82,8 @@ def _makePredictor(self, features):
             z2 = Input((self.noise_dim,), name="z2_in")
             ins += [z1, z2]
 
-        if self.skip_connections:
-            h, s32, s16, s8 = encoder([img0_in, img_in])
-        else:
-            h = encoder([img_in])
-            h0 = encoder(img0_in)
+        h = encoder([img_in])
+        h0 = encoder(img0_in)
 
         # =====================================================================
         # Actually get the right outputs
@@ -132,7 +124,7 @@ def _makePredictor(self, features):
         # =====================================================================
         # And adversarial model
         loss = wasserstein_loss if self.use_wasserstein else "binary_crossentropy"
-        weights = [0.01, 0.01, 1.] if self.use_wasserstein else [100., 100., 1.]
+        weights = [0.1, 0.1, 1.] if self.use_wasserstein else [100., 100., 1.]
 
         model = Model(ins, [image_out, image_out2, is_fake])
         model.compile(
@@ -175,8 +167,7 @@ def _makeImageDiscriminator(self, img_shape):
         option = Input((1,),name="disc_options")
         option2 = Input((1,),name="disc2_options")
         ins = [img0, img, option, option2, img_goal, img_goal2]
-        dr = self.dropout_rate
-        dr = 0
+        dr = self.dropout_rate*0.
 
         x0 = AddConv2D(img0, 64, [4,4], 1, dr, "same", lrelu=True, bn=False)
         xobs = AddConv2D(img, 64, [4,4], 1, dr, "same", lrelu=True, bn=False)
@@ -209,9 +200,11 @@ def _makeImageDiscriminator(self, img_shape):
             x = Flatten()(x)
             x = AddDense(x, 1, "linear", 0., output=True, bn=False)
         else:
-            x = AddConv2D(x, 1, [1,1], 1, 0., "same", activation="sigmoid",
-                bn=False)
-            x = GlobalAveragePooling2D()(x)
+            #x = AddConv2D(x, 1, [1,1], 1, 0., "same", activation="sigmoid",
+            #    bn=False)
+            #x = GlobalAveragePooling2D()(x)
+            x = Flatten()(x)
+            x = AddDense(x, 1, "sigmoid", 0., output=True, bn=False)
 
         discrim = Model(ins, x, name="image_discriminator")
         self.lr *= 2.