Merge pull request #52 from antoinedemathelin/master

fix: modify BaseAdaptDeep because of batch and dataset length issues

Author: antoinedemathelin

adapt/base.py

@@ -958,9 +958,10 @@ def fit(self, X, y=None, Xt=None, yt=None, domains=None, **fit_params):
         epochs = fit_params.get("epochs", 1)
         batch_size = fit_params.pop("batch_size", 32)
         shuffle = fit_params.pop("shuffle", True)
+        buffer_size = fit_params.pop("buffer_size", None)
         validation_data = fit_params.pop("validation_data", None)
         validation_split = fit_params.pop("validation_split", 0.)
-        validation_batch_size = fit_params.pop("validation_batch_size", batch_size)
+        validation_batch_size = fit_params.get("validation_batch_size", batch_size)
 
         # 2. Prepare datasets
 
@@ -998,8 +999,7 @@ def fit(self, X, y=None, Xt=None, yt=None, domains=None, **fit_params):
                     for dom in range(self.n_sources_))
                 )
 
-            dataset_src = tf.data.Dataset.zip((dataset_Xs, dataset_ys))
-            
+            dataset_src = tf.data.Dataset.zip((dataset_Xs, dataset_ys))            
         else:
             dataset_src = X
 
@@ -1029,47 +1029,62 @@ def fit(self, X, y=None, Xt=None, yt=None, domains=None, **fit_params):
             self._initialize_networks()
             if isinstance(Xt, tf.data.Dataset):
                 first_elem = next(iter(Xt))
-                if (not isinstance(first_elem, tuple) or
-                not len(first_elem)==2):
-                    raise ValueError("When first argument is a dataset. "
-                                     "It should return (x, y) tuples.")
+                if not isinstance(first_elem, tuple):
+                    shape = first_elem.shape
                 else:
                     shape = first_elem[0].shape
+                if self._check_for_batch(Xt):
+                    shape = shape[1:]
             else:
                 shape = Xt.shape[1:]
             self._initialize_weights(shape)
 
-#         validation_data = self._check_validation_data(validation_data,
-#                                                       validation_batch_size,
-#                                                       shuffle)
+        
+        # 3.5 Get datasets length
+        self.length_src_ = self._get_length_dataset(dataset_src, domain="src")
+        self.length_tgt_ = self._get_length_dataset(dataset_tgt, domain="tgt")
+        
 
         # 4. Prepare validation dataset
         if validation_data is None and validation_split>0.:
             if shuffle:
-                dataset_src = dataset_src.shuffle(buffer_size=1024)
-            frac = int(len(dataset_src)*validation_split)
+                dataset_src = dataset_src.shuffle(buffer_size=self.length_src_,
+                                                  reshuffle_each_iteration=False)
+            frac = int(self.length_src_*validation_split)
             validation_data = dataset_src.take(frac)
             dataset_src = dataset_src.skip(frac)
-            validation_data = validation_data.batch(batch_size)
+            if not self._check_for_batch(validation_data):
+                validation_data = validation_data.batch(validation_batch_size)
+        
+        if validation_data is not None:
+            if isinstance(validation_data, tf.data.Dataset):
+                if not self._check_for_batch(validation_data):
+                    validation_data = validation_data.batch(validation_batch_size)
 
+        
         # 5. Set datasets
         # Same length for src and tgt + complete last batch + shuffle
-        try:
-            max_size = max(len(dataset_src), len(dataset_tgt))
-            max_size = np.ceil(max_size / batch_size) * batch_size
-            repeat_src = np.ceil(max_size/len(dataset_src))
-            repeat_tgt = np.ceil(max_size/len(dataset_tgt))
-
-            dataset_src = dataset_src.repeat(repeat_src)
-            dataset_tgt = dataset_tgt.repeat(repeat_tgt)
-            
-            self.total_steps_ = float(np.ceil(max_size/batch_size)*epochs)
-        except:
-            pass
-        
         if shuffle:
-            dataset_src = dataset_src.shuffle(buffer_size=1024)
-            dataset_tgt = dataset_tgt.shuffle(buffer_size=1024)
+            if buffer_size is None:                
+                dataset_src = dataset_src.shuffle(buffer_size=self.length_src_,
+                                                  reshuffle_each_iteration=True)
+                dataset_tgt = dataset_tgt.shuffle(buffer_size=self.length_tgt_,
+                                                  reshuffle_each_iteration=True)
+            else:
+                dataset_src = dataset_src.shuffle(buffer_size=buffer_size,
+                                                  reshuffle_each_iteration=True)
+                dataset_tgt = dataset_tgt.shuffle(buffer_size=buffer_size,
+                                                  reshuffle_each_iteration=True)
+        
+        max_size = max(self.length_src_, self.length_tgt_)
+        max_size = np.ceil(max_size / batch_size) * batch_size
+        repeat_src = np.ceil(max_size/self.length_src_)
+        repeat_tgt = np.ceil(max_size/self.length_tgt_)
+
+        dataset_src = dataset_src.repeat(repeat_src).take(max_size)
+        dataset_tgt = dataset_tgt.repeat(repeat_tgt).take(max_size)
+
+        self.total_steps_ = float(np.ceil(max_size/batch_size)*epochs)
 
         # 5. Pretraining
         if not hasattr(self, "pretrain_"):
@@ -1097,14 +1112,14 @@ def fit(self, X, y=None, Xt=None, yt=None, domains=None, **fit_params):
             pre_verbose = prefit_params.pop("verbose", verbose)
             pre_epochs = prefit_params.pop("epochs", epochs)
             pre_batch_size = prefit_params.pop("batch_size", batch_size)
-            pre_shuffle = prefit_params.pop("shuffle", shuffle)
             prefit_params.pop("validation_data", None)
-            prefit_params.pop("validation_split", None)
-            prefit_params.pop("validation_batch_size", None)
 
             # !!! shuffle is already done
-            dataset = tf.data.Dataset.zip((dataset_src, dataset_tgt)).batch(pre_batch_size)
-                
+            dataset = tf.data.Dataset.zip((dataset_src, dataset_tgt))
+            
+            if not self._check_for_batch(dataset):
+                dataset = dataset.batch(pre_batch_size)
+
             hist = super().fit(dataset, validation_data=validation_data,
                                epochs=pre_epochs, verbose=pre_verbose, **prefit_params)
 
@@ -1121,7 +1136,10 @@ def fit(self, X, y=None, Xt=None, yt=None, domains=None, **fit_params):
             self.history_ = {}
 
         # .7 Training
-        dataset = tf.data.Dataset.zip((dataset_src, dataset_tgt)).batch(batch_size)
+        dataset = tf.data.Dataset.zip((dataset_src, dataset_tgt))
+        
+        if not self._check_for_batch(dataset):
+            dataset = dataset.batch(batch_size)
 
         self.pretrain_ = False
 
@@ -1257,7 +1275,8 @@ def compile(self,
             if "_" in name:
                 new_name = ""
                 for split in name.split("_"):
-                    new_name += split[0]
+                    if len(split) > 0:
+                        new_name += split[0]
                 name = new_name
             else:
                 name = name[:3]
@@ -1571,6 +1590,37 @@ def _initialize_weights(self, shape_X):
             X_enc = self.encoder_(np.zeros((1,) + shape_X))
             if hasattr(self, "discriminator_"):
                 self.discriminator_(X_enc)
+                
+    
+    def _get_length_dataset(self, dataset, domain="src"):
+        try:
+            length = len(dataset)
+        except:
+            if self.verbose:
+                print("Computing %s dataset size..."%domain)
+            if not hasattr(self, "length_%s_"%domain):
+                length = 0
+                for _ in dataset:
+                    length += 1
+            else:
+                length = getattr(self, "length_%s_"%domain)
+            if self.verbose:
+                print("Done!")
+        return length
+        
+        
+    def _check_for_batch(self, dataset):
+        if dataset.__class__.__name__ == "BatchDataset":
+            return True
+        if hasattr(dataset, "_input_dataset"):
+            return self._check_for_batch(dataset._input_dataset)
+        elif hasattr(dataset, "_datasets"):
+            checks = []
+            for data in dataset._datasets:
+                checks.append(self._check_for_batch(data))
+            return np.all(checks)
+        else:
+            return False
 
 
     def _unpack_data(self, data):

adapt/feature_based/_fa.py

@@ -6,6 +6,7 @@
 
 import numpy as np
 from sklearn.utils import check_array
+from sklearn.exceptions import NotFittedError
 
 from adapt.base import BaseAdaptEstimator, make_insert_doc
 from adapt.utils import check_arrays
@@ -221,6 +222,11 @@ def transform(self, X, domain="tgt"):
         domain of ``X`` in order to apply the appropriate feature transformation.
         """
         X = check_array(X, allow_nd=True)
+        
+        if not hasattr(self, "n_domains_"):
+            raise NotFittedError("FA model is not fitted yet, please "
+                                 "call 'fit_transform' or 'fit' first.")
+        
         if domain in ["tgt", "target"]:
             X_emb = np.concatenate((np.zeros((len(X), X.shape[-1]*self.n_domains_)),
                                     X,

adapt/instance_based/_kmm.py

@@ -5,6 +5,7 @@
 import numpy as np
 from sklearn.metrics import pairwise
 from sklearn.utils import check_array
+from sklearn.exceptions import NotFittedError
 from sklearn.metrics.pairwise import KERNEL_PARAMS
 from cvxopt import matrix, solvers
 

adapt/instance_based/_tradaboost.py

@@ -384,8 +384,8 @@ def _boost(self, iboost, Xs, ys, Xt, yt,
         self.estimators_.append(estimator)
         self.estimator_errors_.append(estimator_error)
 
-        if estimator_error <= 0.:
-            return None, None
+        # if estimator_error <= 0.:
+        #     return None, None
 
         beta_t = estimator_error / (2. - estimator_error)
 

tests/test_adda.py

@@ -70,4 +70,28 @@ def test_fit():
     # assert np.sum(np.abs(
     #     model.predict(Xs, "source").ravel() - ys)) < 0.01
     assert np.sum(np.abs(np.ravel(model.predict_task(Xs, domain="src")) - ys)) < 11
+    assert np.sum(np.abs(model.predict(Xt).ravel() - yt)) < 25
+    
+    
+def test_nopretrain():
+    tf.random.set_seed(0)
+    np.random.seed(0)
+    encoder = _get_encoder()
+    task = _get_task()
+    
+    src_model = Sequential()
+    src_model.add(encoder)
+    src_model.add(task)
+    src_model.compile(loss="mse", optimizer=Adam(0.01))
+    
+    src_model.fit(Xs, ys, epochs=100, batch_size=34, verbose=0)
+    
+    Xs_enc = src_model.predict(Xs)
+    
+    model = ADDA(encoder, task, _get_discriminator(), pretrain=False,
+                 loss="mse", optimizer=Adam(0.01), metrics=["mae"],
+                 copy=False)
+    model.fit(Xs_enc, ys, Xt, epochs=30, batch_size=34, verbose=0)
+    assert np.abs(model.encoder_.get_weights()[0][1][0]) < 0.2
+    assert np.sum(np.abs(np.ravel(model.predict(Xs)) - ys)) < 25
     assert np.sum(np.abs(model.predict(Xt).ravel() - yt)) < 25

tests/test_base.py

@@ -28,6 +28,10 @@
 yt = 0.2 * Xt[:, 0].ravel()
 
 
+def _custom_metric(yt, yp):
+    return tf.shape(yt)[0]
+
+
 class DummyFeatureBased(BaseAdaptEstimator):
 
     def fit_transform(self, Xs, **kwargs):
@@ -239,4 +243,54 @@ def test_multisource():
     model.fit(Xs, ys, Xt=Xt, domains=np.random.choice(2, len(Xs)))
     model.predict(Xs)
     model.evaluate(Xs, ys)
-    assert model.n_sources_ == 2
+    assert model.n_sources_ == 2
+
+    
+def test_complete_batch():
+    model = BaseAdaptDeep(Xt=Xt[:3], metrics=[_custom_metric])
+    model.fit(Xs, ys, batch_size=120)
+    assert model.history_["cm"][0] == 120
+    
+    model = BaseAdaptDeep(Xt=Xt[:10], yt=yt[:10], metrics=[_custom_metric])
+    model.fit(Xs[:23], ys[:23], batch_size=17, buffer_size=1024)
+    assert model.history_["cm"][0] == 17
+    assert model.total_steps_ == 2
+    
+    dataset = tf.data.Dataset.zip((tf.data.Dataset.from_tensor_slices(Xs),
+                                   tf.data.Dataset.from_tensor_slices(ys.reshape(-1,1))
+                                  ))
+    Xtt = tf.data.Dataset.from_tensor_slices(Xt)
+    model = BaseAdaptDeep(Xt=Xtt, metrics=[_custom_metric])
+    model.fit(dataset, batch_size=32, validation_data=dataset)
+    assert model.history_["cm"][0] == 32
+    
+    model = BaseAdaptDeep(Xt=Xtt.batch(32), metrics=[_custom_metric])
+    model.fit(dataset.batch(32), batch_size=48, validation_data=dataset.batch(32))
+    assert model.history_["cm"][0] == 25
+    
+    def gens():
+        for i in range(40):
+            yield Xs[i], ys[i]
+            
+    dataset = tf.data.Dataset.from_generator(gens,
+                                             output_shapes=([2], []),
+                                             output_types=("float32", "float32"))
+    
+    def gent():
+        for i in range(50):
+            yield Xs[i], ys[i]
+            
+    dataset2 = tf.data.Dataset.from_generator(gent,
+                                             output_shapes=([2], []),
+                                             output_types=("float32", "float32"))
+    
+    model = BaseAdaptDeep(metrics=[_custom_metric])
+    model.fit(dataset, Xt=dataset2, validation_data=dataset, batch_size=22)
+    assert model.history_["cm"][0] == 22
+    assert model.total_steps_ == 3
+    assert model.length_src_ == 40
+    assert model.length_tgt_ == 50
+    
+    model.fit(dataset, Xt=dataset2, validation_data=dataset, batch_size=32)
+    assert model.total_steps_ == 2
+    assert model.history_["cm"][-1] == 32

tests/test_ccsa.py

@@ -20,7 +20,7 @@ def test_ccsa():
                 optimizer="adam", metrics=["acc"], gamma=0.1, random_state=0)
     ccsa.fit(Xs, tf.one_hot(ys, 2).numpy(), Xt=Xt[ind],
              yt=tf.one_hot(yt, 2).numpy()[ind], epochs=100, verbose=0)
-    assert np.mean(ccsa.predict(Xt).argmax(1) == yt) > 0.9
+    assert np.mean(ccsa.predict(Xt).argmax(1) == yt) > 0.8
 
     ccsa = CCSA(task=task, loss="categorical_crossentropy",
                 optimizer="adam", metrics=["acc"], gamma=1., random_state=0)

tests/test_cdan.py

@@ -72,8 +72,8 @@ def test_fit_lambda_one_no_entropy():
                  random_state=0, validation_data=(Xt, ytt))
     model.fit(Xs, yss, Xt, ytt,
               epochs=300, verbose=0)
-    assert model.history_['acc'][-1] > 0.9
-    assert model.history_['val_acc'][-1] > 0.9
+    assert model.history_['acc'][-1] > 0.8
+    assert model.history_['val_acc'][-1] > 0.8
 
 
 def test_fit_lambda_entropy():

tests/test_dann.py

@@ -74,7 +74,7 @@ def test_fit_lambda_one():
               epochs=100, batch_size=32, verbose=0)
     assert isinstance(model, Model)
     assert np.abs(model.encoder_.get_weights()[0][1][0] / 
-            model.encoder_.get_weights()[0][0][0]) < 0.07
+            model.encoder_.get_weights()[0][0][0]) < 0.15
     assert np.sum(np.abs(model.predict(Xs) - ys)) < 1
     assert np.sum(np.abs(model.predict(Xt) - yt)) < 2