VariationalAutoencoderRunner.py - Fixed gaussian_sample_size causes incompatible shapes

[daviddao]

I fixed the scaling issue in VariationalAutoencoderRunner.py (See #23). However, running the default example causes following:

Epoch: 0001 cost= 1114.439753835
Epoch: 0002 cost= 662.529461080
Epoch: 0003 cost= 594.752329830
Epoch: 0004 cost= 569.599913920
Epoch: 0005 cost= 556.361018750
Epoch: 0006 cost= 545.052694460
Epoch: 0007 cost= 537.334268253
Epoch: 0008 cost= 530.251896875
Epoch: 0009 cost= 523.817275994
Epoch: 0010 cost= 519.874919247
Epoch: 0011 cost= 514.975155966
Epoch: 0012 cost= 510.715168395
Epoch: 0013 cost= 506.326094318
Epoch: 0014 cost= 502.172605824
Epoch: 0015 cost= 498.612383310
Epoch: 0016 cost= 495.592024787
Epoch: 0017 cost= 493.580289986
Epoch: 0018 cost= 490.370449006
Epoch: 0019 cost= 489.957028977
Epoch: 0020 cost= 486.818214844
W tensorflow/core/common_runtime/executor.cc:1102] 0x27f47b0 Compute status: Invalid argument: Incompatible shapes: [10000,200] vs. [128,200]
         [[Node: Mul = Mul[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](Sqrt, random_normal)]]
W tensorflow/core/common_runtime/executor.cc:1102] 0x542b0b0 Compute status: Invalid argument: Incompatible shapes: [10000,200] vs. [128,200]
         [[Node: Mul = Mul[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](Sqrt, random_normal)]]
         [[Node: range_1/_29 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_226_range_1", tensor_type=DT_INT32, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
W tensorflow/core/common_runtime/executor.cc:1102] 0x542b0b0 Compute status: Invalid argument: Incompatible shapes: [10000,200] vs. [128,200]
         [[Node: Mul = Mul[T=DT_FLOAT, _device="/job:localhost/replica:0/task:0/gpu:0"](Sqrt, random_normal)]]
         [[Node: add_1/_27 = _Recv[client_terminated=false, recv_device="/job:localhost/replica:0/task:0/cpu:0", send_device="/job:localhost/replica:0/task:0/gpu:0", send_device_incarnation=1, tensor_name="edge_225_add_1", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/cpu:0"]()]]
Traceback (most recent call last):
  File "VariationalAutoencoderRunner.py", line 53, in <module>
    print "Total cost: " + str(autoencoder.calc_total_cost(X_test))

It seems that fixing the gaussian_sample_size causes an error everytime we want to evaluate a batch of data where gaussian_sample_size != batch_size.

[daviddao]

@snurkabill any ideas?