def compute_loss(self, x, y): # y[:, :-1] y = lbann.Slice( y, axis=0, slice_points=str_list([0, self.input_feature_dims - 1]), ) y = lbann.Identity(y) # x[:, 1:] x = lbann.Slice( x, slice_points=str_list([1, self.input_feature_dims]), ) x = lbann.Identity(x) # Convert indices in x to one-hot representation # Note: Ignored indices result in zero vectors ignore_mask = lbann.Equal( x, self.constant(self.label_to_ignore, hint_layer=x), ) keep_mask = lbann.LogicalNot(ignore_mask) length = lbann.Reduction(keep_mask, mode='sum') length = lbann.Max(length, self.constant(1, [1])) x = lbann.Add( lbann.Multiply(keep_mask, x), lbann.Multiply(ignore_mask, self.constant(-1, hint_layer=x)), ) x = lbann.Slice(x, slice_points=str_list(range(self.input_feature_dims))) x = [lbann.Identity(x) for _ in range(self.input_feature_dims - 1)] x = [lbann.OneHot(xi, size=self.dictionary_size) for xi in x] x = [ lbann.Reshape(xi, dims=str_list([1, self.dictionary_size])) for xi in x ] x = lbann.Concatenation(x, axis=0) # recon_loss = F.cross_entropy( # y[:, :-1].contiguous().view(-1, y.size(-1)), # x[:, 1:].contiguous().view(-1), # ignore_index=self.pad # ) # Note: Ideally we'd shift y by y.max(-1) for numerical stability shifts = lbann.MatMul( lbann.Max(y, self.constant(0, hint_layer=y)), self.constant( 1 / math.sqrt(self.dictionary_size), [self.dictionary_size, self.dictionary_size], ), ) y = lbann.Subtract(y, shifts) z = lbann.MatMul( lbann.Exp(y), self.constant(1, [self.dictionary_size, 1]), ) z = lbann.Log(z) z = lbann.MatMul( lbann.Reshape(keep_mask, dims=str_list([1, -1])), z, ) recon_loss = lbann.MatMul( lbann.Reshape(y, dims=str_list([1, -1])), lbann.Reshape(x, dims=str_list([1, -1])), transpose_b=True, ) recon_loss = lbann.Subtract(z, recon_loss) recon_loss = lbann.Reshape(recon_loss, dims=str_list([1])) recon_loss = lbann.Divide(recon_loss, length) return recon_loss
def compute_loss(self, x, y): # y[:, :-1] y = lbann.Slice( y, axis=0, slice_points=str_list([0, self.input_feature_dims-1]), ) y = lbann.Identity(y) # x[:, 1:] x = lbann.Slice( x, slice_points=str_list([1, self.input_feature_dims]), ) x = lbann.Identity(x) # Figure out entries in x to ignore ignore_mask = lbann.Equal( x, self.constant(self.label_to_ignore, hint_layer=x), ) keep_mask = lbann.LogicalNot(ignore_mask) length = lbann.Reduction(keep_mask, mode='sum') length = lbann.Max(length, self.constant(1, [1])) # Convert entries in x to indices in y # Note: Ignored entries correspond to an index of -1. offsets = [ row*self.dictionary_size for row in range(self.input_feature_dims-1) ] offsets = lbann.Weights( initializer=lbann.ValueInitializer(values=str_list(offsets)), optimizer=lbann.NoOptimizer(), ) offsets = lbann.WeightsLayer( dims=str_list([self.input_feature_dims-1]), weights=offsets, ) y_inds = lbann.Add(x, offsets) y_inds = lbann.Add( lbann.Multiply(keep_mask, y_inds), lbann.Multiply( ignore_mask, self.constant(-1, hint_layer=y_inds), ), ) # recon_loss = F.cross_entropy( # y[:, :-1].contiguous().view(-1, y.size(-1)), # x[:, 1:].contiguous().view(-1), # ignore_index=self.pad # ) # Shift y for numerical stability # Note: We'd prefer to shift by y.max(-1) shifts = lbann.MatMul( lbann.Max(y, self.constant(0, hint_layer=y)), self.constant( 1 / math.sqrt(self.dictionary_size), [self.dictionary_size, self.dictionary_size], ), ) y = lbann.Subtract(y, shifts) # Compute log of softmax denominator and sum z = lbann.MatMul( lbann.Exp(y), self.constant(1, [self.dictionary_size, 1]), ) z = lbann.Log(z) z = lbann.MatMul( lbann.Reshape(keep_mask, dims=str_list([1, -1])), z, ) z = lbann.Reshape(z, dims=str_list([1])) # Compute cross entropy recon_loss = lbann.Gather( lbann.Reshape(y, dims=str_list([-1])), y_inds, ) recon_loss = lbann.Reduction(recon_loss, mode='sum') recon_loss = lbann.Subtract(z, recon_loss) recon_loss = lbann.Divide(recon_loss, length) return recon_loss