Make MultiHeadAttention use masks from query and value tensors (#7951)

BUG
* Use describeWithFlags

* Separate tests out of for loop

* Make MHA use masks from query and value tensors

* Fix lint

* Refactor mask computation into a separate function

Author: mattsoulanille

tfjs-core/src/tensor.ts

@@ -275,6 +275,8 @@ export class Tensor<R extends Rank = Rank> implements TensorInfo {
   kept = false;
   /** The id of the scope this tensor is being tracked in. */
   scopeId: number;
+  /** The keras mask that some keras layers attach to the tensor */
+  kerasMask?: Tensor;
 
   /**
    * Number of elements to skip in each dimension when indexing. See
@@ -442,6 +444,9 @@ export class Tensor<R extends Rank = Rank> implements TensorInfo {
     if (this.isDisposed) {
       return;
     }
+    if (this.kerasMask) {
+      this.kerasMask.dispose();
+    }
     trackerFn().disposeTensor(this);
     this.isDisposedInternal = true;
   }

tfjs-layers/src/base_callbacks.ts

@@ -488,7 +488,8 @@ export function standardizeCallbacks(
   }
   // Convert custom callback configs to custom callback objects.
   const callbackConfigs =
-      generic_utils.toList(callbacks) as CustomCallbackArgs[];
+      generic_utils.toList<BaseCallback | CustomCallbackArgs>(
+        callbacks) as CustomCallbackArgs[];
   return callbackConfigs.map(
       callbackConfig => new CustomCallback(callbackConfig, yieldEvery));
 }

tfjs-layers/src/engine/topology.ts

@@ -751,19 +751,19 @@ export abstract class Layer extends serialization.Serializable {
    */
   protected assertInputCompatibility(inputs: Tensor|Tensor[]|SymbolicTensor|
                                      SymbolicTensor[]): void {
-    inputs = generic_utils.toList(inputs);
+    const inputsList = generic_utils.toList(inputs);
     if (this.inputSpec == null || this.inputSpec.length === 0) {
       return;
     }
     const inputSpec = generic_utils.toList(this.inputSpec);
-    if (inputs.length !== inputSpec.length) {
+    if (inputsList.length !== inputSpec.length) {
       throw new ValueError(
           `Layer ${this.name} expects ${inputSpec.length} inputs, ` +
-          `but it received ${inputs.length} input tensors. ` +
+          `but it received ${inputsList.length} input tensors. ` +
           `Input received: ${inputs}`);
     }
-    for (let inputIndex = 0; inputIndex < inputs.length; inputIndex++) {
-      const x = inputs[inputIndex];
+    for (let inputIndex = 0; inputIndex < inputsList.length; inputIndex++) {
+      const x = inputsList[inputIndex];
       const spec: InputSpec = inputSpec[inputIndex];
       if (spec == null) {
         continue;
@@ -954,20 +954,8 @@ export abstract class Layer extends serialization.Serializable {
     // Ensure inputs are all the same type.
     const inputsList = generic_utils.toList(inputs);
 
-    let allAreSymbolic = true;
-    for (const input of inputsList) {
-      if (!(input instanceof SymbolicTensor)) {
-        allAreSymbolic = false;
-        break;
-      }
-    }
-    let noneAreSymbolic = true;
-    for (const input of inputsList) {
-      if (input instanceof SymbolicTensor) {
-        noneAreSymbolic = false;
-        break;
-      }
-    }
+    const allAreSymbolic = checkAllSymbolic(inputs);
+    const noneAreSymbolic = checkNoneSymbolic(inputs);
 
     if (allAreSymbolic === noneAreSymbolic) {
       throw new ValueError(
@@ -1017,8 +1005,13 @@ export abstract class Layer extends serialization.Serializable {
 
       // Actually call the layer, collecting output(s), mask(s), and shape(s).
       if (noneAreSymbolic) {
-        let output = this.call(inputs as Tensor | Tensor[], kwargs);
-        // TODO(michaelterry): Compute the outputMask
+        let output = this.call(inputs, kwargs);
+
+        // Apply masks to the output tensors if the layer supports it.
+        if (this.supportsMasking) {
+          // TODO(mattsoulanille): pass the input tensors' masks to computeMask
+          this.setMaskMetadata(inputs, output);
+        }
 
         // If the layer returns tensors from its inputs, unmodified,
         // we copy them to avoid loss of tensor metadata.
@@ -1073,8 +1066,7 @@ export abstract class Layer extends serialization.Serializable {
           If the input tensor(s) had no previous history,
           this does nothing.
         */
-        this.addInboundNode(
-            inputs as SymbolicTensor | SymbolicTensor[], output, null, null,
+        this.addInboundNode(inputs, output, null, null,
             inputShape, outputShape, kwargs);
         this._refCount++;
 
@@ -1395,6 +1387,32 @@ export abstract class Layer extends serialization.Serializable {
     return mask;
   }
 
+  private setMaskMetadata(inputs: Tensor|Tensor[], outputs: Tensor|Tensor[],
+                          previousMask?: Tensor|Tensor[]): void {
+    if (!this.supportsMasking) {
+      return;
+    }
+
+    const outputMasks = this.computeMask(inputs, previousMask);
+    if (outputs instanceof Array && outputMasks instanceof Array) {
+      if (outputs.length !== outputMasks.length) {
+        throw new Error(`${this.name} outputs ${outputs.length} tensors `
+          + `but ${outputMasks.length} masks for those tensors`);
+      }
+      for (let i = 0; i < outputs.length; i++) {
+        outputs[i].kerasMask = outputMasks[i];
+      }
+    } else if (outputMasks instanceof Array) {
+      throw new Error(`{this.name} outputs a single tensor `
+        + `but ${outputMasks.length} masks`);
+    } else if (outputs instanceof Array) {
+      throw new Error(`{this.name} outputs ${outputs.length} tensors `
+        + `but only one mask`);
+    } else {
+      outputs.kerasMask = outputMasks;
+    }
+  }
+
   /**
    * Internal method to create an inbound node for the layer.
    *
@@ -1642,3 +1660,29 @@ export function getSourceInputs(
     }
   }
 }
+
+type MaybeSymbolic = SymbolicTensor | Tensor;
+
+function checkAllSymbolic(tensors: MaybeSymbolic | MaybeSymbolic[]
+                         ): tensors is SymbolicTensor | SymbolicTensor[] {
+  let allAreSymbolic = true;
+  for (const tensor of generic_utils.toList(tensors)) {
+    if (!(tensor instanceof SymbolicTensor)) {
+      allAreSymbolic = false;
+      break;
+    }
+  }
+  return allAreSymbolic;
+}
+
+function checkNoneSymbolic(tensors: MaybeSymbolic | MaybeSymbolic[]
+                          ): tensors is Tensor | Tensor[] {
+  let noneAreSymbolic = true;
+  for (const tensor of generic_utils.toList(tensors)) {
+    if (tensor instanceof SymbolicTensor) {
+      noneAreSymbolic = false;
+      break;
+    }
+  }
+  return noneAreSymbolic;
+}

tfjs-layers/src/layers/nlp/multihead_attention.ts

@@ -20,7 +20,7 @@
  */
 
 /* Original source: keras/layers/attention/multi_head_attention.py */
-import { Tensor, einsum, linalg, mul, ones, serialization, tidy } from '@tensorflow/tfjs-core';
+import { Tensor, einsum, linalg, logicalAnd, mul, ones, serialization, tidy } from '@tensorflow/tfjs-core';
 // tslint:disable-next-line: no-imports-from-dist
 import { arraysEqual } from '@tensorflow/tfjs-core/dist/util_base';
 
@@ -813,12 +813,20 @@ export class MultiHeadAttention extends Layer {
     return tidy(() => {
       let autoMask: Tensor;
 
+      const queryMask = query.kerasMask;
+      const valueMask = value.kerasMask;
+      if (queryMask != null) {
+        autoMask = queryMask.expandDims(2); // Shape is [B, T, 1]
+      }
+      if (valueMask != null) {
+        const mask = valueMask.expandDims(1); // Shape is [B, 1, S]
+        autoMask = autoMask ? logicalAnd(autoMask, mask) : mask;
+      }
       if (useCausalMask) {
         // the shape of the causal mask is [1, T, S]
         const mask = this.computeCausalMask(query, value);
-        autoMask = mask;
+        autoMask = autoMask ? logicalAnd(autoMask, mask) : mask;
       }
-
       if (autoMask != null) {
         // Merge attentionMask & automatic mask, to shape [B, T, S]
         attentionMask = attentionMask ?