python - 如何在Keras / TensorFlow中可视化RNN / LSTM梯度？(How to visualize RNN/LSTM gradients in Keras/TensorFlow?)

Question

I've come across research publications and Q&A's discussing a need for inspecting RNN gradients per backpropagation through time (BPTT) - ie, gradient for each timestep .

(我遇到过研究出版物，并且Q＆A讨论了需要检查每个反向传播的时间（BPTT）的RNN梯度，即每个时间步的梯度。)

The main use is introspection : how do we know if an RNN is learning long-term dependencies ?

(主要用途是自省 ：我们如何知道RNN是否正在学习长期依赖关系 ？)

A question of its own topic, but the most important insight is gradient flow :

(关于其自身主题的问题，但最重要的见解是梯度流 ：)

• If a non-zero gradient flows through every timestep, then every timestep contributes to learning - ie, resultant gradients stem from accounting for every input timestep, so the entire sequence influences weight updates

  (如果每个时间步长都流过一个非零的梯度，那么每个时间步长都有助于学习 -即，由于每个输入时间步长都产生了最终的梯度，因此整个序列都会影响权重更新)

• Per above, an RNN no longer ignores portions of long sequences , and is forced to learn from them

  (在上面，RNN 不再忽略长序列的部分 ，而是被迫向它们学习)

... but how do I actually visualize these gradients in Keras / TensorFlow?

(...但是我实际上如何在Keras / TensorFlow中可视化这些渐变呢？)

Some related answers are in the right direction, but they seem to fail for bidirectional RNNs, and only show how to get a layer's gradients, not how to meaningfully visualize them (the output is a 3D tensor - how do I plot it?)

(一些相关的答案是朝着正确的方向发展的，但对于双向RNN而言，它们似乎失败了，仅显示了如何获取图层的渐变，而不是如何有意义地可视化它们（输出是3D张量-我如何绘制它？）)

  ask by OverLordGoldDragon translate from so

Answer 1

Gradients can be fetched wrt weights or outputs - we'll be needing latter.

(可以从权重或输出中获取渐变-我们将需要后者。)

Further, for best results, an architecture-specific treatment is desired.

(此外，为了获得最佳结果，需要特定于体系结构的处理。)

Below code & explanations cover every possible case of a Keras/TF RNN, and should be easily expandable to any future API changes.

(以下代码和说明涵盖了Keras / TF RNN的所有可能情况 ，并且应该可以轻松扩展为将来的任何API更改。)

---

Completeness : code shown is a simplified version - the full version can be found at my repository, See RNN (this post included w/ bigger images);

(完整性 ：显示的代码是简化版本-完整版本可以在我的存储库中找到， 请参阅RNN （此文章带有大图）；)

included are:

(包括：)

• Greater visual custsomizability

  (更好的视觉客体性)

• Docstrings explaining all functionality

  (解释所有功能的文档字符串)

• Support for Eager, Graph, TF1, TF2, and from keras & from tf.keras

  (支持Eager，Graph，TF1，TF2， from keras和from tf.keras)

• Activations visualization

  (激活可视化)

• Weights gradients visualization (coming soon)

  (权重梯度可视化（即将推出）)

• Weights visualization (coming soon)

  (重量可视化（即将推出）)

---

I/O dimensionalities (all RNNs):

(I / O维度 （所有RNN）：)

• Input : (batch_size, timesteps, channels) - or, equivalently, (samples, timesteps, features)

  (输入 ：（ (batch_size, timesteps, channels) -或等效地， (samples, timesteps, features))

• Output : same as Input, except:

  (输出 ：与输入相同，除了：)

  • channels / features is now the # of RNN units , and:

    (channels / features现在是RNN单位的数量 ，并且：)

  • return_sequences=True --> timesteps_out = timesteps_in (output a prediction for each input timestep)

    (return_sequences=True > timesteps_out = timesteps_in （为每个输入时间步长输出预测）)

  • return_sequences=False --> timesteps_out = 1 (output prediction only at the last timestep processed)

    (return_sequences=False > timesteps_out = 1 （仅在处理的最后一个时间步输出预测）)

---

Visualization methods :

(可视化方法 ：)

• 1D plot grid : plot gradient vs. timesteps for each of the channels

  (一维绘图网格 ：每个通道的绘图梯度与时间步长)

• 2D heatmap : plot channels vs. timesteps w/ gradient intensity heatmap

  (2D热图 ：使用梯度强度热图绘制通道与时间步的关系)

• 0D aligned scatter : plot gradient for each channel per sample

  (0D对齐散点图 ：每个样本每个通道的绘图梯度)

• histogram : no good way to represent "vs. timesteps" relations

  (直方图 ：没有好的方法来表示“与时间步长”的关系)

• One sample : do each of above for a single sample

  (一个样本 ：对单个样本执行上述每个操作)

• Entire batch : do each of above for all samples in a batch;

  (整批 ：对一批中的所有样品执行上述每个操作；)

  requires careful treatment

  (需要仔细治疗)

# for below examples
grads = get_rnn_gradients(model, x, y, layer_idx=1) # return_sequences=True
grads = get_rnn_gradients(model, x, y, layer_idx=2) # return_sequences=False

---

EX 1: one sample, uni-LSTM, 6 units -- return_sequences=True , trained for 20 iterations

(EX 1：一个样本，uni-LSTM，6个单位 return_sequences=True ，经过20次迭代训练)
show_features_1D(grads[0], n_rows=2)

• Note : gradients are to be read right-to-left , as they're computed (from last timestep to first)

  (注意 ：在计算梯度（从最后一个时间步到第一个时间）时，应从右到左读取)

• Rightmost (latest) timesteps consistently have a higher gradient

  (最右边（最新）的时间步长始终具有较高的梯度)

• Vanishing gradient : ~75% of leftmost timesteps have a zero gradient, indicating poor time dependency learning

  (消失的梯度 ：最左边的时间步中约75％的梯度为零，表明时间依赖性学习较差)

---

EX 2: all (16) samples, uni-LSTM, 6 units -- return_sequences=True , trained for 20 iterations

(EX 2：所有（16）个样本，uni-LSTM，6个单位 return_sequences=True ，经过20次迭代训练)
show_features_1D(grads, n_rows=2)
show_features_2D(grads, n_rows=4, norm=(-.01, .01))

• Each sample shown in a different color (but same color per sample across channels)

  (每个样本以不同的颜色显示（但每个样本在通道之间具有相同的颜色）)

• Some samples perform better than one shown above, but not by much

  (一些样本的效果优于上面显示的样本，但效果不佳)

• The heatmap plots channels (y-axis) vs. timesteps (x-axis);

  (热图绘制通道（y轴）与时间步长（x轴）的关系；)

  blue=-0.01, red=0.01, white=0 (gradient values)

  (蓝色= -0.01，红色= 0.01，白色= 0（渐变值）)

---

EX 3: all (16) samples, uni-LSTM, 6 units -- return_sequences=True , trained for 200 iterations

(例3：所有（16）个样本，uni-LSTM，6个单位 return_sequences=True ，经过200次迭代训练)
show_features_1D(grads, n_rows=2)
show_features_2D(grads, n_rows=4, norm=(-.01, .01))

• Both plots show the LSTM performing clearly better after 180 additional iterations

  (这两个图均显示，经过180次迭代后，LSTM的性能明显更好)

• Gradient still vanishes for about half the timesteps

  (渐变仍然消失了大约一半的时间)

• All LSTM units better capture time dependencies of one particular sample (blue curve, all plots) - which we can tell from the heatmap to be the first sample.

  (所有LSTM单位都可以更好地捕获一个特定样品（蓝色曲线，所有曲线）的时间依赖性-从热图中可以看出这是第一个样品。)

  We can plot that sample vs. other samples to try to understand the difference

  (我们可以绘制该样本与其他样本的关系图，以试图了解差异)

---

EX 4: 2D vs. 1D, uni-LSTM : 256 units, return_sequences=True , trained for 200 iterations

(EX 4：2D与1D，uni-LSTM ：256个单位， return_sequences=True ，训练了200次迭代)
show_features_1D(grads[0])
show_features_2D(grads[:, :, 0], norm=(-.0001, .0001))

• 2D is better suited for comparing many channels across few samples

  (2D更适合比较少数样本中的多个通道)

• 1D is better suited for comparing many samples across a few channels

  (1D更适合于跨几个通道比较许多样本)

---

EX 5: bi-GRU, 256 units (512 total) -- return_sequences=True , trained for 400 iterations

(EX 5：bi-GRU，256个单位（总共512个） return_sequences=True ，训练了400次迭代)
show_features_2D(grads[0], norm=(-.0001, .0001), reflect_half=True)

• Backward layer's gradients are flipped for consistency wrt time axis

  (向后翻转图层的渐变以确保时间轴的一致性)

• Plot reveals a lesser-known advantage of Bi-RNNs - information utility : the collective gradient covers about twice the data.

  (该图揭示了Bi-RNN的一个鲜为人知的优势- 信息实用程序 ：集体梯度覆盖了大约两倍的数据。)

  However , this isn't free lunch: each layer is an independent feature extractor, so learning isn't really complemented

  (但是 ，这不是免费的午餐：每一层都是独立的特征提取器，因此学习并没有真正的补充)

• Lower norm for more units is expected, as approx.

  (期望更多单位的norm较低，大约为)

  the same loss-derived gradient is being distributed across more parameters (hence the squared numeric average is less)

  (相同的基于损耗的梯度分布在更多参数上（因此平方的平方数较小）)

---

EX 6: 0D, all (16) samples, uni-LSTM, 6 units -- return_sequences=False , trained for 200 iterations

(EX 6：0D，所有（16）个样本，uni-LSTM，6个单位 return_sequences=False ，训练了200次迭代)
show_features_0D(grads)

• return_sequences=False utilizes only the last timestep's gradient (which is still derived from all timesteps, unless using truncated BPTT), requiring a new approach

  (return_sequences=False仅使用最后一个时间步的梯度（除非使用截断的BPTT，否则它仍从所有时间步得出），需要一种新方法)

• Plot color-codes each RNN unit consistently across samples for comparison (can use one color instead)

  (在样本之间对每个RNN单元一致地进行颜色编码以进行比较（可以使用一种颜色代替）)

• Evaluating gradient flow is less direct and more theoretically involved.

  (梯度流的评估不那么直接，而且在理论上涉及更多。)

  One simple approach is to compare distributions at beginning vs. later in training: if the difference isn't significant, the RNN does poorly in learning long-term dependencies

  (一种简单的方法是在训练的开始阶段和以后的阶段比较分布：如果差异不大，则RNN在学习长期依赖项方面表现不佳)

<img src="https://stackoom.com/link/aHR0cHM6Ly9pLnN0YWNrLmltZ3VyLmNvbS82OTNFTy5wbmc=" width="560" referrerpolicy="no-