Here the problem is due to batch normalization layers.
During model training, the batch is generally normalized by its mean and variance.
In the training phase, the batch is normalized w.r.t. it's mean and variance. But in the testing phase, the batch is normalized w.r.t. the changing average of observed mean and variance.
In this problem, the number of observed batches is small e.g., 5 batches in your case. In the BatchNormalization layer, the moving_mean is initialized to be 0 and moving_variance is initialized to be 1.
Here the default momentum is 0.99 (given), you can update the moving averages before they converge to the "real" mean and variance.
So the prediction can be wrong in the early stage but would be correct after 1000 epochs.
You can verify it by forcing the BatchNormalization layers to operate in "training mode".
You can notice in the following example that during training, the accuracy is 1 and the loss is close to zero:
3/3 [==============================] - 19s 6s/step - loss: 1.4624 - acc: 0.3333
3/3 [==============================] - 0s 63ms/step - loss: 0.6051 - acc: 0.6667
3/3 [==============================] - 0s 57ms/step - loss: 0.2168 - acc: 1.0000
3/3 [==============================] - 0s 56ms/step - loss: 1.1921e-07 - acc: 1.0000
3/3 [==============================] - 0s 53ms/step - loss: 1.1921e-07 - acc: 1.0000
To evaluate the model, we'll observe high loss and low accuracy because after 5 updates, the moving averages are still pretty close to the initial values:
3/3 [==============================] - 3s 890ms/step
But, if we manually specify the "learning phase" variable and let the BatchNormalization layers use the "real" batch mean and variance, the result becomes the same as what's observed in fit().
sample_weights = np.ones(3)
learning_phase = 1 # 1 means "training"
ins = [imgs, y, sample_weights, learning_phase]
You can also change the momentum to a smaller value to verify.
For example, add momentum=0.01 to all the batch norm layers in ResNet50, the prediction after 20 epochs is:
array([[ 1.00000000e+00, 1.34882026e-08, 3.92139575e-22],
[ 0.00000000e+00, 1.00000000e+00, 0.00000000e+00],
[ 8.70998792e-06, 5.31159838e-10, 9.99991298e-01]], dtype=float32)
Hope this answer helps.
If you wish to learn about Machine Learning then visit this Machine Learning Course.